Model training method, car washing method of car washing robot and related device

Abstract

The embodiment of the invention provides a model training method, a car washing method of a car washing robot and a related device, and relates to the technical field of intelligent car washing robots. The training method comprises the following steps: loading a vehicle washing environment simulation model, a vehicle simulation model and a vehicle washing robot simulation model; under the environment simulated by the car washing environment simulation model, simulating acquisition operation through the car washing robot simulation model, and obtaining simulation cleaning parameters required for cleaning the car simulation model; the simulation cleaning parameters are input into the initial vehicle washing strategy model, and a vehicle washing strategy of the vehicle simulation model is generated; a vehicle washing strategy is adopted to control the vehicle washing robot simulation model to simulate cleaning operation on the vehicle simulation model in a simulation environment, and a cleaning result is obtained; and performing parameter adjustment on the initial car washing strategy model according to the washing result. The method can reduce training time consumption and training difficulty.

Description

Technical Field

[0001] This application relates to the field of intelligent car wash robot technology, and in particular to a model training method, a car wash robot method, and related devices. Background Technology

[0002] With the development of robotics technology, various types of robots are being promoted and applied in real-world scenarios. Car wash robots are robots used to clean vehicles and can be used in car washes, repair shops, or vehicle charging stations. For example, an embodied intelligent car wash robot includes components such as a walking section and a robotic arm. The walking section allows the embodied intelligent car wash robot to move freely in the environment; the cleaning components of the robotic arm can perform cleaning operations such as spraying, wiping, or dusting on the vehicle, thereby cleaning it.

[0003] A car wash strategy can be understood as a cleaning plan developed by a car wash robot for a vehicle to be cleaned. For example, a car wash strategy may include the division of the area to be cleaned for the vehicle, the movement route of the car wash robot during cleaning, the motion trajectory of the robotic arm during cleaning, and cleaning motion parameters such as the force and angle of the robotic arm's movements.

[0004] Car wash robots can include artificial intelligence models (hereinafter referred to as models) used to generate car wash strategies for vehicles to be cleaned. By training the model, its ability to generate car wash strategies can be improved, enabling it to formulate car wash strategies that are more conducive to improving cleaning results. However, due to various factors, some methods for training car wash robot models suffer from problems such as long training time and high training difficulty. Summary of the Invention

[0005] This application provides a model training method, a car washing robot method, and related apparatus to solve the problems of long training time and high training difficulty when training the model of a car washing robot in some methods.

[0006] In a first aspect, embodiments of this application provide a model training method, the method comprising:

[0007] Load the car wash environment simulation model, vehicle simulation model, and car wash robot simulation model;

[0008] In the environment simulated by the car wash environment simulation model, the car wash robot simulation model simulates the data collection operation to obtain the simulation cleaning parameters required for the vehicle cleaning simulation model.

[0009] Input the simulation cleaning parameters into the initial car wash strategy model to generate the car wash strategy of the vehicle simulation model;

[0010] A car wash strategy is used to control a car wash robot simulation model in a simulated environment to simulate cleaning operations on a vehicle simulation model and obtain cleaning results.

[0011] The parameters of the initial car wash strategy model are adjusted based on the cleaning results.

[0012] In one possible implementation, the environmental simulation model includes a basic environmental simulation model and at least one variant environmental simulation model; loading the car wash environmental simulation model includes:

[0013] During the first training round of multiple training rounds, the basic environment simulation model is loaded;

[0014] During training cycles other than the first round, a variant environment simulation model is loaded.

[0015] In one possible implementation, the method further includes:

[0016] Based on the real car wash environment, an environmental simulation model corresponding to the real car wash environment is constructed to obtain the basic environmental simulation model;

[0017] By adjusting the environmental factor parameters of the basic environmental simulation model, at least one variant environmental simulation model is obtained. The environmental factor parameters include at least one of the following: weather disturbance parameters, illumination disturbance parameters, and obstacle disturbance parameters.

[0018] In one possible implementation, the method further includes:

[0019] Obtain the target cleaning parameters corresponding to the target car wash case in a real car wash scenario. The target car wash case is the car wash case in which the cleaning result of the car wash robot does not meet the preset cleaning result among multiple car wash cases in a real car wash scenario.

[0020] The target cleaning parameters are input into the initial car wash strategy model after parameter adjustment, and the car wash strategy corresponding to the target cleaning parameters is generated.

[0021] The car wash robot simulation model is controlled by the car wash strategy corresponding to the target cleaning parameters to simulate the cleaning of the vehicle simulation model corresponding to the target car wash case, and the cleaning results are obtained.

[0022] Based on the cleaning results, the parameters of the initial car wash strategy model with adjusted parameters are adjusted.

[0023] In one possible implementation, the method further includes:

[0024] After the initial car wash strategy model is trained, a new car wash strategy model is obtained by knowledge distillation based on the initial car wash strategy model. The model parameter size of the initial car wash strategy model is larger than that of the car wash strategy model.

[0025] Secondly, embodiments of this application also provide a car washing method using a car washing robot. The car washing robot includes an initial car washing strategy model or a car washing strategy model obtained after training through the first aspect and / or various possible implementations of the first aspect. The car washing method includes:

[0026] In any real car wash scenario, the car wash robot's sensors collect the actual cleaning parameters required to clean the target vehicle.

[0027] Input the actual cleaning parameters into the initial car wash strategy model or the car wash strategy model to generate the car wash strategy for the target vehicle.

[0028] The target vehicle is cleaned in any real-world car wash scenario using the target vehicle's car wash strategy.

