Multi-logistics-robot system based on edge training and training and distribution method

Abstract

The invention provides a multi-logistics-robot system based on edge training and a training and distribution method, and belongs to the technical field of logistics robots, and the system comprises logistics robots and an operation center. The distribution method comprises the steps that the logistics robot completes a specified distribution task according to scheduling of an operation center; the logistics robot can automatically form and record operation process data and upload the operation process data to the operation center in real time. The training method comprises the steps of determining targets covered in an actual application scene according to operation process data of each logistics robot; a classification trainer is adopted to classify targets of a single type; calling at least one classification trainer to execute knowledge distillation training on the logistics robot; and according to an order dispatching demand of logistics distribution, a logistics robot trained for an actual application scene is deployed, and the corresponding order dispatching demand is responded. The accuracy of logistics robot training is improved, and the training efficiency is improved.

Description

Technical Field

[0001] The present application belongs to the technical field of logistics robots, and specifically relates to a multi-logistics robot system, training and distribution method based on edge training. Background Art

[0002] Logistics robots are usually used in closed office buildings, parks, communities, campuses, etc. that are inconvenient for deliverymen and couriers to enter, to achieve logistics delivery of parcels, letters, food and other items.

[0003] In the existing technology, logistics robots need to have the ability to classify and recognize targets, so that in the process of logistics distribution, they can autonomously identify the surrounding targets as: delivery items, human bodies, vehicles, obstacles, building structures, traffic markings and signs, paths, etc. The target classification and recognition module of logistics robots generally has a multi-layer convolutional neural network and a classifier that outputs category probability distribution. Then, according to the results of target classification and recognition, the logistics robot performs adaptive operations for different target categories, such as approaching, stopping, avoiding, detouring, accelerating and decelerating, etc.

[0004] The target classification and recognition module of the logistics robot should be pre-trained using the training set before being put into use. However, in the actual scenarios faced by the logistics robot, the composition and form of various targets are complex and the changes are dynamic, so continuous enhanced training is required to improve the accuracy and response speed of target classification and recognition.

[0005] For enhanced training, since the scenarios covered by the deployment locations of each batch of logistics robots and even the scenarios covered by the distribution tasks assigned to each robot are quite different and diverse, the training sets provided and updated by the manufacturer are also difficult to fully cover these scenarios. Therefore, there is a bottleneck in using the training set to enhance the robot's pertinence for actual scenarios. On the other hand, the use of training sets also causes logistics robots to be trained on some scenarios that will not be encountered or are rare in actual applications, resulting in poor training pertinence and low efficiency. Summary of the invention

[0006] Based on the above technical problems, this application proposes a multi-logistics robot system, training and distribution method based on edge training.

[0007] In a first aspect, the present application proposes a multi-logistics robot system based on edge training, comprising: a logistics robot and an operation center in the cloud; the logistics robot communicates with the operation center in real time;

[0008] The logistics robot is used to realize the autonomous execution of all operation links in the logistics distribution process, integrates various types of sensors and intelligent controllers, can automatically generate and record operation process data, and upload the operation process data to the operation center in real time;

[0009] The operation center is used to classify the targets covered in the actual application scenarios according to the operation process data of each logistics robot, perform knowledge distillation training on the logistics robots based on the classification trainers corresponding to different classifications, and call the trained logistics robots so that the logistics robots can complete the designated delivery tasks.

[0010] The operation center includes: a scene analysis module, a classification training combination module, a training coordination module and a robot scheduling module;

[0011] The scenario analysis module is used to determine the actual application scenario of the logistics robot according to the operation process data of each logistics robot, and then determine the targets covered in the actual application scenario, and classify the targets;

[0012] The classification training combination module includes a plurality of classification trainers, each of which can classify a single type of target, and each of which has a training sample set corresponding to the target of this type;

[0013] The training coordination module is used to call at least one classification trainer in the classification training combination module to perform knowledge distillation training on the logistics robot;

[0014] The robot scheduling module is used to determine the actual application scenarios corresponding to the order dispatching requirements of logistics distribution, and then deploy logistics robots trained for the actual application scenarios to respond to the corresponding order dispatching requirements.

