control device

Abstract

To suppress an increase in incompatibility of automatic control caused by a model change.SOLUTION: A control device generates a first path and a second path according to a first control model and a second control model, evaluates a divergence between the generated first path and the generated second path by using compatibility criteria, generates a final path from the first path and the second path by reflecting an evaluation result, and controls movement of a mobile body according to the generated final path. The compatibility criteria are adjusted according to user feedback so that the first path is given increased priority, the less a user performs an operation of intervention, and the second path is given increased priority, the more the user performs the operation of intervention.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] This disclosure relates to control technologies for moving bodies such as autonomous vehicles.

Background Art

[0002] In Patent Document 1, a system for autonomous vehicle control configured to determine vehicle commands from routes, GPS data, and sensor data using a trained neural network has been proposed.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] One objective of this disclosure is to provide a control technology for suppressing an increase in the incompatibility of automatic control caused by model changes.

Means for Solving the Problems

[0005] A control device according to a first aspect of this disclosure comprises a storage unit for storing a first control model and a second control model, and a control unit. The first control model is newly deployed in relation to the second control model. The control unit is configured to generate a first path using the first control model, generate a second path using the second control model, evaluate the discrepancy between the generated first and second paths using a suitability criterion, generate a final path from the first and second paths reflecting the results of the evaluation of the discrepancy, and control the movement of the moving object according to the generated final path. The suitability criterion is adjusted by user feedback such that the final path is generated with priority given to the first path as less user intervention is performed, and the final path is generated with priority given to the second path as more user intervention is performed. At least one of the first and second control models may be composed of a neural network, and deep learning may be used as the machine learning method. [Effects of the Invention]

[0006] According to this disclosure, it is expected that the number of non-conformities in automatic control caused by model changes will be reduced. [Brief explanation of the drawing]

[0007] [Figure 1] Figure 1 schematically illustrates an example of a scenario in which this disclosure applies. [Figure 2] Figure 2 schematically shows an example of the data structure of the intervention history in this disclosure. [Figure 3] Figure 3 schematically shows an example of the hardware configuration of the control device of this disclosure. [Figure 4] Figure 4 shows an example of a processing procedure related to control by the control device of this disclosure. [Modes for carrying out the invention]

[0008] Conventionally, rule-based autonomous driving systems are known. Furthermore, according to methods such as those described in Patent Document 1, an autonomous driving system can be constructed using a trained machine learning model. However, the present inventor has found the following problems with these conventional methods. For example, consider a scenario where a model change is required, such as changing from a rule-based model to a machine learning model, or updating a machine learning model. This model change results in... The content of control instructions (paths) can change under the same environment (for example, the timing of lane changes may change). In some cases, model changes can significantly alter the content of instructions. For example, if a major model change is made, such as changing the model structure (for example, changing from a rule-based model to a machine learning model, or adopting a machine learning model with a different structure), or adding or retraining a large amount of new training data, the content of instructions is likely to change significantly. When the content of instructions changes significantly in this way, the autonomous driving control provided by the new control model after the change may not necessarily be suitable for the user compared to the original control model before the change. For example, a user may feel uncomfortable with the autonomous driving control provided by the new control model because they are accustomed to the autonomous driving control provided by the original control model (i.e., it may no longer be suitable for the user). In other words, due to model changes, the autonomous driving control provided by the control model may no longer be suitable for the user (it may have been suitable with the original control model, but not with the new control model). If the autonomous driving control provided by the control model is not suitable for the user, user intervention will occur, autonomous driving will not be performed, and manual driving will be performed. An increase in the unsuitability of autonomous driving due to model changes may lead to more frequent user interventions. Furthermore, this problem can occur regardless of the type of vehicle. Moreover, this problem is not limited to situations involving vehicle control. The same issues arise with other moving objects besides vehicles when it comes to controlling their movement. Therefore, similar problems can occur when controlling any moving object other than a vehicle.

[0009] In contrast, the control device according to the first aspect of this disclosure comprises a storage unit for storing a first control model and a second control model, and a control unit. The first control model is newly deployed in relation to the second control model. The control unit is configured to generate a first path using the first control model, generate a second path using the second control model, evaluate the discrepancy between the generated first and second paths using a suitability criterion, generate a final path from the first and second paths reflecting the results of the discrepancy evaluation, and control the movement of the moving object according to the generated final path. The suitability criterion is adjusted by user feedback such that the less user intervention is performed, the more the first path is prioritized in generating the final path, and the more user intervention is performed, the more the second path is prioritized in generating the final path.