[0029] Thirdly, embodiments of this application also provide a model training apparatus, the apparatus comprising:

[0030] The loading module is used to load the car wash environment simulation model, the vehicle simulation model, and the car wash robot simulation model.

[0031] The data acquisition module is used to acquire the simulation cleaning parameters required by the vehicle washing simulation model by simulating the data acquisition operation through the car wash robot simulation model in the environment simulated by the car wash environment simulation model.

[0032] The generation module is used to input the simulation cleaning parameters into the initial car wash strategy model and generate the car wash strategy of the vehicle simulation model.

[0033] The cleaning module is used to control the car wash robot simulation model in a simulated environment to simulate cleaning operations on the vehicle simulation model and obtain cleaning results.

[0034] The optimization module is used to adjust the parameters of the initial car wash strategy model based on the cleaning results.

[0035] In one possible implementation, the environmental simulation model includes a basic environmental simulation model and at least one variant environmental simulation model; the loading module is specifically used for:

[0036] During the first training round of multiple training rounds, the basic environment simulation model is loaded;

[0037] During training cycles other than the first round, a variant environment simulation model is loaded.

[0038] In one possible implementation, the device further includes a building module for:

[0039] Based on the real car wash environment, an environmental simulation model corresponding to the real car wash environment is constructed to obtain the basic environmental simulation model;

[0040] By adjusting the environmental factor parameters of the basic environmental simulation model, at least one variant environmental simulation model is obtained. The environmental factor parameters include at least one of the following: weather disturbance parameters, illumination disturbance parameters, and obstacle disturbance parameters.

[0041] In one possible implementation, the device further includes an acquisition module, used to: acquire target cleaning parameters corresponding to a target car wash case in a real car wash scenario, wherein the target car wash case is a car wash case in which the cleaning result of the car wash robot does not meet the preset cleaning result among multiple car wash cases in a real car wash scenario.

[0042] The generation module is also used to: generate a car wash strategy corresponding to the target cleaning parameters in the initial car wash strategy model after the target cleaning parameters are input and adjusted;

[0043] The cleaning module is also used to: control the car wash robot simulation model to simulate cleaning the vehicle simulation model corresponding to the target car wash case using the car wash strategy corresponding to the target cleaning parameters, and obtain the cleaning results;

[0044] The optimization module is also used to adjust the parameters of the initial car wash strategy model after parameter adjustment based on the cleaning results.

[0045] In one possible implementation, the apparatus further includes a distillation module for:

[0046] After the initial car wash strategy model is trained, a new car wash strategy model is obtained by knowledge distillation based on the initial car wash strategy model. The model parameter size of the initial car wash strategy model is larger than that of the car wash strategy model.

[0047] Fourthly, embodiments of this application also provide a car washing device for a car washing robot. The car washing robot includes an initial car washing strategy model or a car washing strategy model obtained after training through the first aspect and / or various possible implementations of the first aspect. The car washing device includes:

[0048] The sensor module is used to collect the actual cleaning parameters required to clean the target vehicle in any real car wash scenario through the car wash robot's sensors.

[0049] The processing module is used to input the actual cleaning parameters into the initial car wash strategy model or the car wash strategy model to generate the car wash strategy for the target vehicle.

[0050] The execution module is used to clean the target vehicle in any real-world car wash scenario using the target vehicle's car wash strategy.

[0051] Fifthly, embodiments of this application provide an electronic device, including: a memory and a processor; the memory stores computer-executable instructions; the processor executes the computer-executable instructions stored in the memory, causing the processor to perform the methods described in the first aspect, various possible implementations of the first aspect, and / or the second aspect.

[0052] In a sixth aspect, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect, various possible implementations of the first aspect, and / or the method of the second aspect.

[0053] In a seventh aspect, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect, various possible implementations of the first aspect, and / or the method of the second aspect.

[0054] This application provides a model training method, a car wash robot, and related apparatus. The model training method, by loading a car wash environment simulation model, a vehicle simulation model, and a car wash robot simulation model, recreates a realistic car wash application scenario in a high-fidelity simulation world, establishing a rapid transfer channel from simulation to reality. Thus, in any training round, there is no need to build a real car wash environment or use real vehicles; various environment simulation models and vehicle model simulation models can be selected within the simulated environment. By simulating data collection operations and inputting the collected simulated cleaning parameters into the initial car wash strategy model, the model's ability to generate car wash strategies can be trained. Furthermore, by adjusting the parameters of the initial car wash strategy model based on the cleaning results, the model parameters can be optimized, improving the ability to generate car wash strategies. Therefore, this method significantly reduces training difficulty, allows for rapid start and completion of training, shortens training time, and thus achieves the effects of reducing training time and difficulty. Attached Figure Description

[0055] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0056] Figure 1 A schematic flowchart illustrating the model training method provided in this application embodiment;

[0057] Figure 2 A training system architecture diagram provided for embodiments of this application;

[0058] Figure 3 A schematic diagram of the environmental disturbance library provided in the embodiments of this application;

[0059] Figure 4A schematic diagram of the model training method provided in the embodiments of this application. Figure 1 ;

[0060] Figure 5 A schematic diagram of the model training method provided in the embodiments of this application. Figure 2 ;

[0061] Figure 6 A schematic flowchart of a car washing method using a car washing robot provided in an embodiment of this application;

[0062] Figure 7 This is a schematic diagram of the structure of the model training device provided in the embodiments of this application;

[0063] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0064] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0065] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims.

[0066] The collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved in the technical solutions of this application comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0067] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0068] In the embodiments of this application, the use of terms such as "first" and "second" is to distinguish between identical or similar items that have essentially the same function and effect. For example, "first electronic device" and "second electronic device" are merely used to distinguish different electronic devices and do not limit their order of execution. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different.

