Group control system and group control method

Abstract

To provide a group control system and a group control method that are able to appropriately control a plurality of moving bodies.SOLUTION: A group control system 10 includes: an acceleration data acquisition unit that acquires acceleration data of each moving body 200; a position estimation unit 102 that estimates a current position of each moving body 200; and a moving-body control unit 103 that repeatedly performs a process of inputting first state information including the acceleration data and the estimated position of each moving body 200 and second state information indicating a future state of each moving body 200, to a deep layer reinforcement learning model M1 and controlling an action of each moving body 200, based on output from the deep layer reinforcement learning model M1. The second state information input to the deep layer reinforcement learning model M1 indicates a future state of each moving object 200 predicted using an output obtained by inputting seed information to a hostile generation network M2.SELECTED DRAWING: Figure 1

Description

Technical Field

[0001] The present disclosure relates to a group control system and a group control method.

Background Art

[0002] In recent years, the development of systems using a plurality of mobile bodies such as robots capable of autonomous driving has advanced. Along with this, various technologies regarding group control for controlling a plurality of mobile bodies have been proposed. For example, Patent Document 1 discloses a group control system capable of avoiding collisions between mobile bodies.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] For example, when the error between the position of each mobile body grasped by a device performing group control of mobile bodies and the actual position of each mobile body increases, it may not be possible to appropriately control the mobile bodies. Against such a background, a technology capable of appropriately controlling a plurality of mobile bodies is demanded.

[0005] The present disclosure has been made against the background of the above circumstances, and an object thereof is to provide a group control system and a group control method capable of appropriately controlling a plurality of mobile bodies.

Means for Solving the Problems

[0006] One aspect of the present disclosure for achieving the above objective is a swarm control system for controlling a plurality of mobile bodies capable of autonomously traveling within a predetermined area, comprising: an acceleration data acquisition unit that acquires acceleration data detected by sensors provided on each mobile body; a position estimation unit that estimates the current position of each mobile body; and a mobile body control unit that inputs first state information including the acceleration data and estimated position of each mobile body, and second state information representing the future state of each mobile body, into a deep reinforcement learning model, and repeatedly executes a process to control the behavior of each mobile body based on the output from the deep reinforcement learning model, wherein the second state information input into the deep reinforcement learning model represents the future state of each mobile body predicted using the output obtained by inputting seed information into a generative adversarial network, and the seed information includes information output from the deep reinforcement learning model in the previous process.

[0007] In one embodiment described above, the deep reinforcement learning model may be a model in which deep reinforcement learning has been performed in an environment in which rewards are given based on the difference between the position of each moving object estimated by the position estimation unit and the actual position of each moving object.

[0008] In one embodiment described above, the second state information may include the position and acceleration data of each moving body at a future point in time.

[0009] In one embodiment described above, the generative adversarial network outputs a plurality of images showing the position of each of the moving objects at different points in the future, and the second state information may include the acceleration data at a future point in time calculated from the positions of each of the moving objects at different points in time shown in the plurality of images.

[0010] Another aspect of the present disclosure for achieving the above objective is a group control method for controlling a plurality of mobile bodies capable of autonomously traveling within a predetermined area, the method comprising: acquiring acceleration data detected by sensors provided on each mobile body; estimating the current position of each mobile body; inputting first state information, including the acceleration data and the estimated position of each mobile body, and second state information, representing the future state of each mobile body, into a deep reinforcement learning model; repeatedly executing a process to control the behavior of each mobile body based on the output from the deep reinforcement learning model, wherein the second state information input into the deep reinforcement learning model represents the future state of each mobile body predicted using the output obtained by inputting seed information into a generative adversarial network, and the seed information includes information output from the deep reinforcement learning model in the previous processing. [Effects of the Invention]

[0011] This disclosure provides a group control system and a group control method that can appropriately control multiple moving objects. [Brief explanation of the drawing]

[0012] [Figure 1] This is a block diagram showing an example of the configuration of a group control system according to an embodiment. [Figure 2] This is a schematic diagram showing an example of the configuration of a machine learning model used by the mobile unit control unit. [Figure 3] This is a flowchart illustrating an example of the operation flow of an information processing device. [Figure 4] This is a block diagram showing an example of the hardware configuration of an information processing device. [Modes for carrying out the invention]

[0013] The embodiments will now be described with reference to the drawings. Figure 1 is a block diagram showing an example of the configuration of a group control system 10 according to the embodiment. As shown in Figure 1, the group control system 10 includes a plurality of mobile bodies 200 and an information processing device 100 that manages these plurality of mobile bodies 200. The information processing device 100 and each mobile body 200 are connected wirelessly.