[0010] In a first aspect of this disclosure, both a new control model (first control model) and the original control model (second control model) are retained. The conformity criteria are adjusted based on user feedback to prioritize the second control model as user intervention operations are performed over automatic control by the first control model, and to prioritize the first control model as little intervention operation is performed. This increases the proportion of use of the second control model when user intervention occurs, and increases the proportion of use of the first control model when user intervention does not occur. As a result, by limiting the use of the first control model to a range where no nonconformity occurs, it is possible to suppress the increase in automatic control nonconformity caused by model changes.

[0011] Furthermore, as another form of the control device according to the above embodiment, one aspect of this disclosure may be an information processing method that implements all or part of the above components, a program, or a machine-readable storage medium that stores such a program. Here, a machine-readable storage medium is a medium that stores information such as programs by electrical, magnetic, optical, mechanical, or chemical action.

[0012] [1. Application Examples] Figure 1 schematically shows an example of a scenario in which the present disclosure is applied. The control device 1 according to this embodiment is one or more computers configured to control the automatic movement of a target mobile body M. In this embodiment, the control device 1 is mounted on the mobile body M and holds a first control model 30 and a second control model 35. Each control model (30, 35) controls the automatic movement of the mobile body M. It is configured to derive a path in order to execute the command. The first control model 30 is a new addition to the second control model 35.

[0013] In this embodiment, the control device 1 generates a first pass 40 using a first control model 30. The control device 1 also generates a second pass 45 using a second control model 35. The control device 1 evaluates the discrepancy between the generated first pass 40 and second pass 45 using a conformance criterion 20. The control device 1 generates a final pass 50 from the first pass 40 and second pass 45, reflecting the results of the discrepancy evaluation.

[0014] If there is no user intervention operation 55, the control device 1 controls the movement (movement) of the mobile body M according to the generated final path 50. On the other hand, if there is a user intervention operation 55, the control device 1 discards the generated final path 50 or overlaps with the final path 50 and controls the movement of the mobile body M according to the user intervention operation 55. The conformity criterion 20 is adjusted by user feedback so that the less user intervention operation 55 is performed, the more the first path 40 is prioritized when generating the final path 50, and the more user intervention operation 55 is performed, the more the second path 45 is prioritized when generating the final path 50.

[0015] In this embodiment, the control device 1 maintains both a new control model (first control model 30) and the original control model (second control model 35). The first control model 30 is newer than the second control model 35. Therefore, the first control model 30 is basically used preferentially over the second control model 35. That is, in the generation of the final pass 50, the first pass 40 is prioritized. However, control by the new control model (first control model 30) may not be suitable for users who are accustomed to control by the original control model (second control model 35). In particular, if there is a large discrepancy between the first pass 40 and the second pass 45, control by the first pass 40 is likely to be unsuitable (cause discomfort) for users accustomed to the second pass 45. If it is unsuitable, user intervention operations 55 may occur.

[0016] In this embodiment, the user intervention operation 55 is used as an indicator to evaluate whether the control by the new control model (first control model 30) is suitable for the user. If an intervention occurs, the control by the new control model is evaluated as not being suitable for the user, so the suitability criterion 20 is adjusted to increase the weighting of the original control model (second control model 35) (i.e., prioritize the second pass 45). On the other hand, if no intervention occurs, the control by the new control model is evaluated as being suitable for the user, so the suitability criterion 20 is adjusted to increase the weighting of the new control model (first control model 30) (i.e., prioritize the first pass 40). This makes it possible to narrow the range of use of the first control model 30 to a range where no unsuitability occurs, and as a result, it is possible to suppress the increase in unsuitability of automatic control caused by model changes.

[0017] (Mobile) If it can be automatically moved by mechanical control, the type of the moving body M may be appropriately selected according to the embodiment. The moving body M may be, for example, a movable device such as a vehicle, an aircraft, a ship, or a robot device. The aircraft may be at least one of a drone or other unmanned aircraft and a manned aircraft. In one example, as shown in FIG. 1, the moving body M may be a vehicle. The type of the vehicle (number of wheels, power source, size, etc.) may be arbitrarily selected. As a typical example, the moving body M may be an automobile having an automatic driving ability of level 2 or higher.

[0018] (Controlling the operation) In one example, controlling the operation (movement) of the target moving body M may be constituted by directly controlling the target moving body M. In another example, the moving body M may be provided with a dedicated control device such as a controller, for example. In this case, controlling the operation of the target moving body M by the control device 1 may be constituted by indirectly controlling the target moving body M by giving a derived result to the dedicated control device. The control device 1 may be deployed at an arbitrary location. In one example, as shown in FIG. 1, the control device 1 may be mounted on the moving body M. In another example, the control device 1 may be arranged away from the moving body M and remotely control the moving body M. The control device 1 may be configured to be switchable between the automatic control mode and the manual control mode of the moving body M and from the manual control mode to the automatic control mode at an arbitrary timing by an arbitrary operation.