[0069] In this application embodiment, "at least one" refers to one or more, and "more than one" refers to two or more. "And / or" describes the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following associated objects have an "or" relationship.

[0070] For example, different types of vehicles differ in many ways, such as vehicle structure, body size, and the complexity of body curves. For instance, there are many differences between pickup trucks, SUVs, and sedans. To enable a car wash robot to adapt to various types and even specific models of vehicles and to develop effective washing strategies for each type, the robot typically needs to collect real-world data from each type of vehicle and then train to generate washing strategies based on that data. Through multiple training cycles for each vehicle type, the model's experience and ability to generate washing strategies are continuously optimized, allowing the car wash robot to develop better strategies in real-world applications.

[0071] Furthermore, car wash robots need to collect images or 3D point cloud data of the vehicles to be washed using various sensors in order to generate a car wash strategy. Environmental factors during the car wash process significantly impact the accuracy of sensor data collection, thus affecting the robot's perception of the vehicle. Therefore, besides vehicle model differences being a crucial factor in generating a car wash strategy, environmental factors are another important consideration. For example, in rainy, foggy weather or environments with significant changes in lighting, sensors such as LiDAR and vision cameras may distort, resulting in errors in the collected data and affecting the model's ability to generate a superior car wash strategy.

[0072] Training a car wash robot model in the real world requires using various types of vehicles and conducting real-world training in diverse environments. However, the number of car models in real-world scenarios is vast, and creating car wash environments with varying weather, lighting, and obstacles is a challenging and time-consuming task. Therefore, training the car wash robot's controller in the real world presents both significant challenges in terms of training time and difficulty.

[0073] In view of this, this application provides a model training method. This method loads a car wash environment simulation model, a vehicle simulation model, and a car wash robot simulation model to recreate a realistic car wash application scenario in a high-fidelity simulation world, establishing a rapid transfer channel from simulation to reality. Thus, in any training round, there is no need to build a real car wash environment or use real vehicles; various environment simulation models and vehicle model simulation models can be selected within the simulation environment. By simulating data collection operations and inputting the collected simulation cleaning parameters into the initial car wash strategy model, the model's ability to generate car wash strategies can be trained. Furthermore, by adjusting the parameters of the initial car wash strategy model based on the cleaning results, the model parameters can be optimized, improving the ability to generate car wash strategies. Therefore, this method can greatly reduce training difficulty, quickly start and complete training, shorten training time, and thus achieve the effects of reducing training time and training difficulty.

[0074] It is understood that the model training method provided in this application can be used to train car wash strategy models for various intelligent car wash robots, including but not limited to embodied intelligent car wash robots. The training method in this application helps to build an efficient intelligent car wash robot simulation training platform. This platform allows for the training of various intelligent car wash robots with low difficulty and high efficiency, and can improve the performance of intelligent car wash robots.

[0075] The technical solutions of this application will be described in detail below with reference to specific embodiments. The specific embodiments described below can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments. The embodiments of this application will be described below with reference to the accompanying drawings.

[0076] Figure 1 This is a flowchart illustrating the model training method provided in an embodiment of this application. The execution entity of this method can be an electronic device with corresponding data storage and computing capabilities, such as a computer, server, or server cluster. Figure 1 As shown, the method includes:

[0077] S101 loads the car wash environment simulation model, vehicle simulation model, and car wash robot simulation model.

[0078] For example, the car wash environment simulation model can be a virtual car wash environment built through a simulation engine, the vehicle simulation model can be a virtual vehicle built through a simulation engine, and the car wash robot simulation model can also be a virtual car wash robot built through a simulation engine. For instance, each simulation model is obtained by performing a high-fidelity model simulation in a simulation engine based on a real car wash environment, a real vehicle, and a real car wash robot, and each simulation model can be a three-dimensional simulation model.

[0079] A simulation engine can be any engine capable of building simulation models and simulating the mechanical interactions between robots and objects. Examples include high-fidelity simulation engines and robot simulation software. For instance, a high-fidelity simulation engine can reflect the mechanical interactions, such as collisions and friction, between a car wash robot and objects like cars or obstacles in a simulated environment.

[0080] The simulation engine can include physics-related physics engines. These include: a rigid body dynamics model to simulate the contact force between the robotic arm's brush head and the vehicle body, with a friction coefficient adjustable from 0.2 to 0.8; a fluid model to simulate the hydrodynamic characteristics of water spray, with water pressure ranging from 0.5 to 3 MPa and atomization angle from 60° to 120°; and a robotic arm kinematics model to implement complete joint constraints, such as velocity, acceleration, or torque limits.

[0081] Before training, a simulation model of the car wash robot can be pre-built using simulation model building software or a high-fidelity simulation engine. Multiple vehicle simulation models can be built for various vehicle types, or multiple environment simulation models can be built based on one or more real car wash environments. This results in a simulation model library. During training, any one or more simulation models from this library can be called and loaded in real time as needed, eliminating the need for training with real-world models, thus improving training efficiency and reducing training difficulty.

[0082] Digital twin technology is understood as constructing a real-time, synchronous virtual simulation model of a physical entity using technologies such as sensors, the Internet of Things (IoT), and artificial intelligence (AI), enabling digital mapping and optimization decision support for the entire lifecycle of real-world objects. Building a simulation model using a simulation engine and then using that model to simulate the car wash process can be understood as follows: Applying digital twin technology to the simulation engine, a high-fidelity car wash scenario simulation environment is constructed, including a physical engine with models of robotic arm dynamics and contact mechanics; and sensor simulation models including LiDAR noise and point cloud characteristics. Within this high-fidelity simulation environment, the model used to generate car wash strategies for the car wash robot is effectively trained.

