Automated guided vehicle and method of training an automated guided vehicle

Abstract

An automated guided vehicle and a method of training an automated guided vehicle, the automated guided vehicle comprising: a mobile base comprising a powertrain to drive the automated guided vehicle in an autonomous navigation mode within a facility; a camera to capture image data of objects within the facility and comprising a content filter; and a primary control module to receive the image data and further to: execute a recognition neural network program to recognize targets in the image data; execute, under user control, a supervised program to train the recognition neural network program, the supervised program to receive the image data and recognize optical labels attached to the targets in the image data, generate a supervised result, the optical labels recognized by the content filter of the camera, and the targets with the optical labels classified into a first category; and use the supervised result to adjust at least one weight of at least one node of the recognition neural network program. The automated guided vehicle can improve the ability to recognize people or other targets.

Description

[0001] This application is a divisional application of the invention patent application filed on November 15, 2018, with application number 201880038567.4 and entitled "Real-time supervised machine learning system and method for on-site environment". Technical Field

[0002] This application relates to automated guided vehicles, and in particular to automated guided vehicles capable of performing supervised learning in real time. Background Technology

[0003] An automated guided vehicle (AGV) is a mobile robot capable of autonomous navigation using methods such as tags, wires, metal strips on the ground, camera images, magnets, lasers, or any combination thereof. AGVs are commonly used in industrial applications, such as transporting materials in factories or warehouses. In recent years, the number of AGVs and the variety of their applications have increased significantly.

[0004] Automated guided vehicles (AGVs) operating in warehouses typically rely on computer vision (CV) to identify a target (such as a vehicle or person) and its environment (such as a warehouse). In current practice, the operating system or similar control program of an AGV is trained by the AGV's manufacturer before delivery to the customer's facility. That is, the AGV's manufacturer trains the AGV's neural network to identify targets using a default dataset (or a pre-defined, pre-labeled dataset), such as training the AGV's control program to identify people, objects, environments, etc.

[0005] However, in practical applications, the performance of the trained neural network of the automated guided vehicle (AGV) often falls short of expectations, especially when the actual (or real-time) environment differs significantly from the default dataset used for training. Indeed, the default training dataset often contains factors that differ significantly from the actual environment, such as warehouse lighting conditions, staff uniforms (or other clothing), and the appearance of shelves. Such differences can lead to low recognition rates and poor user experiences. However, for AGV manufacturers, recording the actual facility status and creating new datasets for each customer to adapt to various environments is extremely difficult.

[0006] Therefore, there is a need in the art to improve automated guided vehicle (AGV) systems, and in particular, an AGV system that can perform training in a real-world environment after the AGV is delivered to a customer's facility. Summary of the Invention

[0007] To address the shortcomings of the aforementioned prior art, the main objective of this invention is to provide an automated guided vehicle (AGV) comprising: (i) a mobile base including a powertrain for driving the AGV within a facility in an autonomous navigation mode; (ii) a first camera for acquiring first image data of objects within the facility; (iii) a second camera for acquiring second image data of objects within the facility, wherein the second camera includes a content filter; and (iv) a main control module for receiving the first image data and the second image data transmitted from the first camera and the second camera, respectively. The main control module is further configured to execute a recognition neural network program. The neural network program is used to identify targets in the first image data. The main control module is also configured to execute a supervision program under user control. The supervision program receives the second image data and identifies tags attached to targets in the second image data. The supervision program generates a supervision result, wherein a first target with a first tag is classified into a first category, and the supervision program uses the supervision result to adjust at least one weight of at least one node in the recognition neural network program.

[0008] In one embodiment, the tag is identified by the content filter.

[0009] In another embodiment, the content filter is a color filter used to identify a unique color associated with a label.

[0010] In another embodiment, the content filter is used to identify a unique pattern associated with a tag.

[0011] In another embodiment, the monitoring program classifies the first target with the first label into the first category according to the instructions input by the user, and thereby generates a monitoring result.