[0019] (Path) The path (output of the control model) may not be particularly limited as long as it can control the operation of the moving body M, and may be appropriately determined according to the embodiment. In one example, the path (40, 45, 50) may be constituted by one or more control commands. The control command may be configured to indicate the control amount of the moving body M. In another example, the path (40, 45, 50) may be configured to indicate the future movement path of the moving body M and may be used to derive one or more control commands. Accordingly, one or more control commands may be determined from the final path 50 by an arbitrary method. The control model (30, 35) may be referred to as a path planner.

[0020] The control command relates to the operation of the moving body M. The configuration of the control command may be appropriately selected according to the embodiment. In one example, the control command may be constituted by acceleration, deceleration, steering, or a combination thereof. Acceleration and deceleration may include gear shifting. The control command may be configured to indicate a control amount (control instruction value, control output amount) of the moving body M, such as an accelerator control amount, a brake control amount, a steering wheel steering angle, etc. Further, the control command may further include a command related to the operation of the moving body M. As an example, when the moving body M is a vehicle, the control command may include vehicle operations such as a wiper, hazard, horn, communication processing (e.g., transmitting data to a center, making an emergency call, etc.).

[0021] (Discard or overlap) Discard (ignore) may be to immediately switch from automatic control to manual control, that is, to immediately switch from automatic control by the final path 50 to manual control according to the intervention operation 55 by the user. On the other hand, overlap may be to gradually switch from automatic control to manual control (user control).

[0022] (Control model) The control models (30, 35) are constructed to derive a path according to the environment of the moving body M. The environment is an event observed in at least one of the moving body M itself and its surroundings. In one example, at least a part of the environment may be observed by one or more sensors S disposed 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), radar, LiDAR (Light Detection And Ranging), sonar (ultrasonic sensor), infrared sensor, GNSS (Global Navigation Satellite System) / GPS (Global Positioning Satellite) module, etc.

[0023] If a path can be derived from the environment of the mobile body M, the input / output format of the control models (30, 35) may be appropriately selected according to the embodiment. In one example, at least one of the control models (30, 35) may be configured to derive a path from observation data (sensor data) of sensor S at one or more time points. In another example, at least one of the control models (30, 35) may be configured to derive a path from the recognition result of the surrounding environment. In this case, the control device 1 may further include an analysis model that infers the recognition result of the surrounding environment from the observation data of sensor S. Alternatively, at least one of the control models (30, 35) may include such an analysis model. The analysis model may be arbitrarily configured using a trained machine learning model or the like. Other information may be optionally added to the input of at least one of the control models (30, 35). It is possible. At least one of the control models (30, 35) may be configured to further accept input of arbitrary information such as set speed, speed limit, map data, and navigation information (route data). If the moving object M is a vehicle, the arbitrary information may include driving data. The arbitrary information may be appropriately obtained from devices such as navigation devices and on-board sensors. Furthermore, the control models (30, 35) may be configured to output a path directly, or to output a path indirectly. In the latter case, a path may be obtained by performing arbitrary information processing (interpretation processing) on ​​the output of the control models (30, 35).

[0024] The control model (30, 35) may consist of at least one of a trained machine learning model and a rule-based model. The rule-based model is configured to match a given input (e.g., environmental information such as observation data and recognition results of the surrounding environment) against rules and derive a path according to the matching result (according to the matching rules). The rules may be set manually or at least partially automatically. The machine learning model is configured to have one or more computational parameters that can be adjusted by machine learning. One or more computational parameters are used in the computation of the desired inference (in this disclosure, path derivation). Machine learning is the process of adjusting (optimizing) the values ​​of the computational parameters using training data. The machine learning model may consist of, for example, a neural network, a support vector machine, a regression model, a decision tree model, etc. The machine learning method may be appropriately selected depending on the machine learning model adopted (e.g., backpropagation, etc.). Machine learning may include supervised learning, unsupervised learning, and reinforcement learning.

[0025] In one example, at least one of the control models (30, 35) may be composed of a neural network at least partially. The structure of the neural network may be determined as appropriate depending on the embodiment. The structure of the neural network 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. The neural network may also include any layers such as fully connected layers, convolutional layers, pooling layers, deconvolutional layers, unpooling layers, normalization layers, dropout layers, LSTM (Long short-term memory), etc. The neural network may have any mechanism such as an attention mechanism. The neural network may include any model such as a GNN (Graph neural network), a diffusion model, a generative model (e.g., Generative Adversarial Network, Transformer, etc.). When used in Dell, the weights of the connections between each node in the control model and the thresholds of each node are examples of computational parameters. Note that when employing a machine learning model, the control model may consist of an end-to-end model structure.