[0014] The mobile unit 200 is a device capable of autonomous movement and is also referred to as an autonomous mobile robot. For example, the mobile unit 200 is an AMR (Autonomous Mobile Robot) and may perform tasks such as transporting luggage. The mobile unit 200 performs various tasks that involve movement. For example, the tasks performed may be transport tasks, such as moving from one point to another for the purpose of transporting luggage, or charging tasks, such as moving to a location with charging equipment to charge the rechargeable battery that powers the mobile unit 200. In this embodiment, each mobile unit 200 includes, in particular, an IMU (Inertial Measurement Unit) 201, a LiDAR (Light Detection And Ranging) sensor 202, a position detection unit 203, a driving control unit 204, and a communication processing unit 205, as shown in Figure 1.

[0015] The IMU201 is a sensor that detects the acceleration of the moving object 200, detecting acceleration in each of the three orthogonal axes (X, Y, and Z axes), as well as angular acceleration around each of these three axes.

[0016] The LiDAR sensor 202 is a sensor that uses laser light to measure the shape and distance of objects around the moving object 200.

[0017] The position detection unit 203 detects the position of the moving object 200 by comparing the measurement data from the LiDAR sensor 202 with map data of the environment in which the moving object 200 is moving. More specifically, the position detection unit 203 detects the position of the moving object 200 at the time of measurement by the position detection sensor (LiDAR sensor 202).

[0018] The driving control unit 204 controls the autonomous driving of the mobile body 200. The driving control unit 204 controls the driving of the mobile body 200 based on the position of the mobile body 200 detected by the position detection unit 203 and the travel path to the destination point for task execution assigned by the information processing device 100. In other words, the driving control unit 204 controls the driving of the mobile body 200 so that it moves to the destination point. If the travel path assigned by the information processing device 100 is a set of a series of waypoints to the destination point, that is, if only discrete waypoints are notified to the mobile body 200 as part of the travel path and the travel path to each waypoint is not notified, the driving control unit 204 may create a specific travel path to the next waypoint to be visited and drive the mobile body 200 along the created travel path.

[0019] The communication processing unit 205 performs information transmission and reception processing with the information processing unit 100. In this embodiment, the communication processing unit 205 transmits the position of the moving object 200 detected by the position detection unit 203 to the information processing unit 100. The communication processing unit 205 also transmits the acceleration data detected by the IMU 201 to the information processing unit 100. In other words, the communication processing unit 205 transmits the acceleration data of the moving object 200 to the information processing unit 100. The transmission of the position of the moving object 200 detected by the position detection unit 203 and the acceleration data detected by the IMU 201 to the information processing unit 100 is performed periodically.

[0020] Next, the information processing apparatus 100 will be described. The information processing apparatus 100 is a device that controls a plurality of mobile bodies 200 capable of autonomously traveling within a predetermined area. In particular, the information processing apparatus 100 assigns tasks to the mobile bodies 200 present within the predetermined area and notifies the mobile bodies 200 of the movement paths for executing the tasks. As shown in FIG. 1, in the present embodiment, the information processing apparatus 100 particularly includes a mobile body information acquisition unit 101, a position estimation unit 102, a movement control unit 103, a model storage unit 104, and a model learning unit 105. Note that, in the configuration shown in FIG. 1, the configuration in which the information processing apparatus 100 includes the model learning unit 105 is shown, but the model learning unit 105 may be realized by another device. That is, the learning process described later may be executed by a device different from the information processing apparatus 100.

[0021] The mobile body information acquisition unit 101 acquires information from each mobile body 200. Specifically, the mobile body information acquisition unit 101 acquires the information transmitted by the communication processing unit 205 of the mobile body 200. In the present embodiment, the mobile body information acquisition unit 101 acquires the position of the mobile body 200 detected by the position detection unit 203. Focusing on the function of acquiring the position of the mobile body 200, the mobile body information acquisition unit 101 may be referred to as a position acquisition unit. Further, the mobile body information acquisition unit 101 acquires the acceleration data of each mobile body 200. Focusing on the function of acquiring the acceleration data, the mobile body information acquisition unit 101 may be referred to as an acceleration data acquisition unit.