[0083] S102, in the environment simulated by the car wash environment simulation model, the car wash robot simulation model simulates the data acquisition operation to obtain the simulation cleaning parameters required for the vehicle cleaning simulation model.

[0084] For example, a car wash robot may include a controller, which can be understood as a hardware device, software device, or a combination of software and hardware device for controlling and managing the car wash robot. For instance, the controller is a control chip or control system, which contains a model for generating car wash strategies. This model can be an artificial intelligence model or a neural network.

[0085] In addition to the controller, a car wash robot can also include sensors for collecting cleaning parameters. These include sensors such as LiDAR sensors, six-dimensional force sensors, and vision cameras. LiDAR sensors can collect point cloud data of objects in the environment, such as the vehicle to be washed, thus helping the car wash robot perceive its environment, the vehicle to be washed, and its positional relationship with other objects. Six-dimensional force sensors can collect the values ​​of the forces generated between the robotic arm and the vehicle during washing, allowing for appropriate force application or avoidance. Vision cameras can acquire images of the environment and objects, which also helps the robot perceive objects in its environment.

[0086] When constructing a simulation model of a car wash robot, high-fidelity sensor simulation models that reproduce real sensors can also be built within the car wash robot simulation model. For example, in addition to constructing a geometric simulation model of the car wash robot, sensor simulation models of sensors such as LiDAR sensors, six-dimensional force sensors, and vision cameras can be added to the geometric simulation model. These simulated sensor models can realize the functions that real sensors can perform in the real world within the simulation environment.

[0087] Furthermore, perturbations can be added to sensor simulation models to simulate data acquisition errors or distortions that occur when real sensors collect data in real-world scenarios. For example, the point cloud density error of a lidar sensor simulation model can be set to ±5%, its maximum ranging distance to 20m, and the sensor can be simulated to have visibility less than 50m under perturbation conditions such as rain and fog. Similarly, the noise of a six-dimensional force sensor simulation model can be set to ±0.5N, the sampling rate to 1kHz, and the vibration interference under perturbation simulation to be 5-200Hz. Furthermore, the resolution of a vision camera simulation model can be set to 1920×1080 pixels, and it can be configured as a distortion simulation model, with perturbation simulations including strong light glare or water stain reflections. Perturbations in sensor simulation models can also be achieved by adding random noise.

[0088] By constructing a highly realistic sensor simulation model, the simulation parameters required for the car wash robot to clean the vehicle can be collected in the simulated car wash environment of the simulation engine. This can highly replicate the data collection operation of the car wash robot in the real car wash process, thus enabling better training results through digital twin training.

[0089] In a realistic car wash environment, the types of realistic cleaning parameters can include vehicle structural parameters, such as the shape and size parameters of the vehicle roof, windshield, doors, and handles; they can also include the distance and orientation parameters between the car wash robot and the various surfaces of the vehicle to be cleaned, which can be understood as the pose parameters of the car wash robot and the vehicle; and they can also include the pose parameters of fixed obstacles such as walls and tracks, or moving obstacles such as pedestrians in a real car wash environment. In the simulated car wash environment of the simulation engine, the types of simulated cleaning parameters can be consistent with the types of realistic cleaning parameters in a real car wash environment.

[0090] By simulating data collection operations using a car wash robot simulation model, the necessary simulation cleaning parameters for the vehicle washing simulation model can be acquired through sensor simulation models. For example, actions such as sending and / or receiving acquisition signals by various sensor simulation models within the car wash robot simulation model can be used to collect data from the vehicle simulation model, thereby obtaining the required simulation cleaning parameters.

[0091] S103: Input the simulation cleaning parameters into the initial car wash strategy model to generate the car wash strategy of the vehicle simulation model.

[0092] For example, the initial car wash strategy model can be understood as the model used to generate the car wash strategy before parameter adjustment. This initial car wash strategy model can be, for example, an artificial intelligence model of reinforcement learning neural network or optimization algorithm, or other neural networks, etc.

[0093] The initial car wash strategy module can interact with the simulation engine. After obtaining the cleaning parameters, these parameters can be input into the initial car wash strategy module through a data interaction channel for analysis and prediction. In this embodiment, the cleaning parameters can be understood as simulation cleaning parameters, real cleaning parameters, or target cleaning parameters, and these parameters can be of the same type. Simulation cleaning parameters can be understood as cleaning parameters collected by the sensor simulation model in a simulation environment. Real cleaning parameters can be understood as cleaning parameters collected by real sensors in a real environment. Target cleaning parameters can be understood as cleaning parameters for a target car wash case, also collected by real sensors during a real car wash process.

[0094] After collecting the simulation cleaning parameters, these parameters can be input into the initial car wash strategy module for data calculation, strategy analysis, and result prediction, thereby generating a car wash strategy. This car wash strategy is generated for the current training cycle and includes the area division planning of the vehicle simulation model's cleaning area, the movement route of the car wash robot simulation model during cleaning, the motion trajectory of the robotic arm simulation model in the car wash robot simulation model during cleaning, and cleaning action parameters such as the force and angle of the motion.

[0095] S104 uses a car wash strategy to control the car wash robot simulation model in a simulated environment to simulate the cleaning operation of the vehicle simulation model and obtain the cleaning result.