[0026] Furthermore, the first control model 30 and the second control model 35 are constructed for the same purpose. That is, the first control model 30 and the second control model 35 are constructed to perform the same type of movement in the same or overlapping environment. Since the control required may differ from scene to scene, the first control model 30 and the second control model 35 may be prepared for each scene. For example, if the moving object M is a vehicle, the first control model 30 and the second control model 35 may perform, for example, lane changes, lane keeping, emergency stopping (EDSS: Emergency Driving Stop System), merging yield, and automatic parking. It's fine to have separate versions for each scene, such as those involving cars.

[0027] (Newly deployed) The first control model 30 may be a new control model provided for the second control model 35. That is, the first control model 30 may be the new control model after the change, and the second control model 35 may be the original control model before the change. In one example, the change in the control model may be a system replacement. Accordingly, the change in the control model may be, for example, replacing a rule-based model with a trained machine learning model. This may include replacing a machine learning model with a rule-based model, or modifying the structure of a machine learning model. Another example is that a model change may be an upgrade. Accordingly, a model change may include, for example, deploying a new rule-based model with modified rules, or deploying a new trained machine learning model updated through additional training or retraining. Note that each control model (30, 35) (model data) may be provided from an external server or the like as appropriate.

[0028] In one example, the first control model 30 may be composed of a trained machine learning model. The second control model 35 may be composed of a rule-based model. The structures of the trained machine learning model and the rule-based model may differ. Therefore, the characteristics of automatic control of the trained machine learning model and the rule-based model may differ from each other. As a result, when an automatic control system is updated from a rule-based model to a trained machine learning model, the control by the trained machine learning model may not be suitable for users who are accustomed to the control by the rule-based model. According to one example of this embodiment, even when updating such an automatic control system, the arbitration process based on the suitability criteria 20 can be expected to suppress the increase in automatic control incompatibility caused by the model change.

[0029] In another example, the first control model 30 may consist of an updated, trained machine learning model. On the other hand, the second control model 35 may consist of an unupdated, trained machine learning model. The update may include at least one of the above system replacement and / or version upgrade. The update may include any processing that can change the output of the control model. The update may include, for example, modifying parameter values ​​through additional training or retraining, or changing the model structure. According to this example, it is possible to suppress the increase in automatic control misfits caused by model updates.

[0030] The number of control models held by the control device 1 is not necessarily limited to two. The number of first control models 30 is not limited to one. The number of second control models 35 is not limited to one. Using the first control model 30 may include using one or more first control models 30. Using the second control model 35 may include using one or more second control models 35. In another example, the control device 1 may hold three or more control models. Corresponding to the first control models 30 and the second control models 35, a hierarchy may be defined among the three or more control models in which control models with intervention are given priority and control models with no intervention are given priority. The control device 1 may be configured to perform mediation processing among the three or more control models according to the conformity criteria 20 and generate the final pass 50. That is, generating the final pass 50 from the first pass 40 and the second pass 45 may include generating the final pass 50 from three or more passes. Typically, the first control model 30 is a control model one generation later than the second control model 35. However, the relationship between the first control model 30 and the second control model 35 is not limited to this example. The generational difference between the first control model 30 and the second control model 35 may be two or more generations.

[0031] (Conformity criteria) The conformity criterion 20 is defined to evaluate whether the automatic control by the first control model 30 is suitable for the user in relation to the discrepancy between the first control model 30 and the second control model 35. The conformity criterion 20 may be set as appropriate, provided that such conformity can be evaluated. Simply put, the conformity criterion 20 may consist of thresholds for physical indicators. For example, the conformity criterion 20 may consist of thresholds for speed, acceleration, timing of change, or a combination thereof for movement (straight, right turn, left turn, etc.). In this case, adjusting the conformity criterion 20 may consist of updating (i.e., modifying) the thresholds. The conformity criterion 20 may consist of movement scenes (lane changes, etc.) and movement environments (e.g., location, route, The system may be defined to evaluate user suitability after identifying at least one of the surrounding environment (etc.).

[0032] (Generating the final path) The generation of the final path 50 from the first path 40 and the second path 45 may be comprised of (I) selecting one of the first path 40 and the second path 45 as the final path 50 according to the evaluation results, or (II) generating the final path 50 by integrating the first path 40 and the second path 45. The final path 50 is the final derived path and is used for controlling the mobile body M.