[0012] In another embodiment, the first category includes "person recognition".

[0013] In another embodiment, the first category includes "the target being followed".

[0014] In another embodiment, the supervisory procedure updates a default dataset associated with the recognition neural network procedure with the supervision results and relevant real-world information.

[0015] In one embodiment, the real-world information is related to the clothing of people in the first and second image data.

[0016] In another embodiment, the real-world information is associated with the underlying physical structure of the facility in the first and second image data.

[0017] Another primary objective of the present invention is to provide a method for training an automated guided vehicle (AGV) comprising a mobile base having a powertrain and used to drive the AAV in an autonomous navigation mode within a facility. The method comprises: (i) acquiring first image data of objects within the facility using a first camera; (ii) acquiring second image data of objects within the facility using a second camera, wherein the second camera includes a content filter; (iii) receiving first image data and second image data transmitted from the first camera and the second camera respectively using a main control module; (iv) executing a recognition neural network program using the main control module and pre-labeled image data stored in a default dataset, the recognition neural network program being able to operate on and identify targets in the first image data using the pre-labeled image data; (v) executing a supervision program using the main control module under user control, the supervision program being used to receive the second image data and identify labels attached to targets in the second image data, the supervision program being used to generate a supervision result, wherein a first target with a first label is classified into a first category; and (vi) using the supervision result to adjust at least one weight of at least one node in the recognition neural network program.

[0018] Before describing specific embodiments of the present invention, certain words and phrases used in this patent document should be defined here: the terms "comprising" and "including" and similar phrases mean "including but not limited to"; the term "or" is inclusive, meaning "and / or"; the terms "associated with" and "related to" and similar phrases may mean "including, contained in, interconnected with, contained, included in, connected to, coupled with, intercommunication with, cooperate with, intertwined with, juxtaposed with, adjacent to, constrained by, or having, possessing the properties of something", etc.; the term "controller" means any device, system, or part thereof capable of controlling at least one action, such a device may be implemented in hardware, firmware, or software, or some combination of at least two of them. It is worth noting that the functions associated with any particular controller may be local or remote, centralized or distributed. Definitions of certain words and phrases are provided in this patent document, and those skilled in the art should understand that in many or most cases, such definitions also apply to the words and phrases defined in the preceding and subsequent instances. Attached Figure Description

[0019] To help understand the present invention and its effects, please refer to the following description and corresponding drawings, wherein the same reference numerals denote the same parts.

[0020] Figure 1 This is a schematic diagram of an automated guided vehicle in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of an image in a default dataset and a real-time, real-world environment, as described in an embodiment of the present invention.

[0022] Figure 3 This is a schematic diagram of an operator or user control screen for an automated guided vehicle system, as described in an embodiment of the present invention.

[0023] Figure 4 This is a detailed schematic diagram of a specific subsystem of an automated guided vehicle system in an embodiment of the present invention.

[0024] Figure 5 This is an example of real-time supervised learning based on the principles of this invention.

[0025] Figure 6 This is a flowchart of a real-time supervised learning method in an embodiment of the present invention.

[0026] Figure 7 This is an example of real-time supervised learning based on the principles of this invention, using "person recognition".

[0027] Figure 8 This is an example of real-time supervised learning based on the principles of this invention, using "person following recognition".

[0028] Figure 9 This is an example of using a camera to perform real-time supervised learning based on the principles of the present invention. Detailed Implementation

[0029] The following is related Figures 1 to 9 The descriptions and embodiments used to illustrate the principles of the invention in this patent document are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art will understand that the principles of the invention can be implemented by any suitable automated guided vehicle system.

[0030] This disclosure describes an automated guided vehicle (AGV) system that enables the AGV's neural network program to self-improve when the AGV enters a new environment inconsistent with the neural network program's default dataset. This is achieved through a real-time supervised process that updates the default dataset. The user (or operator) of the AGV initiates a network training mode for the AGV to perform supervised learning. This supervised process relies on optical tags selected by the user, which can be identified by a trained computer vision camera on the AGV to update specific functions (e.g., person recognition, person following, autonomous navigation, obstacle avoidance, etc.).