[0022] The position estimation unit 102 estimates the current position of each mobile body 200. Although the position information of the mobile body 200 is acquired by the mobile body information acquisition unit 101, the position indicated by the acquired position information is not necessarily the current position of the mobile body 200. That is, when the position acquired by the mobile body information acquisition unit 101 is the position of the mobile body 200 detected at a time point retroactively from the current time point, due to the movement of the mobile body 200 during the period from the position detection time point to the current time point, the current position of the mobile body 200 is different from the position acquired by the mobile body information acquisition unit 101.

[0023] In this embodiment, the position estimation unit 102 estimates the current position of each mobile body 200 based on a predetermined moving speed for the mobile body 200 and the position of the mobile body 200 at the time of position detection, as notified by the mobile body 200. The predetermined moving speed (hereinafter referred to as the planned moving speed) is a constant speed and may differ for each mobile body 200, or it may be a speed common to all mobile bodies 200. In this embodiment, the mobile body 200 accelerates until it reaches the planned moving speed, and once the speed of the mobile body 200 reaches the planned moving speed, it moves while maintaining the planned moving speed. More specifically, the position estimation unit 102 estimates the current position of the mobile body 200 by assuming that the mobile body 200 moved at the planned moving speed along the movement path assigned to the mobile body 200, starting from the position of the mobile body 200 at the time of position detection.

[0024] Incidentally, when the mobile body 200 moves from point A to another point B, it is not possible to move at the planned speed for the entire distance from point A to point B. That is, acceleration is necessary near point A, and the speed of the mobile body 200 at that time is less than or equal to the planned speed. Similarly, deceleration is necessary near point B, and the speed of the mobile body 200 at that time is also less than or equal to the planned speed. For this reason, the position estimated by the position estimation unit 102 may differ from the actual position of the mobile body 200. The possibility of such a discrepancy increases as the travel environment becomes a narrower area. This is because the distance over which the planned speed can be maintained becomes shorter. In this embodiment, as will be described later, the mobile body control unit 103 controls the behavior (movement) of the mobile body 200 so that the mobile body 200 acts (moves) in a way that minimizes the discrepancy between the position estimated by the position estimation unit 102 and the actual position of the mobile body 200. This makes it possible to minimize the discrepancy between the position grasped by the information processing device 100 and the actual position of each mobile body 200, even in confined areas, and to appropriately manage the positions of multiple mobile bodies 200. In other words, it enables the appropriate operation of the mobile bodies 200 in confined areas.

[0025] The mobile unit control unit 103 controls the actions of each mobile unit 200. Specifically, the mobile unit control unit 103 assigns tasks to the mobile units 200 and assigns a travel path for executing those tasks. The mobile unit control unit 103 then notifies each mobile unit 200 of the assigned task and travel path. The mobile unit control unit 103 may also notify only the travel path. In this embodiment, the travel path consists of a series of waypoints to the destination.

[0026] The mobile unit control unit 103 assigns various tasks that occur in the environment to the mobile units 200, and also assigns a movement path to the mobile units 200 for executing those tasks. Each mobile unit 200 is assigned one of a predetermined set of tasks. Each mobile unit 200 is also assigned one of a predetermined set of movement paths. In this embodiment, the mobile unit control unit 103 first assigns tasks that occur in the environment to each mobile unit 200 based on its current position, etc. Then, the mobile unit control unit 103 adjusts the tasks and movement paths assigned to each mobile unit 200 using a trained machine learning model stored in the model storage unit 104. In other words, the mobile unit control unit 103 controls the behavior of each mobile unit 200 using the machine learning model. Specifically, the mobile unit control unit 103 repeatedly performs the following process. This process involves inputting first state information and second state information into a deep reinforcement learning model and controlling the behavior of each mobile unit 200 based on the output from this deep reinforcement learning model. Here, the first state information includes acceleration data of each mobile body 200 acquired by the mobile body information acquisition unit 101 and the position of each mobile body 200 estimated by the position estimation unit 102. Specifically, the acceleration data in the first state information is the most recently acquired acceleration data (latest acceleration data), and may also include acceleration data acquired earlier. Similarly, the position in the first state information is specifically the most recently estimated position (latest estimated position), and may also include positions estimated earlier. The second state information represents the future state of each mobile body 200 predicted using the output of the generative adversarial network. Here, the future state may be, for example, a few minutes ahead, or preferably a few seconds ahead. In this embodiment, the second state information includes the position and acceleration data of each mobile body 200 at a future point in time. Note that the second state information may include only the position of each mobile body 200 at a future point in time, or it may include other information (for example, the velocity of each mobile body 200 at a future point in time).