[0033] If (I) is adopted, the priority may be to select the preferred path. As a simple example, consider a case where the conformity criterion 20 is given as a threshold for the speed of movement. In this case, the control device 1 may select the second path 45 as the final path 50 if the speed difference between the first path 40 and the second path 45 exceeds the threshold of the conformity criterion 20, and may select the first path 40 as the final path 50 otherwise. The less intervention operation 55 occurs, the larger the threshold of the conformity criterion 20 may be adjusted to make the first path 40 more likely to be selected as the final path 50. On the other hand, the more intervention operation 55 occurs, the smaller the threshold of the conformity criterion 20 may be adjusted to make the second path 45 more likely to be selected as the final path 50.

[0034] If (II) is adopted, the priority may be to increase the integration ratio. The conformity criterion 20 may include this integration ratio together with or in place of the threshold above. Adjusting the conformity criterion 20 may include adjusting the value of the integration ratio. The less intervention operation 55 occurs, the higher the ratio of the first path 40 may be, and the more intervention operation 55 occurs, the higher the ratio of the second path 45 may be. Evaluating the deviation in conformity criterion 20 may include determining the integration ratio of each path (40, 45) according to conformity criterion 20. The integration ratio may be determined according to the degree of deviation between the first path 40 and the second path 45, or it may be determined independently of the degree of deviation. Integration may be performed using any calculation such as sum, average, weighted average, etc. In one example, in the above case, the control device 1 may determine the ratio of each pass (40, 45) according to the conformity criterion 20 and generate the final pass 50 by integrating the first pass 40 and the second pass 45 at the determined ratio. In another example, if the conformity criterion 20 adopts a form that includes the above threshold, in the above case, the control device 1 may generate the final pass 50 by integrating the first pass 40 and the second pass 45 if the speed difference between the first pass 40 and the second pass 45 exceeds the threshold of the conformity criterion 20, and otherwise may select the first pass 40 as the final pass 50.

[0035] Furthermore, evaluating the deviation according to the conformance criterion 20 and generating the final pass 50 according to the evaluation result may be performed using at least one of rule-based or computational model (machine learning model)-based methods. If a rule-based method is adopted, the control device 1 may evaluate the deviation between the first pass 40 and the second pass 45 according to the conformance criterion 20 and perform the above process (I) or (II) according to the evaluation result of the deviation. If a computational model-based method is adopted, the control device 1 may perform the deviation evaluation and the above process (I) or (II) as a series of processes of the computational model. In this case, the conformance criterion 20 may be incorporated into the computational model.

[0036] Furthermore, in this embodiment, when a user intervention operation 55 occurs, the suitability criteria 20 are adjusted to increase the usage ratio of the second control model 35 (the second pass 45). Increasing the usage ratio means making it easier to select the second pass 45 (for example, by lowering the threshold) in case (I) above, and increasing the integration ratio of the second pass 45 in case (II) above. In one example, after increasing the usage ratio of the second control model 35 in this way... The conformity criterion 20 may be adjusted to gradually increase the proportion of use of the first control model 30, to the extent that no user intervention is required. This may gradually eliminate the user's discomfort with the control provided by the new control model (first control model 30) and allow the user to enjoy the benefits of the new control model.

[0037] Furthermore, even after the priority of the second path 45 has been increased due to a user intervention operation 55, user intervention operations 55 may continue to occur. In this case, in one example, the control device 1 may continue to increase the degree to which it prioritizes the second path 45. In another example, the control device 1 may use a calculation model to generate the final path 50 from the first path 40 and the second path 45. The control device 1 may adjust the values ​​of the parameters of its calculation model (for example, the integration ratio of the first path 40 and the second path 45) so that it obtains an output that matches the path resulting from the intervention operation 55. Note that if automatic control that matches the user is not executed even when the second path 45 is given maximum priority, it is possible that the second control model 35 is not matched to the user in the first place, and that the newly deployed first control model 30 is also not matched to the user. As a result, it may be difficult for the first control model 30 and the second control model 35 to generate a path that matches the user. Therefore, in one example, if such a situation occurs, the control device 1 may transmit information indicating the environmental conditions and intervention operation (or path) at the time of intervention to a computer (e.g., an external server) that provides the control model, prompting further updates to the control model.

[0038] (User) The term "user" may refer to a specific user or an unspecified user. In one example, the conformity criteria 20 may be different for each user. In this case, the conformity criteria 20 for the target user may reflect only the feedback of that user. In another example, the conformity criteria 20 may not be different. In this case, feedback from any user, including users other than the target user, may be reflected in the conformity criteria 20, and the resulting conformity criteria 20 may be applied when the target user uses the mobile device M. Note that if the mobile device M is a vehicle, the user is typically the driver.