[0031] The automated guided vehicle (AGV) uses a first camera to provide visual input data to its recognition neural network and a second camera with an optical filter for use in supervised learning. The second camera can identify optical labels (e.g., bright neon pink or green stickers) in the captured images. This allows the supervision process to monitor the results of a function (e.g., person recognition), enabling the weights of the neural network's nodes to be adjusted based on the supervision results. In this way, the default dataset can be improved into an updated dataset that better matches the real-time environment in which the AGV operates.

[0032] Figure 1 This is a schematic diagram of an Automated Guided Vehicle (AGV) 100 according to an embodiment of the present invention. The AGV 100 includes a self-driving mobile base 105 and a vertical control console, including a lower control console 115 and an upper control console 116. The lower control console 115 includes a barcode scanning unit 120 and side cameras 145A and 145B. The upper control console 116 includes a touch screen 125, a main camera 140, and a wireless communication unit 130. In a preferred embodiment, the main camera 140 may include a set of cameras, including camera 140A and camera 140B. Camera 140B may include a color filter for training the recognition neural network, which assists in controlling the operation of the AGV 100. Cameras 140A and 140B face forward, and side cameras 145A and 145B face the left and right sides of the AGV 100, respectively. The main cameras 140A and 140B, together with the side cameras 145A and 145B, provide a 270-degree field of view. Alternatively, a rear-facing camera (not shown) can be added to the rear of the upper part of the console 116 or the movable base 105 to provide a 360-degree field of view.

[0033] The mobile base 105 includes multiple guardrails 110, multiple ultrasonic sensors including ultrasonic sensors 150A-150F, multiple laser sensors including a LiDAR sensor 160, a main control unit 170, an electromechanical module 175, drive wheels 180, and stabilizing wheels 185. The guardrails 110 are located around the upper surface of the mobile base 105. When the automated guided vehicle 100 moves within a warehouse or other facility, the guardrails 110 prevent inventory items that are pushed or jostled during movement from sliding off the upper surface of the mobile base 105.

[0034] The touchscreen 125 is used to display information and allow the user (or operator) to control the automated guided vehicle 100. The touchscreen 125 is just one type of user interface. Alternatively, the automated guided vehicle 100 can include any suitable user input device to provide instructions, maps of the facility, route information, inventory information, or inventory platforms, etc. This allows the user to control the automated guided vehicle 100 both automatically and manually. This is particularly useful in network training mode; in network training mode, supervised learning can be used to update and improve the default dataset of the automated guided vehicle 100, as detailed below.

[0035] The main control unit 170 controls the overall operation of the automated guided vehicle 100. The main control unit 170 receives distance information about surrounding objects from ultrasonic sensors 150A-150F, lidar sensor 160, and front-facing main cameras 140A, 140B and side cameras 145A, 145B. In response to the received distance information, the main control unit 170 guides the automated guided vehicle 100 to move within its location within the factory or warehouse.

[0036] The mobile base 105 includes multiple drive wheels 180 and multiple stabilizing wheels 185 controlled by a main control unit 170 for moving and steering the automated guided vehicle 100. In one exemplary embodiment, the motor module 175 includes one or more electric motors controlled by the main control module 170 to rotate the drive wheels 185. The motor module 175 may also include a lifting system for adjusting the height of the mobile base 105 to load or unload inventory items.

[0037] The mobile base 105 may include two drive wheels 180 and four stabilizing wheels 185. The stabilizing wheels 185 may be casters located at the four corners of the mobile base 105. The drive wheels 180 may be located below the mobile base 105, between the front and rear stabilizing wheels 185. Each drive wheel 180 is operated by a motor controlled by the main control module 175 and rotates in any specific direction and at a variable speed. For example, the drive wheels 180 may rotate and move the automated guided vehicle 100 forward, backward, and laterally in the XY plane on the ground. The drive wheels 180 may be controlled to rotate at different speeds and in different directions to steer or rotate the automated guided vehicle 100.