[0027] Figure 2 is a schematic diagram showing an example of the configuration of the machine learning model M used by the mobile unit control unit 103. The mobile unit control unit 103 repeatedly performs the process of acquiring the output of the deep reinforcement learning model M1 in the machine learning model M as time progresses. Each time the mobile unit control unit 103 acquires the output of the deep reinforcement learning model M1, it controls the actions of each mobile unit 200 based on this output. The deep reinforcement learning model M1 is a Deep Q-Network composed of an input layer, a hidden layer, and an output layer. As described above, the input layer receives first state information and second state information for each mobile unit 200. However, if the second state information is not available, dummy information is input as the second state information. The output of the deep reinforcement learning model M1 is the Q value of each action in a given state (i.e., the state input to the deep reinforcement learning model M1). In this embodiment, the first state information includes not only the acceleration data of each mobile body 200 acquired by the mobile body information acquisition unit 101 and the position of each mobile body 200 estimated by the position estimation unit 102, but also the task currently assigned to each mobile body 200 and the movement path currently assigned to each mobile body 200 for the execution of that task. In this embodiment, the action associated with the Q value is a pair of the task to be assigned to each mobile body 200 and the movement path to be assigned to each mobile body 200. That is, each node in the output layer of the deep reinforcement learning model M1 corresponds to the action of each mobile body 200 (the task and movement path of each mobile body 200). Therefore, the recommended task and movement path for each mobile body 200 can be obtained from the output of the deep reinforcement learning model M1. The mobile body control unit 103 changes the task and movement path to be assigned to each mobile body 200 according to the output result of the deep reinforcement learning model M1. The mobile body control unit 103 then notifies each mobile body 200 of the new assignment.

[0028] As shown in Figure 2, the machine learning model M is a model composed of a deep reinforcement learning model M1 and a generative adversarial network M2. In particular, the output of the deep reinforcement learning model M1 is used as seed information input to the generative adversarial network M2. The output of the generative adversarial network M2 is used as the second state information input to the deep reinforcement learning model M1. That is, the second state information input to the deep reinforcement learning model M1 represents the future state of each mobile object 200 predicted using the output obtained by inputting the seed information to the generative adversarial network M2. Here, the seed information includes the first state information input to the deep reinforcement learning model M1 in the previous process in the repeated process (for example, the position of each mobile object 200) and the information output from the deep reinforcement learning model M1 in the previous process. That is, the seed information includes the state of each mobile object 200 at a certain point in time and the action recommended at that point in time. Here, the state of each mobile object 200 at a given time, used as seed information, includes at least the position of each mobile object 200, and may also include acceleration. Thus, the second state information input to the deep reinforcement learning model M1 in the nth iteration of the repeated process (where n is an integer greater than or equal to 2) is obtained by inputting the information output from the deep reinforcement learning model M1 in the (n-1)th iteration into the generative adversarial network M2.

[0029] The adversarial generative network M2 outputs an image showing the position of each mobile object 200 at a future point in time, based on this seed information. More specifically, the seed information is input to the generator of the adversarial generative network M2, and an image is output from the generator. The output image is a map of a predetermined area (i.e., the area where each mobile object 200 moves), and is an image of the map with the position of each mobile object 200 plotted. In this embodiment, in order to also obtain acceleration data for each mobile object 200 at a future point in time, the adversarial generative network M2 outputs multiple images showing the position of each mobile object 200 at different points in the future. If time-series data of the future position of each mobile object 200 is obtained in this way, the velocity data and acceleration data of each mobile object 200 can be easily calculated. Therefore, the mobile object control unit 103 may use the output of the adversarial generative network M2 to predict not only the position of each mobile object 200 but also the acceleration of each mobile object 200 as a future state. Thus, in this embodiment, the second state information input to the deep reinforcement learning model M1 includes acceleration data for a future time point calculated from the positions of each moving object 200 at different time points shown in multiple images output by the generative adversarial network M2. If the acceleration data of each moving object 200 at a future time point is not used as the second state information, in other words, if only the positions of each moving object 200 at a future time point is used as the second state information, the generative adversarial network M2 may generate a single image showing the positions of each moving object 200 at one future time point.

[0030] The machine learning model M described above is generated in advance by the model learning unit 105 and stored in the model storage unit 104. The model learning unit 105 performs the process of generating the machine learning model M. That is, the model learning unit 105 performs the training phase processing of the machine learning model M.