[0039] (feedback) In one example, user feedback (whether or not intervention operation 55 was performed) may be immediately reflected in the suitability criterion 20. In another example, user feedback may be stored as intervention history 60, and the obtained intervention history 60 may be reflected retrospectively in the suitability criterion 20. User feedback may be reflected in the adjustment of the suitability criterion 20 (for example, updating the thresholds mentioned above) after identifying at least one of the movement scene and the movement environment. In this case, the conditions under which the intervention occurred (scene, environment, etc.) may be identified as appropriate.

[0040] When user feedback is accumulated as intervention history 60, the items of information to be saved as intervention history 60 are not particularly limited, as long as the suitability criteria 20 can be adjusted, and may be appropriately selected depending on the embodiment. For example, when the suitability criteria 20 is adjusted by identifying the conditions under which the intervention occurred, intervention history 60 may include information indicating the conditions under which the intervention occurred. Also, when a trained machine learning model is adopted or has the potential to be adopted as a control model (30, 35), intervention history 60 may include information for generating training data that can be used to train the machine learning model. For example, in order to identify the model to be improved, intervention history 60 may include information (e.g., identifier, etc.) for identifying the control model that was used when the intervention occurred. Furthermore, intervention history 60 may further include the user's intervention operation 55 or the path resulting from that operation 55. Information indicating the environmental conditions under which the intervention occurred, included in intervention history 60, may be used as training data (input data), and the user's intervention operation 55 or the path resulting from that operation 55 may be used as correct data (teacher signals, labels). In this case, the control model, which is composed of machine learning models, may be updated (retrained, additionally trained, etc.) based on the obtained intervention history 60. The machine learning for updating is the control The system may be executed within device 1, or it may be executed on a computer other than control device 1 (for example, an external server). The data format of the intervention history 60 may be determined as appropriate depending on the embodiment. The intervention history 60 may be stored in any database format.

[0041] Figure 2 schematically shows an example of the data structure of the intervention history 60 according to this embodiment. In the example in Figure 2, the records of the intervention history 60 include a timestamp, identification information of the control model used (used model), the environmental conditions in which the intervention occurred, and information indicating the operation of the intervention. The timestamp indicates the date and time the intervention occurred. The date and time indicated by the timestamp may be used for deleting old records, identifying records to be reflected in the conformance criteria 20 (when the intervention history within a certain period is reflected in the conformance criteria 20), etc. Note that records of the intervention history 60 may be generated for each intervention operation. The unit of intervention operation indicated by the record may be arbitrarily determined. In one example, one record (sample of the intervention history) may be generated for each intervention operation. In another example, one record may be generated for multiple interventions.

[0042] [2 Example Configurations] Figure 3 schematically shows an example of the hardware configuration of the control device 1 according to this embodiment. The control device 1 according to this embodiment is a computer in which a control unit 11, a storage unit 12, an external interface 13, an input device 14, an output device 15, and a drive 16 are electrically connected.

[0043] The control unit 11 includes a CPU (Central Processing Unit) and RAM (Random Access Memory). The memory unit 12 includes, for example, a hard disk drive, a solid-state drive, etc., and is configured to perform arbitrary information processing based on the program and various data. The control unit 11 (CPU) is an example of a processor resource. The memory unit 12 may be composed of, for example, a hard disk drive, a solid-state drive, etc. The memory unit 12 (and RAM, ROM) is an example of a memory resource and is an example of a memory unit in this disclosure. In this embodiment, the memory unit 12 stores various information such as a control program 81, first model data 300, second model data 350, conformity criteria information 200, intervention history data 600, etc.

[0044] The control program 81 is a program that causes the control device 1 to execute information processing related to the control of the mobile body M (Figure 4, described later). The control program 81 includes a series of instructions for said information processing. The first model data 300 is configured to show information related to the first control model 30. The second model data 350 is configured to show information related to the second control model 35. In one example, if a machine learning model is adopted for the control model, the model data (300, 350) may include information showing the values ​​of the computation parameters adjusted by machine learning. Depending on the case, the model data (300, 350) may further include information showing the configuration of the machine learning model (e.g., the structure of the neural network). In another example, if a rule-based model is adopted for the control model, the model data (300, 350) may include information showing the rule-based model (rules). The conformity criterion information 200 may be configured as appropriate to show the conformity criterion 20. The intervention history data 600 may be configured as appropriate to show the intervention history 60.