[0015] The operation process data includes: path records, surrounding scene target data during the logistics distribution process, and action execution data.

[0016] The classification trainer is equivalent to the teacher network in knowledge distillation, and the logistics robot is equivalent to the student network in knowledge distillation;

[0017] The training coordination module can call the corresponding classification trainer for each logistics robot's actual application scenario and perform knowledge distillation training on the logistics robot.

[0018] The training coordination module can also call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, and other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group;

[0019] The training coordination module can form an edge training group of logistics robots with the same actual application scenarios according to the actual application scenarios of logistics robots, use logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios as teacher networks in knowledge distillation, and use scene target data recorded when performing previous logistics delivery tasks as training sample sets; and use other logistics robots as student networks in knowledge distillation.

[0020] The logistics robot is connected to the wireless broadband Internet of Things, has uplink and downlink network communication capabilities, and can maintain real-time communication with the cloud-based operation center. During the entire operation process, the logistics robot keeps uploading the operation process data in real time.

[0021] In the second aspect, the present application proposes a multi-logistics robot training method based on edge training, comprising the following steps:

[0022] According to the operation process data of each logistics robot, determine the actual application scenario of the logistics robot, and then determine the goals covered in the actual application scenario;

[0023] Use a classification trainer to classify a single type of target;

[0024] Call at least one classification trainer to perform knowledge distillation training on the logistics robot;

[0025] According to the order dispatching requirements of logistics distribution, the actual application scenarios corresponding to the order dispatching requirements are determined, and then logistics robots trained for the actual application scenarios are deployed to respond to the corresponding order dispatching requirements.

[0026] The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot comprises the following steps:

[0027] The classification trainer is used as the teacher network in knowledge distillation, and the logistics robot is used as the student network in knowledge distillation;

[0028] For each target covered in the actual application scenario of the logistics robot, the corresponding classification trainer is called to perform knowledge distillation training on the logistics robot.

[0029] The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot also includes the following steps:

[0030] Call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, while other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group;

[0031] According to the actual application scenarios of logistics robots, logistics robots with the same actual application scenarios are formed into an edge training group, and logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios are used as teacher networks in knowledge distillation, and the scene target data recorded when performing previous logistics delivery tasks are used as training sample sets; other logistics robots are used as student networks in knowledge distillation.

[0032] In a third aspect, the present application proposes a multi-logistics robot delivery method based on edge training, comprising the following steps:

[0033] According to the dispatch of the operation center, the logistics robot completes the designated delivery tasks;

[0034] Logistics robots can automatically generate and record operation process data, and upload the operation process data to the operation center in real time.

[0035] Beneficial technical effects:

[0036] This application proposes a multi-logistics robot system, training and distribution method based on edge training, which breaks through the technical bottleneck of traditional methods. Based on the classification trainers corresponding to different classifications, knowledge distillation training is performed on the logistics robot. Not only can the classification trainer in the operation center be used as the teacher network in knowledge distillation, and the logistics robot can be used as the student network in knowledge distillation; for the targets covered in the actual application scenarios of each logistics robot, the corresponding classification trainer is called to perform knowledge distillation training on the logistics robot. It is also possible to call two or more logistics robots to form an edge training group, in which some logistics robots serve as the teacher network in knowledge distillation, and other logistics robots are equivalent to the student network in knowledge distillation, thereby realizing mutual training of logistics robots in the edge training group. This application improves the accuracy of logistics robot training and improves the efficiency of training. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 This is a principle block diagram of a multi-logistics robot system based on edge training according to an embodiment of the present application;

[0038] Figure 2 This is a flow chart of a multi-logistics robot system training method based on edge training in an embodiment of the present application;

[0039] Figure 3 This is a flow chart of a multi-logistics robot system distribution method based on edge training according to an embodiment of the present application. DETAILED DESCRIPTION

[0040] The present application is further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present application.