[0031] The model learning unit 105 first performs training of the deep reinforcement learning model M1 (hereinafter referred to as the first learning process) using only the first state information as input. Specifically, the model learning unit 105 performs deep reinforcement learning of the deep reinforcement learning model M1 using an environment in which rewards are set so that rewards are given based on the difference between the position of each mobile object 200 estimated by the position estimation unit 102 and the actual position of each mobile object 200. That is, the model learning unit 105 updates the parameters of the deep reinforcement learning model M1 by repeating the following process. This process inputs the first state information obtained when each mobile object 200 is made to act in the environment in which the above-mentioned rewards are set to input to the deep reinforcement learning model M1, and causes each mobile object 200 to execute an action selected based on the output obtained from the deep reinforcement learning model M1 (i.e., movement according to the assigned task and movement path). As a result, the deep reinforcement learning model M1 is trained to minimize the discrepancy between the position of each mobile object 200 estimated by the position estimation unit 102 and the actual position of each mobile object 200. In setting the reward, for example, a larger reward is obtained the smaller the difference between the position of each mobile object 200 estimated by the position estimation unit 102 and the actual position of each mobile object 200. The difference referred to here may be, for example, the average of the position differences for each mobile object 200, or the sum of the position differences for each mobile object 200.

[0032] Next, the model learning unit 105 trains the generative adversarial network M2 to generate the second state information described above (hereinafter referred to as the second learning process). Specifically, the model learning unit 105 trains the generative adversarial network M2 by comparing the images generated by the generator with images as real samples using a discriminator. In this embodiment, the real samples are time-series data of the positions of each moving object 200 that have been actually observed, that is, a set of images showing the position of each moving object 200 at each time point. For example, the model learning unit 105 may use the time-series data of the positions of each moving object 200 observed during the first learning process as real samples to perform the second learning process. Alternatively, the model learning unit 105 may use time-series data of seed information obtained during the first learning process to generate the image to be output to the discriminator using the generator. This trains the generative adversarial network M2 to generate images showing the future positions of each moving object 200.

[0033] After training the adversarial generative network M2, the model learning unit 105 retrains the deep reinforcement learning model M1 that has undergone the first training process (hereinafter referred to as the third training process). That is, after it becomes possible to generate the second state information, the model learning unit 105 performs the third training process. In this training process, the model learning unit 105 trains the deep reinforcement learning model M1 using the first and second state information as input. This training process is the same as the first training process, except that the second state information is added as input to the deep reinforcement learning model M1. In order to generate the second state information, the output of the adversarial generative network M2 after the second training process is completed is used. In the third training process as well, the second state information that is input to the deep reinforcement learning model M1 in the nth iteration of the repeated process is the information obtained by inputting the information output from the deep reinforcement learning model M1 in the (n-1)th iteration into the adversarial generative network M2. If the second state information cannot be obtained (first processing), dummy information is input as the second state information. As a result, the deep reinforcement learning model M1 is trained so that the discrepancy between the position of each moving object 200 estimated by the position estimation unit 102 and the actual position of each moving object 200 is further reduced. With this, the training of the machine learning model M is completed. The model learning unit 105 stores the trained machine learning model M in the model storage unit 104.

[0034] Next, the operation flow of the information processing device 100 described above will be explained with reference to the flowchart shown in Figure 3. First, in step S10, the deep reinforcement learning model M1 is trained (first training process) without using the future state (second state information). Next, in step S11, the generative adversarial network M2 is trained (second training process). Next, in step S12, the deep reinforcement learning model M1 is trained (third training process) using the future state as well. With these processes, the training of the machine learning model M is completed. After that, in step S13, the group control system 10 is operated using the trained machine learning model M.

[0035] Next, an example of the hardware configuration of the information processing device 100 will be explained using Figure 4. The information processing device 100 has a network interface 150, a memory 151, and a processor 152.

[0036] The network interface 150 is used to communicate with other devices (mobile devices 200). The network interface 150 may include, for example, a network interface card (NIC).

[0037] Memory 151 is composed of, for example, a combination of volatile memory and non-volatile memory. Memory 151 is used to store programs executed by the processor 152, data used for processing, etc. Memory 151 may include multiple memories. Furthermore, the model storage unit 104 described above may be implemented by a storage device such as memory 151.