[0038] Based on the principles of the present invention, the main control module 175 may include a recognition neural network that can analyze visual data collected by, for example, cameras 140A and 140B, and use the visual data to perform specific functions, such as a person recognition function, a person following function, an autonomous navigation function, an obstacle avoidance function, and other routine functions of existing automated guided vehicle systems.

[0039] As is well known, a neural network contains multiple layers of computational nodes, or neurons, with different layers interconnected. A neural network transforms data until the data can be categorized into an output result. Each node multiplies an initial value by a weight, adds the output result to other values ​​entering the same node, adjusts the result value by a bias chosen by the node, and then standardizes the output result using an activation function.

[0040] A neural network employs an iterative learning process where records (rows) are presented to the network individually, and the weights associated with the input values ​​are adjusted sequentially. After the records are presented, the process is typically repeated (or executed iteratively). During this learning phase, the neural network trains itself to correctly predict the class labels of input samples by adjusting the weights. The advantage of a neural network is its high tolerance for noisy data and its ability to classify patterns that have not yet been trained on. The most popular neural network algorithm is the backpropagation algorithm. Once a neural network has been built for a specific application, it is considered trained. To begin the process, initial weights can be randomly selected before starting the training or learning process.

[0041] The network uses the weights and the functionality of the hidden layers located in the neural nodes to process the records in the default dataset (or training set) one by one, and then compares the output with the expected output. Errors are propagated back through the neural network, and the weights are adjusted accordingly and used for the next record. This process is repeated to correct the weights of the nodes. During training, the same set of data will be processed multiple times due to the continuous adjustment of connection weights.

[0042] Figure 2 This is a schematic diagram of images from a default dataset and a real-time, real-world environment in an embodiment of the present invention. Image 211 represents the interior of a warehouse in the default dataset 210. Image 212 represents staff in the default dataset 210. Image 221 is a captured image showing the actual warehouse in the real-world environment 220. Image 222 is a captured image showing the actual staff in the real-world environment 220.

[0043] The real-world warehouse in image 221 and the real-world workers in image 222 are significantly different from the default warehouse in image 211 and the default workers in image 212 used to train the recognition neural network of the automated guided vehicle 100. Therefore, the image recognition neural network of the automated guided vehicle 100 cannot recognize real-world shelves and storage racks due to the different structural types of shelves and storage racks, and it also cannot recognize real-world workers due to the different clothing or uniforms.

[0044] Figure 3 This is a schematic diagram of an operator or user control panel on a touchscreen 125 of an automated guided vehicle (AGV) system, as described in an embodiment of the present invention. The control panel includes multiple menu options, including navigation mode 311, human-following mode 312, manual control mode 313, and network training mode 314. By selecting network training mode 314, the user of the AGV 100 can initiate a supervised program to train the recognition neural network to identify new images and update the default dataset to better reflect the real-world environment.

[0045] Figure 4This is a detailed schematic diagram of a specific subsystem of an automated guided vehicle (AGV) system 100 in an embodiment of the present invention. The specific subsystem includes, but is not limited to, some components of the main control module 170 and some components of the motion module 175. A data bus 499 provides a communication path between the touchscreen 125, camera 145, processor 405, memory 410, distance sensor 420, contact sensor 425, storage 430, lifting system 440, and powertrain 435. The lifting system 440 and powertrain 435 may be components within the motion module 175. The processor 405, memory 410, and storage 435 may be components within the main control module 170.

[0046] Memory 410 may include a combination of dynamic random access memory (DRAM) and static random access memory (SRAM). Memory 410 contains a dataset 411, a supervision program 412, and a recognition neural network program 415. The dataset 411, supervision program 412, and recognition neural network program 415 may typically be stored in storage 430, which may be, for example, flash RAM or an optical disc. Processor 405 loads the dataset 411, supervision program 412, and recognition neural network program 415 into memory 410 upon power-on. According to the principles of the invention, processor 405 executes neural network program 413 and supervision program 412 to train the recognition neural network in the real-world environment.