[0041] In the first aspect, the present application proposes a multi-logistics robot system based on edge training, such as Figure 1 As shown, including:

[0042] A logistics robot and an operation center in the cloud; the logistics robot communicates with the operation center in real time;

[0043] The logistics robot is used to realize the autonomous execution of all operation links in the logistics distribution process, integrates various types of sensors and intelligent controllers, can automatically generate and record operation process data, and upload the operation process data to the operation center in real time;

[0044] The operation center is used to classify the targets covered in the actual application scenarios according to the operation process data of each logistics robot, perform knowledge distillation training on the logistics robots based on the classification trainers corresponding to different classifications, and call the trained logistics robots so that the logistics robots can complete the designated delivery tasks.

[0045] The logistics robot itself is an intelligent device with the necessary hardware and software configurations. The operations it performs autonomously include: task acceptance, item inspection, path planning, lidar surrounding scene perception, navigation and obstacle avoidance.

[0046] The operation center includes: a scene analysis module, a classification training combination module, a training coordination module and a robot scheduling module;

[0047] The scenario analysis module is used to determine the actual application scenario of the logistics robot according to the operation process data of each logistics robot, and then determine the targets covered in the actual application scenario and classify the targets; for example, according to the path data of the logistics robot, it is determined that the actual application scenario of the logistics robot is outdoor scene delivery, and then it is determined that the targets covered in the actual application scenario of the logistics robot are mainly human body\vehicle\obstacle\traffic markings. For logistics robots for indoor scene logistics delivery, the actual application scenarios mainly include human body\building structure\corridor path.

[0048] The classification training combination module includes multiple classification trainers, each of which can classify a single type of target, including delivery items\human body\vehicle\obstacle\building structure\traffic markings and signs\corridor paths; and each classification trainer has a training sample set corresponding to the target of this type;

[0049] The training coordination module is used to call at least one classification trainer in the classification training combination module, and can use one classification trainer or multiple classification trainers for combined training to perform knowledge distillation training on the logistics robot;

[0050] The robot scheduling module is used to determine the actual application scenario (outdoor logistics delivery or indoor material delivery) corresponding to the order dispatching demand according to the order dispatching demand of logistics distribution, and then deploy logistics robots trained for the actual application scenario to respond to the corresponding order dispatching demand.

[0051] The operation process data includes: path records, surrounding scene target data during the logistics distribution process, and action execution data.

[0052] The classification trainer is equivalent to the teacher network in knowledge distillation, and the logistics robot is equivalent to the student network in knowledge distillation.

[0053] The training coordination module can call the corresponding classification trainer for the targets covered in the actual application scenarios of each logistics robot, and perform knowledge distillation training on the logistics robot. For example, if the actual application scenarios of the logistics robot for outdoor logistics distribution mainly include targets of the types of human body\vehicle\obstacle\traffic markings, then the corresponding classification trainers of the types of human body\vehicle\obstacle\traffic markings are called to train the logistics robot so that the logistics robot has the classification and recognition capabilities of the above types of targets; for the logistics robot for indoor logistics distribution, if the actual application scenarios mainly include targets of the types of human body\building structure\corridor path, then the classification trainers of the types of human body\building structure\corridor path are called to train the logistics robot so that the logistics robot has the classification and recognition capabilities of the above types of targets.

[0054] The training coordination module can also call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, while other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group.

[0055] The training coordination module can form an edge training group of logistics robots with the same actual application scenarios according to the actual application scenarios of the logistics robots, use the logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios as the teacher network in the knowledge distillation, and use the scene target data recorded when performing previous logistics delivery tasks as the training sample set; and use other logistics robots as the student network in the knowledge distillation.

[0056] The robots dispatched by the robot dispatching module in this application to perform delivery tasks have all been fully trained for the actual application scenarios involved in the delivery tasks (outdoor logistics delivery or indoor material delivery);

[0057] However, due to the dynamic variability of actual application scenarios and the fact that the logistics robots assigned to each scenario are not constant, for example, there are many delivery tasks in indoor scenarios recently, and the original indoor scene robots are not enough. It is necessary to add some robots that originally performed outdoor scenes to perform indoor scenes. At this time, enhanced training or supplementary training is required. Enhanced training refers to enhancing and strengthening the scenarios that have been trained in the past, and supplementary training refers to training for scenarios that have not been trained in the past.