[0096] For example, the car wash strategy can also be fed back to the simulation engine through the interaction channel between the initial car wash strategy model and the simulation engine. The simulation engine can then perform a cleaning operation on the vehicle simulation model based on the area division planning, the movement route of the car wash robot simulation model, and the cleaning action parameters included in the car wash strategy. Furthermore, the simulation engine can obtain the cleaning result based on the cleaning effect after cleaning. This cleaning result can include information characterizing the cleaning effect.

[0097] For example, the cleaning result may include a cleaning score and / or evaluation values ​​from multiple dimensions. The cleaning score can be an overall assessment of the cleaning effect, for example, a maximum score of 100 points. The cleaning score can be obtained by comprehensively calculating the evaluation values ​​from multiple dimensions.

[0098] Evaluation values ​​across multiple dimensions can include those for collision count, cleaning coverage (e.g., the ratio of cleaned area to total area to be cleaned), energy consumption, and total cleaning time. For example, if the collision count threshold is 3, and the actual number of collisions exceeds 3, the highest evaluation value for this dimension is 70 points; the higher the actual number of collisions, the lower the evaluation value for that dimension. If the actual number of collisions is 0, the evaluation value can be 100. Other examples include a cleaning coverage threshold of 98%, an energy consumption threshold of 5 kW•h, and a total cleaning time threshold of 15 minutes. Evaluation values ​​can be calculated for each dimension based on actual conditions, and finally, a weighted calculation using fixed or dynamic weights can be performed to obtain the cleaning score.

[0099] S105, Adjust the parameters of the initial car wash strategy model based on the cleaning results.

[0100] For example, the objects of parameter adjustment can be the model parameters of the initial car wash strategy model, or network parameters. Network parameters can be the weights of each network layer in a neural network, and / or the size of the convolutional kernels in convolutional layers, etc. When adjusting the network parameters of the initial car wash strategy model based on the cleaning results, for example, the loss value can be calculated based on the loss function, and the network parameters can be iteratively optimized and adjusted through backpropagation. Alternatively, the parameters can be adjusted iteratively based on the objective optimization conditions as constraints. Alternatively, other training mechanisms can be combined for parameter adjustment.

[0101] Through one or more rounds of training jointly conducted with the simulation engine and the initial car wash strategy model, the network parameters of the initial car wash strategy model can be continuously iteratively optimized. This improves the generation experience and generation capability with lower training difficulty and time. The parameter-adjusted initial car wash strategy model can process real cleaning parameters collected by real sensors in a real car wash environment, thereby generating a car wash strategy that is more suitable for the specific car model in each real environment. Using this car wash strategy is conducive to achieving better cleaning results.

[0102] The model training method provided in this application, by loading a car wash environment simulation model, a vehicle simulation model, and a car wash robot simulation model, recreates a realistic car wash application scenario in a high-fidelity simulation world, establishing a rapid transfer channel from simulation to reality. Thus, in any training round, there is no need to build a real car wash environment or use real vehicles; various environment simulation models and vehicle model simulation models can be selected within the simulated environment. By simulating data collection operations and inputting the collected simulated cleaning parameters into the initial car wash strategy model, the model's ability to generate car wash strategies can be trained. Furthermore, by adjusting the parameters of the initial car wash strategy model based on the cleaning results, the model parameters can be optimized, improving the ability to generate car wash strategies. Therefore, this method significantly reduces training difficulty, allows for rapid start and completion of training, shortens training time, and thus achieves the effects of reducing training time and difficulty.

[0103] Existing car wash robots have significant limitations in their scene adaptability. Scene adaptability can be understood as the ability of a car wash robot to adapt to changes in factors such as car model and car wash environment. Car wash robots should at least be compatible with mainstream car models on the market, adapting to different body curves and sizes of different models, and achieving thorough cleaning without blind spots. Therefore, certain requirements are placed on the scene applicability and even the adaptive capability of car wash robots.

[0104] Typically, when car wash robots are deployed to new scenarios, they require manual adjustments to adapt to changes in vehicle models and washing environments, resulting in low scenario adaptability for existing car wash robots. However, manual adjustments involve data collection and optimization of washing strategies for each new vehicle model, leading to lengthy adjustment cycles. For example, this requires manually collecting point cloud data of the vehicles and calibrating the wiping surfaces to data from LiDAR and vision cameras. It may also necessitate adjusting path planning algorithms based on point cloud data to ensure the robotic arm can wipe all surfaces without leaving any blind spots, while also meeting robotic arm constraints.

[0105] The model trained using the method described in this application can be trained multiple times on the initial car wash strategy model in a simulation environment. During these multiple training rounds, the combination of vehicle simulation models of multiple car models and environmental simulation models of multiple environments enables repetitive, low-difficulty, and low-time-cost training conditions. This allows the trained model to possess high scene adaptability, thereby reducing reliance on manual adjustments when the car wash robot is deployed to new scenes and making it more conducive to performing car wash tasks.

[0106] For example, the perception reliability of a car wash robot decreases under environmental disturbances. Environmental disturbances such as rain, fog, or changes in lighting can cause a decrease in the accuracy or distortion of the robot's sensors. For instance, rain, fog, or changes in lighting can distort the vision camera or LiDAR. Sensor distortion leads to a decrease in cleaning success rate; therefore, applications of car wash robots require a certain degree of environmental adaptability.

[0107] To provide more environmental simulation models for the training process, one or more variant environmental simulation models can be obtained by adding or adjusting perturbation factors to the basic environmental simulation model. This allows for the rapid acquisition of multiple environmental simulation models, which can then be used for training to enable the model to have a higher environmental generalization ability.