[0047] Figure 5 This is an example of real-time supervised learning based on the principles of the present invention. An image 505 of a real-world environment of a customer facility contains various objects (e.g., people, shelves, storage racks, inventory items), which may not be recognizable by the neural network program 413, such as people wearing special safety vests and helmets. For example, the neural network program 413 trained on the default dataset 411 may be unable to identify workers because the default dataset 411 lacks images of workers wearing vests and helmets.

[0048] According to the principles of this invention, this problem can be solved by adding special optical tags to the people and helmets in the image 505, and using a supervisory program 412 to train the neural network program 413 to identify people with the optical tags. For example, optical tag 511, which may contain a special coating or special color, can be affixed to the shirt of the person on the left side of the image 505. Similarly, optical tags 512 and 513 can be affixed to the back of the vest of the person on the right side of the image 505.

[0049] Camera 140A captures images input to neural network program 413. Camera 140B, using an optical filter capable of sensing tags 511-513, captures images input to supervision program 412. The other camera with the optical filter has a special coating used to identify targets (such as people, shelves, inventory items, etc.) to supervise the output of neural network program 413 and update the weights of the nodes in neural network program 413. The user or operator of the automated guided vehicle 100 can use supervision program 412 to specify targets marked with tags 511-513 as specific categories, such as identified people, people to be followed, obstacles, etc.

[0050] The supervisory process 412 then generates a supervisory result in which the targets labeled 511-513 have been identified as objects associated with a specific category. Block 520, which may be a processing step performed by the supervisory process 412, compares the supervisory result with the "normal" output of the neural network program 413 generated from the default dataset 411. As shown by the dashed line in the diagram from the output of block 520 to the node of the neural network program 413, the supervisory process 412 adjusts the weights of each node in the neural network program 413 based on the comparison and the supervisory result. In this way, the neural network program 413 can be trained using real-world images that can improve and update the default dataset 411.

[0051] Figure 6 This is a flowchart 600 of a real-time supervised learning exercise according to an embodiment of the present invention. In step 605, the user of the automated guided vehicle 100 affixes one or more visual tags to one or more specific targets. In step 610, the user initiates a network training mode from the touchscreen 125 of the automated guided vehicle 100. In step 615, the user selects the configuration settings for training. For example, the user can select the category of the training target: (i) people; (ii) shelves; (iii) inventory items. The user can also select the category of visual features to be identified, such as a special coating, pattern, or other optical label.

[0052] In step 620, the supervisory program 412 executes the selected training method and generates a supervision result. In step 625, the nodes of the neural network are adjusted according to the supervision result to correct the weights of the neural nodes. Next, in step 630, this process can be repeated for other target categories. In this way, the dataset 411 can be repeatedly updated with images of people, shelves, inventory items, etc., corresponding to the real-world environment as new targets for identification.

[0053] Figure 7 In this embodiment of the invention, "person recognition" is used as an example of real-time supervised learning based on the principles of the invention. Figure 7 As shown, it is assumed that the neural network program 413 has been trained to only recognize people wearing casual clothes, such as suits, casual pants, formal shirts, jackets, women's shirts, skirts, etc. It is further assumed that the neural network program 413 has not yet been trained to recognize people wearing helmets.

[0054] However, in image 705, the real-world environment of the customer facility includes people in regular clothing (such as the two people in the middle) and people wearing special safety vests and helmets (such as the people on the left and right). Therefore, the neural network program 413 trained using the default dataset 411 may be able to identify the two people in the middle of image 705, but cannot identify the workers wearing safety vests and helmets on the left and right sides, because these images are missing from the default dataset 411.