[0058] At this time, the scene analysis module is needed to judge the previous surrounding scenes of each robot based on the previous operation process data (mainly path data) of each robot, so as to decide the enhanced training or supplementary training of the robot; there are two ways to enhance training and supplementary training. The first is to call multiple classification trainers in the classification training combination module and combine them to train the logistics robot. This is actually more suitable for supplementary training; the second is to form a marginal training group and use "experienced" logistics robots to train "inexperienced" logistics robots for mutual training, which is more suitable for enhanced training; whether supplementary training can only use the first training method, or whether enhanced training can only use the second training method, is not absolute, and the two can be combined.

[0059] The logistics robot is connected to the wireless broadband Internet of Things, has uplink and downlink network communication capabilities, and can maintain real-time communication with the cloud-based operation center. During the entire operation of the logistics robot, the operation process data is uploaded in real time.

[0060] The above outdoor and indoor scenes are only examples. The actual application scenarios that can be included in practice may include various actual application scenarios such as "hospital, airport, high-speed rail station, shopping mall, school", and the scene targets involved are different.

[0061] The core of this application is to use knowledge distillation for training. Knowledge distillation is usually used in many fields such as speech recognition and natural language processing, and knowledge distillation is one of the important technologies in model compression. These fields are not very relevant to the field of logistics robot technology of this application and cannot be directly borrowed. For example: the first technical problem faced by this application is who is the teacher and who is the student. Knowledge distillation adopts a teacher-student model: a complex and large model is used as a teacher, and the student model structure is relatively simple. The teacher is used to assist the training of the student model. The teacher has a strong learning ability and can transfer the knowledge it has learned to the student model with relatively weak learning ability, so as to enhance the generalization ability of the student model. The complex and bulky but effective teacher model is not online, it is simply a mentor role. The flexible and lightweight student model is actually deployed online for prediction tasks. The traditional method for enhancing the training of logistics robots is the training set provided directly by the manufacturer. However, the scenarios covered by the delivery tasks assigned to each robot are quite different and diverse. It is difficult for the training sets provided and updated by the manufacturer to fully cover these scenarios. It is difficult to fully cover all scenarios directly with the training sets provided by the manufacturer. Therefore, the real "teacher" cannot be the training set provided by the manufacturer. Therefore, this application adopts the operation process data transmitted back to the operation center by all logistics robots, and classifies these data in detail, with a certain classification or a combination of several classifications as the "teachers", and the logistics robots that need to be deployed to perform logistics tasks as the "students", so as to enhance the generalization ability of the student model. At the same time, "experienced" logistics robots can also be used as "teachers", which greatly improves the accuracy and efficiency of logistics robot training.

[0062] The second technical problem is how to effectively apply knowledge distillation to the training of logistics robots. Knowledge distillation is divided into two categories, one is target distillation and the other is feature distillation. This application adopts the feature distillation method in knowledge distillation, because the classification step has been completed before knowledge distillation. The classification step generally uses a classifier with a multi-layer convolutional neural network and an output category probability distribution. Of course, a knowledge distillation classifier can also be used, but this is not a problem that needs to be discussed in this application. No matter how the classification is performed, it does not have much impact on the technical effect of this application. This application uses the feature distillation method in knowledge distillation to quickly and accurately transfer the "experience" in the teacher network to the student network. The specific steps are as follows:

[0063] The first stage: First, select the intermediate layers to be distilled, that is, the hidden layers in the teacher network and the layers to be guided in the student network, so that the layers to be guided in the student network learn the hidden layer outputs of the teacher network;

[0064] Since the output sizes of the two may be different, an additional convolutional layer is added after the layer that needs to be guided so that the output size matches the hidden layer of the teacher. Then the layer that needs to be guided of the student network is trained through knowledge distillation so that the middle layer of the student network learns the output of the hidden layer of the teacher network.