[0045] The external interface 13 may be, for example, a USB (Universal Serial Bus) port, a dedicated port, a wireless communication port, etc., and is configured to connect to an external device by wire or wireless connection. In this embodiment, the control device 1 may be connected to the sensor S via the external interface 13. The input device 14 is, for example, a device for inputting data such as a mouse, keyboard, or control. The output device 15 is, for example, a device for outputting data such as a display or speaker. The input device 14 and the output device 15 may be integrated together by, for example, a touch panel display.

[0046] Drive 16 is for reading various information such as programs stored on the storage medium 91. This is a device. At least one of the above-mentioned control program 81, first model data 300, second model data 350, conformity criteria information 200, and intervention history data 600 may be stored in the storage medium 91 instead of or together with the storage unit 12. The storage medium 91 is configured to store various information (stored programs, etc.) by electrical, magnetic, optical, mechanical, or chemical means so that a machine such as a computer can read the information. The control device 1 may acquire at least one of the above-mentioned control program 81, first model data 300, second model data 350, conformity criteria information 200, and intervention history data 600 from the storage medium 91. The storage medium 91 may be a disk-type storage medium such as a CD or DVD, or a non-disk-type storage medium such as semiconductor memory (e.g., flash memory). The type of drive 16 may be appropriately selected according to the type of storage medium 91.

[0047] Regarding the specific hardware configuration of the control device 1, components can be omitted, replaced, and added as appropriate depending on the embodiment. For example, the control unit 11 may include multiple hardware processors. Hardware processors include microprocessors, FPGAs (field-programmable gate arrays), DSPs (digital signal processors), and ECUs. It consists of components such as an Electronic Control Unit (ECC) and a Graphics Processing Unit (GPU). At least one of the external interface 13, input device 14, output device 15, and drive 16 may be omitted. The control device 1 may be equipped with a communication interface and configured to perform data communication with an external computer. At least one of the input device 14, output device 15, and drive 16 may be connected via an external interface or a communication interface. The control device 1 may be a computer designed specifically for the services provided, a general-purpose computer, a terminal device, etc. If the mobile body M is a vehicle, the control device 1 may be an in-vehicle device.

[0048] [3 Examples of operation] Figure 4 shows an example of a processing procedure for controlling the mobile body M by the control device 1 according to this embodiment. The control unit 11 of the control device 1 executes instructions included in the control program 81 using the CPU. As a result, the control device 1 operates as a computer capable of performing the following information processing. The following processing procedure is an example of a control method executed by the computer. However, the following processing procedure is merely an example, and each step may be modified as much as possible. Furthermore, steps in the following processing procedure can be omitted, replaced, and added as appropriate, depending on the embodiment.

[0049] In step S101, the control unit 11 acquires observation data from the sensor S. The control unit 11 may acquire observation data directly or indirectly from the sensor S. In step S102, the control unit 11 generates the first path 40 and the second path 45 using the first control model 30 and the second control model 35. The calculations of each control model (30, 35) may be performed as appropriate. In one example, if at least one of the control models (30, 35) is a trained machine learning model, the control unit 11 may input at least a portion of the observation data into the trained machine learning model and perform calculations of the trained machine learning model to obtain the path derivation result. In another example, if at least one of the control models (30, 35) is a rule-based model, the control unit 11 may derive a path from the environment indicated by at least a portion of the observation data according to the rules.

[0050] In step S103, the control unit 11 evaluates the discrepancy between the generated first pass 40 and second pass 45 using the conformity criterion 20. In step S104, the control unit 11 generates the final pass 50 from the first pass 40 and second pass 45, reflecting the results of the discrepancy evaluation. In one example, the control unit 11 may select one of the first pass 40 and second pass 45 as the final pass 50 depending on the evaluation results. In another example, the control unit 11 may select the first pass 40 and second pass 45 The final pass 50 may be generated by integrating 45. The processes in steps S103 and S104 may be executed separately or as a series of processes according to the computation model.

[0051] In step S105, the control unit 11 determines whether or not a user intervention operation 55 has been performed. The control unit 11 may accept user intervention operations 55 at any time while controlling the movement of the mobile body M. If no intervention operation 55 is accepted (there is no intervention operation 55), the control unit 11 proceeds to step S106. On the other hand, if an intervention operation 55 is accepted (there is an intervention operation 55), the control unit 11 proceeds to step S107.

[0052] In step S106, the control unit 11 controls the movement of the mobile body M according to the generated final path 50. Once the movement control is complete, the control unit 11 proceeds to the next step S108. In step S107, the control unit 11 discards the generated final path 50 or overlaps with the final path 50 and controls the movement of the mobile body M according to the user intervention operation 55. Once the movement control is complete, the control unit 11 proceeds to the next step S108.