[0108] Figure 2 This is a diagram of the training system architecture provided in an embodiment of this application. Figure 3 This is a schematic diagram of the environmental disturbance library provided in the embodiments of this application. Figure 2 The training system shown can be understood as a training system loaded into an electronic device for training an initial car wash strategy model. The environmental disturbance library can be understood as a collection of various disturbance factors that can be added to a basic environmental simulation model, for example, through mathematical modeling. Figure 4 A schematic diagram of the model training method provided in the embodiments of this application. Figure 1 The following is combined with Figure 2 , Figure 3 and Figure 4 The training method provided in the embodiments of this application will be further described.

[0109] In one possible implementation, the method further includes: constructing an environmental simulation model corresponding to the real car wash environment based on the real car wash environment to obtain a basic environmental simulation model; and obtaining at least one variant environmental simulation model by adjusting the environmental factor parameters of the basic environmental simulation model, wherein the environmental factor parameters include at least one of the following: weather disturbance parameters, light disturbance parameters, and obstacle disturbance parameters.

[0110] For example, the basic environmental simulation model can be an environmental simulation model built in a high-fidelity simulation engine for a specific car wash environment. It can be the basis for a series of variant environmental simulation models. By adjusting the environmental factor parameters, one or more variant environmental simulation models can be obtained. For instance, one basic environmental simulation model can yield 100 or even 1000 variant environmental simulation models. Alternatively, multiple basic environmental simulation models can be built based on multiple real car wash environment simulations, and a large number of variant environmental simulation models can be quickly obtained based on these basic models.

[0111] For example, by adding random noise or other methods to adjust the environmental factor parameters of the basic environmental simulation model, multiple variant environmental variables can be derived from the basic environmental simulation model. For instance, the change in illumination over time, or the rainfall value or its variation, can be randomly adjusted in the basic environmental simulation model. The resulting multiple variant environmental simulation models can be used to simulate the impact of various uncertainties in the actual deployment environment on the perception and cleaning performance of the car wash robot.

[0112] like Figure 3 As shown, based on the basic environmental simulation model, at least one of the following parameters can be adjusted: weather disturbance parameters, illumination disturbance parameters, and obstacle disturbance parameters. Weather disturbance parameters include, for example, rainfall adjustments ranging from 1-50 mm / h, and fog visibility adjustments ranging from 10-100 m. Illumination disturbance parameters include, for example, direct sunlight adjustments ranging from 0-100,000 lux, and shadow contrast adjustments ranging from 0%-90%. Obstacle disturbance parameters include moving obstacles such as pedestrians or vehicles, and may also include static obstacles such as cones or debris. Environmental factor parameters can be randomly generated within their respective ranges when adjusting them.

[0113] For example, in addition to modifying the environmental simulation model, parameters of vehicles and robotic arms can also be adjusted to improve the generalization and adaptability of car wash robots. For instance, one or more variant parameters can be adjusted by referring to scenarios where car wash robots have already been deployed. Variant parameters can be understood as any adjustable parameters on the car wash robot, vehicle, or other objects related to the car wash process, such as the degree of dirt on the vehicle, the volume of the stains, and the ease of stain removal.

[0114] For example, to address parking deviation, a disturbance adjustment of ±30cm translation can be made to the vehicle position; to cover various vehicle types such as urban SUVs and commercial vehicles, a disturbance adjustment of ±0.5m length and ±0.2m width can be made to the vehicle dimensions; to enhance environmental robustness, a disturbance adjustment of ±20dB noise ratio can be made to the sensor noise; to compensate for the aging of the robotic arm equipment, a disturbance adjustment of ±0.5° joint accuracy can be made to the robotic arm error; or other variant parameters can be disturbed and adjusted, which is not limited in this application embodiment.

[0115] like Figure 4 As shown, during the training process, environmental factor parameters can be input into the basic environmental simulation model to generate 100 variant environmental simulation models. These models can then be loaded into the training process. A reinforcement learning training algorithm can be used to train the initial car wash strategy model. During training, the initial car wash strategy model generates a car wash strategy, and its performance can be evaluated. If the strategy meets the requirements, it can be output; otherwise, training continues until the desired strategy is achieved. After meeting the requirements, similar or identical training can be performed on other variant environmental simulation models from the 100 variant models.

[0116] In this embodiment, by adjusting at least one environmental factor parameter, such as weather disturbance parameters, illumination disturbance parameters, and obstacle disturbance parameters, multiple environmental simulation models can be obtained quickly and realistically, improving the efficiency of obtaining multiple environmental simulation models. Furthermore, based on the basic environmental simulation model, it is possible to obtain as many simulated car wash environments as possible that are identical or similar to the real car wash environment, and even to obtain environmental simulation models corresponding to real car wash environments that are difficult to occur, such as those with extreme weather or extreme lighting. By using these environmental simulation models to train the initial car wash strategy model, the adaptability and generalization ability of the initial car wash strategy model can be improved.

[0117] In one possible implementation, the environmental simulation model includes a base environmental simulation model and at least one variant environmental simulation model; loading the car wash environmental simulation model includes: loading the base environmental simulation model during the first training round of multiple training rounds; and loading the variant environmental simulation model during training rounds other than the first training round.

[0118] For example, multiple training rounds can be conducted against a base environment simulation model and one or more variant environment simulation models corresponding to that base environment simulation model. In the first round of training, the base environment simulation model is loaded to simulate the environment depicted by that model. In subsequent rounds, variant environment simulation models are loaded, allowing training to be performed using these variant models obtained by adjusting environmental factor parameters from the base environment simulation model.

[0119] For example, before introducing a variant environment simulation model, the initial car wash strategy model can be quickly converged based on the basic environment simulation model and multiple preset car wash strategies, and the car wash decision can be output based on the basic environment simulation model.