[0065] It is to guide the pre-training student network model according to the loss of the teacher network model. The first h layers of the teacher network are recorded as W Hint , which means guidance. The first g layers of the student network are W Guided , which means being guided. At the beginning of training, the student network is randomly initialized. It is necessary to learn a mapping function W r Make W Guided The dimension matches W Hint , get the parameter initialization value of the student network model in the next stage, and minimize the MSE (Mean Square Error) difference of the two network outputs as the loss (feature distillation), as follows:

[0066]

[0067] Among them, W Hint are the parameters of some layers of the teacher network (i.e., the parameters of other layers of the teacher network except the first layer); W Guided are the parameters of some layers of the student network (i.e., the parameters of other layers of the student network except the first layer); W r is a fully connected layer used to match the output size of the teacher network and the student network, because the hidden layer width of the student network is narrower than that of the teacher network, u h (x; W Hint ) is the current value x and the variable W of the first h layers of the teacher network Hint The error value between g (x; W Guided ) is the current value x and the first g layers of the student network variable W Guided The error value between g (x; W Guided );W r ) is v g (x; W Guided ) and W r The error value between .

[0068] The second stage: Use Soft-target to train the student network.

[0069] After training the layers that need to be guided, the current parameters are used as the initial parameters of the student network, and all layer parameters of the student network are trained using knowledge distillation to make the student network learn the output of the teacher network. Since the teacher's predictions for simple tasks are very accurate, in order to weaken the predicted output and make the information contained richer, the loss function at this time is:

[0070]

[0071] Among them, H refers to the cross entropy loss function; λ is an adjustable parameter to balance the two cross entropies; H(y true , P S ) is the student network P S The output and the corrected true value y true The cross entropy loss between ; Softmax output for the student network Softmax output of the teacher network The cross entropy loss between .

[0072] On the second aspect, this application proposes a multi-logistics robot training method based on edge training, such as Figure 2 As shown, the following steps are included:

[0073] Step S1: Determine the actual application scenario of the logistics robot according to the operation process data of each logistics robot, and then determine the goals covered in the actual application scenario;

[0074] Step S2: using a classification trainer to classify a single type of target;

[0075] Step S3: calling at least one classification trainer to perform knowledge distillation training on the logistics robot;

[0076] Step S4: According to the order dispatching requirements of logistics distribution, determine the actual application scenarios corresponding to the order dispatching requirements, and then deploy logistics robots trained for the actual application scenarios to respond to the corresponding order dispatching requirements.

[0077] The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot comprises the following steps:

[0078] The classification trainer is used as the teacher network in knowledge distillation, and the logistics robot is used as the student network in knowledge distillation;

[0079] For each target covered in the actual application scenario of the logistics robot, the corresponding classification trainer is called to perform knowledge distillation training on the logistics robot.

[0080] The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot also includes the following steps:

[0081] Call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, while other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group;

[0082] According to the actual application scenarios of logistics robots, logistics robots with the same actual application scenarios are formed into an edge training group. Logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios are used as teacher networks in knowledge distillation, and the scene target data recorded when performing previous logistics delivery tasks are used as training sample sets; other logistics robots are used as student networks in knowledge distillation.

[0083] In the third aspect, this application proposes a multi-logistics robot delivery method based on edge training, such as Figure 3 As shown, the following steps are included:

[0084] Step S100: According to the dispatch of the operation center, the logistics robot completes the designated delivery task;

[0085] Step S200: The logistics robot can automatically generate and record operation process data, and upload the operation process data to the operation center in real time.

[0086] The applicant of the present invention has made a detailed explanation and description of the implementation examples of the present invention in conjunction with the drawings in the specification. However, those skilled in the art should understand that the above implementation examples are only preferred implementation schemes of the present invention, and the detailed description is only to help readers better understand the spirit of the present invention, but not to limit the scope of protection of the present invention. On the contrary, any improvements or modifications based on the inventive spirit of the present invention should fall within the scope of protection of the present invention.

Claims

1. A multi-logistics robot system based on edge training, characterized in that: include: A logistics robot and an operation center in the cloud; the logistics robot communicates with the operation center in real time; The logistics robot is used to realize the autonomous execution of all operation links in the logistics distribution process, integrates various types of sensors and intelligent controllers, can automatically generate and record operation process data, and upload the operation process data to the operation center in real time; The operation center is used to classify the targets covered in the actual application scenarios according to the operation process data of each logistics robot, perform knowledge distillation training on the logistics robots based on the classification trainers corresponding to different classifications, and call the trained logistics robots so that the logistics robots can complete the designated delivery tasks.