[0055] as Figure 5 As shown, this problem can be corrected by adding special optical tags 711-714 to the clothing and helmets of the people in image 705, and using a supervision program 412 to train the neural network program 413 to recognize people with the optical tags 711-714. The supervision program 412 generates a supervision result in which targets marked with optical tags 711-714 are considered successfully identified objects of a specific category. Block 720, which can be a processing step performed by the supervision program 412, compares the supervision result with the "normal" output of the neural network program 413 generated from the default dataset 411. As shown by the dashed line from the output of process block 720 to the node of the neural network program 413 in the figure, the supervision program 412 adjusts the weights of each node in the neural network program 413 based on the comparison and the supervision result. In this way, the neural network program 413 is trained by the supervision program 412 to improve and update the default dataset 411 using real-world images.

[0056] Figure 8In one embodiment of the present invention, "person following recognition" is used as an example of real-time supervised learning based on the principles of the present invention. Figure 8 Image 805 represents the real-world environment inside a warehouse. In this case, the desired outcome is that the automated guided vehicle 100 is able to follow a person wearing a jacket and a helmet on the right, rather than a person wearing a vest and a helmet on the left.

[0057] As described above, the automated guided vehicle 100 can accomplish this task by adding special optical tags 811-813 to the clothing and helmet of the person on the right side of image 805, and using a supervision program 412 to train the neural network program 413 to recognize the person with the optical tags 811-813. The supervision program 412 generates a supervision result in which the target marked with optical tags 811-813 is considered a successfully identified object and belongs to the category of "followed target". Block 820, which may be a processing step performed by the supervision program 412, compares the supervision result with the "normal" output of the neural network program 413 generated from the default dataset 411. As shown by the dashed line from the output of process block 820 to the node of the neural network program 413 in the figure, the supervision program 412 adjusts the weights of each node in the neural network program 413 based on the comparison and the supervision result. In this way, the neural network program 413 can be trained by the supervisory program 412 using real-world images that can improve and update the default dataset 411. After training, the neural network program 413 can learn which features are essential for recognizing the target to be followed. Without the training method disclosed herein, the neural network program 413 would easily misjudge and follow the wrong person.

[0058] Figure 9 This is an example of performing real-time supervised learning using only one camera, according to another embodiment of the invention. Figure 9 In this example, color filters or similar content filters are not used on a second camera. Instead, the captured images are provided to the neural network program 413 and the supervisory program 412 via the same camera 140. However, in this case, two-dimensional barcode (QR code) labels 911 and 912 are attached to targets in image 905. The supervisory program 412 is configured to recognize the two-dimensional barcode labels 911 and 912, and thus can identify the targets. On the other hand, the untrained neural network 413 is configured to ignore the two-dimensional barcode labels 911 and 912 and complete the recognition process using the rest of image 905.

[0059] The supervision process 412 generates a supervision result in which the target marked with two-dimensional barcode labels 911, 912 is considered a successfully identified object and belongs to a specific category, such as "person" or "followed target". Block 920, which may be a processing step performed by the supervision process 412, compares the supervision result with the "normal" output of the neural network program 413 generated from the default dataset 411. As shown by the dashed line in the figure from the output of process block 920 to the node of the neural network program 413, the supervision process 412 adjusts the weights of each node in the neural network program 413 based on the comparison and the supervision result.

[0060] The methods and apparatus disclosed in this invention enhance the ability of the automated guided vehicle (AGV) 100's camera to recognize people and better identify what kind of person (or other target) needs to be followed. The disclosed methods and apparatus also enable the camera to better identify and distinguish obstacles or people. If a person obstructs the path, the AGV 100 is more likely to recognize the person and slow down or stop for safety reasons. The disclosed methods and apparatus also enable the camera to better identify shelves and labels on shelves, allowing the AGV 100 to locate itself.