[0120] like Figure 2 As shown, the high-fidelity simulation engine includes a physics engine, sensor simulation models, and an environmental perturbation library. It can simulate and build vehicle simulation models for various car models and various environmental simulation models. The high-fidelity simulation engine can interact with the pre-adaptive training module, allowing for joint training of the controller's initial car wash strategy model. The pre-adaptive training module can combine various variant environmental simulation models to optimize the initial car wash strategy model, adjusting network parameters such as trajectory planning parameters to achieve optimized output of the car wash strategy.

[0121] The input parameters of the pre-adaptive training module can include perception information collected by the simulated LiDAR and simulated vision camera of the sensor simulation model, as well as the robot arm's body information such as joint angles, movement speed, acceleration, and Cartesian posture of the car wash robot simulation model itself, and collision information between the car wash robot simulation model and the vehicle simulation model. The output parameters of the pre-adaptive training module can include adjustment parameters such as the direction and magnitude of network parameter adjustments to the initial car wash strategy model. By refreshing the network parameters of the initial car wash strategy model after each training round, the initial car wash strategy model can generate a car wash strategy that is more conducive to improving the cleaning effect.

[0122] When training the initial car wash strategy model in the pre-adaptive training module, a three-stage learning approach can be adopted. For example, Stage 1 involves training under undisturbed conditions in the basic environmental simulation model, allowing the initial car wash strategy model to converge quickly and improving training speed. Stage 2 can involve gradually increasing the intensity of perturbations, progressively improving the initial car wash strategy model's scene adaptability and environmental generalization ability. Stage 3 can involve training under complex perturbation conditions with multiple environmental factors, such as weather, lighting, and obstacles, where these factors can be dynamic parameters that change over time or space. Stage 3 further enhances the initial car wash strategy model's scene adaptability and environmental generalization ability.

[0123] For example, for one of several basic environmental simulation models, Phase 1 requires training under that basic environmental simulation model to allow the trained initial car wash strategy model to quickly converge to a state applicable in general scenarios. Phase 2 can involve increasing the degree of perturbation under a single basic environmental simulation model to train under varying environmental simulation models, such as rainy days with varying intensity from light rain to heavy rain, making the initial car wash strategy model more applicable even in single extreme car wash environments. In Phase 3, multiple perturbation factors such as heavy rain, fog, low visibility, and weak light can be superimposed. Furthermore, perturbation factors such as the complexity of the vehicle's body surface and its large size and area to be cleaned can also be superimposed, further enhancing the initial car wash strategy model's ability to generate superior car wash strategies.

[0124] In this embodiment, loading the basic environment simulation model during the first training round of multiple training rounds allows for training in a undisturbed simulation environment in the initial stages, reducing the initial training difficulty and enabling the model to converge quickly. In subsequent training rounds, loading a variant environment simulation model gradually increases the dimension and intensity of environmental perturbations in the middle or later stages of training, which is beneficial for improving model capabilities. Therefore, progressive training can be achieved, reducing training difficulty and increasing the training success rate. Furthermore, when optimizing network parameters, a multi-objective optimization approach can be combined to comprehensively determine the direction and magnitude of network parameter adjustments. Multi-objective optimization can be performed based on various dimensions such as collision count, cleaning coverage, energy consumption, and total cleaning time.

[0125] Multiple preset car wash strategies can be derived from a pre-stored knowledge base. These preset strategies can be understood as expert car wash strategies, which can be used to assist in training the initial car wash strategy model. Alternatively, in real-world car wash scenarios, under conditions such as extreme weather, lighting, and / or obstacles, the preset car wash strategy that is most beneficial for improving the car wash effect can be determined from among multiple preset strategies and used as the car wash strategy for cleaning. In this case, the initial car wash strategy model or the car wash strategy output by the car wash strategy model may not be used.

[0126] The knowledge base can be a collection of experiential knowledge constructed based on prior knowledge and / or expert rules. Inputting knowledge from the knowledge base into the initial car wash strategy model helps it establish initial experience in generating car wash strategies, enabling it to grow rapidly. Multiple preset car wash strategies can be, for example, multiple expert car wash strategies capable of handling any vehicle and multiple car wash environments. For instance, these preset car wash strategies could include those capable of handling car wash environments with varying rainfall levels and high applicability; those capable of handling car wash environments with varying visibility and fog levels and high applicability; those capable of handling car wash environments with varying light intensity and high applicability; and those capable of handling car wash environments with numerous obstacles and high applicability, etc.

[0127] For example, a knowledge base can be pre-built before training, and the knowledge in the knowledge base can be continuously updated and expanded in subsequent training or practice, so that the optimized knowledge base can play a greater a priori role and can be promoted and reused in similar fields or scenarios.

[0128] When constructing a knowledge base, various types of knowledge can be stored in corresponding storage formats based on the application scenarios. For example, for the application scenario of quickly matching car wash strategies, knowledge of vehicle model features can be stored in the storage format of geometric primitive combination encoding; for the application scenario of direct calling of similar regions, knowledge of optimal trajectory types can be stored in the storage format of B-spline control point sequences; for the application scenario of real-time diagnosis and recovery, knowledge of fault solution types can be stored in the storage format of decision tree paths; and for the application scenario of adverse weather adaptation, knowledge of environmental disturbance response types can be stored in the storage format of reinforcement learning strategy network parameters. Alternatively, knowledge of corresponding types can be stored in other storage formats for other application scenarios, which is not limited in this embodiment.