2. The multi-logistics robot system based on edge training as claimed in claim 1, characterized in that: The operation center includes: a scene analysis module, a classification training combination module, a training coordination module and a robot scheduling module; The scenario analysis module is used to determine the actual application scenario of the logistics robot according to the operation process data of each logistics robot, and then determine the targets covered in the actual application scenario, and classify the targets; The classification training combination module includes a plurality of classification trainers, each of which can classify a single type of target, and each of which has a training sample set corresponding to the target of this type; The training coordination module is used to call at least one classification trainer in the classification training combination module to perform knowledge distillation training on the logistics robot; The robot scheduling module is used to determine the actual application scenarios corresponding to the order dispatching requirements of logistics distribution, and then deploy logistics robots trained for the actual application scenarios to respond to the corresponding order dispatching requirements.

3. The multi-logistics robot system based on edge training as described in claim 1 or 2, characterized in that: The operation process data includes: path records, surrounding scene target data during the logistics distribution process, and action execution data.

4. The multi-logistics robot system based on edge training as claimed in claim 2, characterized in that: The classification trainer is equivalent to the teacher network in knowledge distillation, and the logistics robot is equivalent to the student network in knowledge distillation; The training coordination module can call the corresponding classification trainer for each logistics robot's actual application scenario and perform knowledge distillation training on the logistics robot.

5. The multi-logistics robot system based on edge training as claimed in claim 2, characterized in that: The training coordination module can also call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, and other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group; The training coordination module can form an edge training group of logistics robots with the same actual application scenarios according to the actual application scenarios of logistics robots, use logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios as teacher networks in knowledge distillation, and use scene target data recorded when performing previous logistics delivery tasks as training sample sets; and use other logistics robots as student networks in knowledge distillation.

6. The multi-logistics robot system based on edge training as claimed in claim 1, characterized in that: The logistics robot is connected to the wireless broadband Internet of Things, has uplink and downlink network communication capabilities, and can maintain real-time communication with the cloud-based operation center. During the entire operation process, the logistics robot keeps uploading the operation process data in real time.

7. A multi-logistics robot training method based on edge training, characterized in that: The steps include: According to the operation process data of each logistics robot, determine the actual application scenario of the logistics robot, and then determine the goals covered in the actual application scenario; Use a classification trainer to classify a single type of target; Call at least one classification trainer to perform knowledge distillation training on the logistics robot; According to the order dispatching requirements of logistics distribution, the actual application scenarios corresponding to the order dispatching requirements are determined, and then logistics robots trained for the actual application scenarios are deployed to respond to the corresponding order dispatching requirements.

8. The multi-logistics robot training method based on edge training as claimed in claim 7, characterized in that: The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot comprises the following steps: The classification trainer is used as the teacher network in knowledge distillation, and the logistics robot is used as the student network in knowledge distillation; For each target covered in the actual application scenario of the logistics robot, the corresponding classification trainer is called to perform knowledge distillation training on the logistics robot.

9. The multi-logistics robot training method based on edge training as claimed in claim 7, characterized in that: The calling of at least one classification trainer to perform knowledge distillation training on the logistics robot also includes the following steps: Call two or more logistics robots to form an edge training group, where some logistics robots serve as teacher networks in knowledge distillation, while other logistics robots are equivalent to student networks in knowledge distillation, thereby achieving mutual training of logistics robots in the edge training group; According to the actual application scenarios of logistics robots, logistics robots with the same actual application scenarios are formed into an edge training group, and logistics robots that perform delivery tasks many times or for a long time in the actual application scenarios are used as teacher networks in knowledge distillation, and the scene target data recorded when performing previous logistics delivery tasks are used as training sample sets; other logistics robots are used as student networks in knowledge distillation.

10. A multi-logistics robot delivery method based on edge training, characterized in that: The steps include: According to the dispatch of the operation center, the logistics robot completes the designated delivery tasks; Logistics robots can automatically generate and record operation process data, and upload the operation process data to the operation center in real time.

Images