[0061] In the foregoing description, the disclosed system and the method of performing machine learning in a field environment are illustrated using a warehouse transport vehicle as an example. However, this presentation is for reference only and is not intended to limit the invention and claims. More broadly, the automated guided vehicle (AGV) can be any autonomous moving vehicle using vision sensors or a self-driving trolley, box, suitcase, etc. For example, other embodiments of the AGV 100 may include motorized self-driving luggage compartments, motorized self-driving shopping carts, etc.

[0062] Vehicles, robotic lawnmowers, unmanned aerial vehicles (UAVs), unmanned underwater vehicles, unmanned cars, autonomous forklifts, transport robots, and so on.

[0063] Although the present invention has been described by way of example of one embodiment, various modifications and variations can be made to the invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automated guided vehicle, characterized in that, Include: A mobile base comprising a powertrain for driving the automated guided vehicle within a facility in an autonomous navigation mode; A camera for capturing image data of objects within the facility, wherein the camera includes a content filter; and A main control module is used to receive image data transmitted from the camera, wherein the main control module is further used to: Execute a recognition neural network program to identify targets in the image data; A supervisory program for training the recognition neural network is executed under user control. The supervisory program receives the image data and identifies optical tags attached to the target in the image data, generating a supervisory result, wherein the optical tags are identified by the content filter of the camera, and the target with the optical tags is classified into a first category. as well as The supervision results are used to adjust at least one weight of at least one node in the recognition neural network program.

2. The automated guided vehicle as described in claim 1, characterized in that, The content filter is a color filter used to identify a unique color associated with an optical label.

3. The automated guided vehicle as described in claim 1, characterized in that, The content filter is used to identify a unique pattern associated with an optical label.

4. The automated guided vehicle as described in claim 1, characterized in that, The monitoring program classifies the target with the optical tag into the first category according to the instructions input by the user, and generates a monitoring result accordingly.

5. The automated guided vehicle as described in claim 4, characterized in that, The first category includes person recognition.

6. The automated guided vehicle as described in claim 4, characterized in that, The first category contains the targets being followed.

7. The automated guided vehicle as described in claim 1, characterized in that, The supervision procedure updates a default dataset associated with the recognition neural network program with the supervision results and relevant real-world information.

8. The automated guided vehicle as described in claim 7, characterized in that, The real-world information is related to the clothing of the people in the image data.

9. The automated guided vehicle as described in claim 7, characterized in that, The real-world information is related to the underlying physical structure of the facility in the image data.

10. A method for training an automated guided vehicle (AGV), the AAV comprising a mobile base having a powertrain for driving the AAV in an autonomous navigation mode within a facility, the method characterized in that it comprises: A camera is used to capture image data of objects within the facility, wherein the camera includes a content filter; A main control module is used to receive image data transmitted from the camera; The main control module is used to execute a recognition neural network program to identify targets in the image data; Under user control, the main control module executes a supervisory program to train the recognition neural network. This supervisory program receives the image data, identifies optical tags attached to the target within the image data, and generates a supervisory result, wherein the optical tags are identified by the camera's content filter. And the target with the optical tag will be classified into a first category; and The supervision results are used to adjust at least one weight of at least one node in the recognition neural network program.

11. The method as described in claim 10, characterized in that, The content filter is a color filter used to identify a unique color associated with an optical label.

12. The method as described in claim 10, characterized in that, The content filter is used to identify a unique pattern associated with an optical label.

13. The method as described in claim 10, characterized in that, The monitoring program classifies the target with the optical tag into the first category according to the instructions input by the user, and generates a monitoring result accordingly.

14. The method as described in claim 13, characterized in that, The first category includes person recognition.

15. The method as described in claim 13, characterized in that, The first category contains the targets being followed.

16. The method as described in claim 10, characterized in that, The supervision procedure updates a default dataset associated with the recognition neural network program with the supervision results and relevant real-world information.

17. The method as described in claim 16, characterized in that, The real-world information is related to the clothing of the people in the image data.

18. The method as described in claim 16, characterized in that, The real-world information is related to the underlying physical structure of the facility in the image data.

Images