[0053] In step S108, the control unit 11 adjusts the suitability criteria 20 based on user feedback. The control unit 11 adjusts the suitability criteria 20 so that the first path 40 is prioritized and the final path 50 is generated less often than the user intervention operation 55 is performed (for example, by increasing the threshold or increasing the integration ratio of the first path 40). On the other hand, the control unit 11 adjusts the suitability criteria 20 so that the second path 45 is prioritized and the final path 50 is generated more often than the user intervention operation 55 is performed (for example, by decreasing the threshold or increasing the integration ratio of the second path 45).

[0054] In step S109, the control unit 11 determines whether or not to terminate control of the mobile body M. The criteria for this determination can be set arbitrarily. For example, while the mobile body M is running, the control unit 11 may determine not to terminate control of the mobile body M, but may determine to terminate control of the mobile body M in response to an arbitrary termination instruction (for example, a termination operation by the user via the input device 14). If it is determined not to terminate control, the control unit 11 returns to step S101 and executes the process again from step S101. On the other hand, if it is determined to terminate control, the processing procedure of the control device 1 according to this example of operation is terminated.

[0055] The processing order of each step is not limited to the example in Figure 4 and may be changed as appropriate depending on the embodiment. For example, the processing in step S108 does not have to be performed every control cycle. In another example, the processing in step S108 may be performed at predetermined intervals (for example, after a predetermined number of times / a predetermined time has elapsed, or after ignition ON / OFF). In yet another example, the processing in step S108 may be performed independently of the processing in steps S101 to S107. When the processing in step S108 (adjustment of conformity criteria 20) is performed retrospectively, the control unit 11 may generate information indicating the user's intervention operation 55 and save the generated information as intervention history 60 in a predetermined storage area. In one example, the predetermined storage area may be RAM, storage unit 12, storage medium 91, etc. If the control device 1 is configured to communicate, the predetermined storage area may be an external computer. The control unit 11 may use the obtained intervention history 60 to adjust the conformity criteria 20 at any timing.

[0056] [Features] In this embodiment, the process in step S108 adjusts the conformity criterion 20 based on user feedback so that the second control model 35 is prioritized more often when an intervention operation 55 is performed, and the first control model 30 is prioritized more often when an intervention operation 55 is not performed. As a result, in the process in step S104, the second path 45 is prioritized more often when an intervention occurs, and the first control model 30 is prioritized more often when an intervention occurs. The first pass 40 is given priority, and the final pass 50 is generated accordingly. As a result, the first control model 30 (first pass 40) is used preferentially, and the situations in which automatic control is performed by the processing in step S106 (situations in which the first control model 30 is used) can be narrowed down to a range in which no inconsistencies occur. Therefore, according to this embodiment, it is possible to suppress the increase in automatic control inconsistencies caused by model changes.

[0057] [4. Variant] While embodiments of this disclosure have been described in detail above, the above description is merely illustrative in all respects of this disclosure. Needless to say, various improvements or modifications can be made without departing from the scope of this disclosure. The processes and means described in this disclosure can be freely combined and implemented as long as no technical inconsistencies arise. [Explanation of symbols]

[0058] 1...Control device, 11...Control unit, 12...Storage unit, 20...Conformance criteria, 30...First control model, 35...Second control model, 40...First pass, 45...Second pass, 50...Final pass, 55...(Intervention) operation, 60...Intervention history M...Mobile object, S...Sensor

Claims

1. A storage unit for storing the first control model and the second control model, and Control unit, A control device comprising, The first control model is newly deployed in relation to the second control model, The control unit, The first path is generated by using the first control model described above. The second path is generated by using the second control model described above. The discrepancy between the generated first and second paths is evaluated using conformity criteria. Based on the results of evaluating the aforementioned discrepancy, the final pass is generated from the first pass and the second pass, and Controlling the movement of the moving object according to the generated final path, It is configured to perform, and The suitability criteria are adjusted based on user feedback such that the first path is prioritized when generating the final path if the user does not perform any intervention actions, and the second path is prioritized when generating the final path if the user does perform any intervention actions. Control device.

2. The first control model is comprised of a trained machine learning model. The second control model is composed of a rule-based model. The control device according to claim 1.

3. The first control model consists of an updated and trained machine learning model. The second control model consists of the previously trained machine learning model. The control device according to claim 1.

4. Generating the final pass from the first pass and the second pass is Selecting either the first pass or the second pass as the final pass, or The final pass is generated by integrating the first pass and the second pass. Composed of, The control device according to claim 1.

5. The aforementioned moving object is a vehicle. The control device according to any one of claims 1 to 4.

Images