[0038] The processor 152 performs processing in the information processing device 100 by reading and executing a program from the memory 151. The processing of each component shown in Figure 2 is performed, for example, by the processor 152 reading and executing a program from the memory 151. The processor 152 may be, for example, a microprocessor, an MPU (Micro Processor Unit), a CPU (Central Processing Unit), or a GPU (Graphics Processing Unit). The processor 152 may include multiple processors. In this way, the information processing device 100 has the functionality of a computer.

[0039] Furthermore, each mobile unit 200 may have the configuration shown in Figure 4. That is, the mobile unit 200 may also have computer functions. In this case, the mobile unit 200 may perform processing by having the processor read a program from memory and execute it.

[0040] The embodiments have been described above. According to the group control system 10, the behavior of each mobile body 200 is controlled based on first state information including the current acceleration data and position of each mobile body 200, and second state information representing the predicted future state of each mobile body 200. Therefore, since the behavior is controlled by considering various state information of each mobile body 200, the behavior of the mobile bodies 200 can be controlled more appropriately than when these are not considered. Therefore, multiple mobile bodies 200 can be controlled appropriately. In particular, the behavior of each mobile body 200 is controlled using a model in which deep reinforcement learning has been performed to reduce the difference between the position grasped by the group control device and the actual position of each mobile body 200. By making each mobile body 200 act according to such a model, the discrepancy between the position grasped by the group control device (information processing device 100) and the actual position of each mobile body 200 can be suppressed, and the positions of multiple mobile bodies 200 can be appropriately managed. As a result, the mobile bodies 200 can be operated appropriately, especially in a small area.

[0041] It should be noted that the present invention is not limited to the embodiments described above, and can be modified as appropriate without departing from the spirit of the invention. For example, at least a part of the configuration of the information processing device 100 (for example, the model learning unit 105) may be mounted on a device other than the information processing device 100. [Explanation of symbols]

[0042] 10 Group Control System 100 Information Processing Devices 101 Mobile object information acquisition unit 102 Position estimation part 103 Mobile Unit Control Unit 104 Model Memory Unit 105 Model Learning Department 150 network interfaces 151 memory 152 processors 200 Mobile Units 201 IMU 202 LiDAR sensor 203 Position detection unit 204 Driving Control Unit 205 Communication Processing Unit M Machine Learning Model M1 Deep Reinforcement Learning Model M2 Generative Adversarial Network

Claims

1. A group control system for controlling multiple mobile objects capable of autonomously traveling within a predetermined area, An acceleration data acquisition unit acquires acceleration data detected by sensors provided on the moving body for each of the moving bodies. A position estimation unit that estimates the current position of each of the aforementioned moving objects, A mobile body control unit inputs first state information, which includes the acceleration data and estimated position of each mobile body, and second state information, which represents the future state of each mobile body, into a deep reinforcement learning model, and repeatedly executes a process to control the behavior of each mobile body based on the output from the deep reinforcement learning model. It has, The second state information input to the deep reinforcement learning model represents the future state of each mobile entity predicted using the output obtained by inputting seed information into a generative adversarial network. The aforementioned seed information includes information output from the deep reinforcement learning model in the previous processing. Group control system.

2. The deep reinforcement learning model is a model in which deep reinforcement learning is performed in an environment in which rewards are given based on the difference between the position of each moving object estimated by the position estimation unit and the actual position of each moving object. The group control system according to claim 1.

3. The second state information includes the position and acceleration data of each moving object at a future point in time. The group control system according to claim 1 or 2.

4. The adversarial generative network outputs multiple images showing the position of each of the moving objects at different points in the future. The second state information includes acceleration data for a future time point calculated from the positions of each moving object at different time points shown in the plurality of images. The group control system according to claim 3.

5. A group control method for controlling multiple mobile bodies capable of autonomously traveling within a predetermined area, Acceleration data detected by sensors provided on the moving body is acquired for each of the moving bodies. The current position of each of the aforementioned moving objects is estimated, The first state information, which includes the acceleration data and estimated position of each of the aforementioned moving bodies, and the second state information, which represents the future state of each of the aforementioned moving bodies, are input to a deep reinforcement learning model, and a process to control the behavior of each of the aforementioned moving bodies is repeatedly executed based on the output from the deep reinforcement learning model. The second state information input to the deep reinforcement learning model represents the future state of each mobile entity predicted using the output obtained by inputting seed information into a generative adversarial network. The aforementioned seed information includes information output from the deep reinforcement learning model in the previous processing. Group control method.

Images