[0129] In addition, after the car wash robot is deployed to a real car wash scenario, if it encounters an extreme environment, when generating a car wash strategy, it can directly select and determine the car wash strategy through a knowledge base or multiple preset car wash strategies, or it can determine the car wash strategy by comparing the car wash strategy generated based on the initial car wash strategy model or the car wash strategy model with at least some preset car wash strategies.

[0130] By combining the transplantation training strategies of multiple preset car wash strategies, the initial car wash strategy model can gain initial experience and continuously learn and generate car wash strategies based on the preset car wash strategies to cope with various variant environment simulation models, thereby improving the scene adaptability and environment generalization of the initial car wash strategy model.

[0131] Once deployed in a car wash scenario, if the car wash robot lacks the ability to continuously upgrade, it will be difficult to maintain high scenario adaptability and environmental generalization. Moreover, successful car wash cases, failed car wash cases, or boundary success cases implemented by the car wash robot in practice all provide information that helps to continuously evolve the capabilities of the car wash robot. Therefore, in order to continuously improve the capabilities of the car wash robot, the initial car wash strategy model can be continuously optimized through continuous self-optimization.

[0132] In one possible implementation, the method further includes: obtaining target cleaning parameters corresponding to a target car wash case in a real car wash scenario, wherein the target car wash case is a car wash case in which the cleaning result of the car wash robot in a real car wash scenario does not meet the preset cleaning result; inputting the target cleaning parameters into the initial car wash strategy model after parameter adjustment to generate a car wash strategy corresponding to the target cleaning parameters; using the car wash strategy corresponding to the target cleaning parameters to control the car wash robot simulation model to simulate cleaning the vehicle simulation model corresponding to the target car wash case and obtain the cleaning result; and adjusting the parameters of the initial car wash strategy model after parameter adjustment according to the cleaning result.

[0133] For example, the target cleaning parameters corresponding to the target car wash case can be obtained through data transmission. The target car wash case is a car wash case in which the cleaning result of the car wash robot in a real car wash scenario does not meet the preset cleaning result.

[0134] The preset cleaning result can be based on a single evaluation value or multiple evaluation values. For example, the preset car wash result is a cleanliness level greater than or equal to 95%. If the cleanliness level of a car wash is less than 95%, then that car wash can be designated as the target car wash; alternatively, it can be further subdivided into cleanliness levels less than 90% as failed cases, and cleanliness levels between 90% and 95% as borderline successful cases. The cleanliness level can be an evaluation value determined based on factors such as cleaning coverage and the degree of stain residue.

[0135] Target cleaning parameters can be understood as the cleaning parameters collected by the car wash robot when performing a target case. The car wash robot can store target car wash cases and their corresponding target cleaning parameters in practice, and can upload the stored target car wash cases and their corresponding target cleaning parameters to the electronic device within a preset period (such as 1 week or 1 month) for further network parameter optimization.

[0136] After acquiring the target car wash case and target cleaning parameters, the electronic device can perform one or more training rounds for any target car wash case. The training process can be the same as the training method for the initial car wash strategy model before deployment, or it can be fine-tuned. Fine-tuning may include, for example, reducing the dimensionality of high-dimensional data in the target cleaning parameters to obtain low-dimensional feature vectors, and using these low-dimensional feature vectors for training. For example, high-dimensional data may be the pose of a robotic arm in Cartesian space, which can be described as 16-dimensional data using a transformation matrix. Through mathematical transformation, the 16-dimensional data can be converted into 7-dimensional data consisting of x, y, z, and quaternions.

[0137] Alternatively, the architecture of the initial car wash strategy model can be adjusted during training. For example, the network layer structure of the initial car wash strategy model can be dynamically added or removed. For instance, when training on target cleaning parameters, if new representative feature vectors can be obtained from the target cleaning parameters, new network layers can be added for these new feature vectors. For example, a branch structure that can analyze and predict the new feature vectors can be added to the backbone network, or a new output layer can be added to the branch structure. This can output multiple car wash strategies, which can then be compared to determine the car wash strategy to be used.

[0138] like Figure 2 As shown, the training system architecture also includes an online incremental learning module, which can interact with the pre-adaptive training module. This online incremental learning module can receive data collected in real-time from actual application scenarios, such as target cleaning parameters. It can also continuously perform incremental training and fine-tuning of the initial car wash strategy model based on real feedback such as the execution status of the car wash robot and sensor data, and can update the knowledge base based on new experiential knowledge.

[0139] In this embodiment, by obtaining the target cleaning parameters corresponding to the target car wash case, the initial car wash strategy model after parameter adjustment can be further trained. This can also be understood as enabling continuous evolutionary training of the initial car wash strategy model deployed in a real-world scenario. This effectively utilizes data collected in real-world scenarios, relying on a data-driven approach to continuously conduct incremental training and improve the capabilities of the car wash robot.

[0140] Based on this, the simulation engine can accelerate the initial adaptation, and through continuous optimization via online learning, it can reduce the deployment and debugging time and cost in new scenarios, and also enable the autonomous performance evolution of the car wash robot.

[0141] Figure 5 A schematic diagram of the model training method provided in the embodiments of this application. Figure 2 Incremental training of the initial car wash strategy model may also include other steps.

[0142] like Figure 5 As shown, after obtaining the target cleaning parameters, data cleaning and labeling can be performed on these parameters. Then, feature alignment operations can be performed on the features extracted from the cleaned and labeled target cleaning parameters to ensure consistency in the range or format of each feature. After inputting the features into the initial car wash strategy model, the car wash strategy obtained in this round of training can be obtained, which can be understood as the simulated car wash strategy during incremental training. The car wash strategy in the target car wash case corresponding to the target cleaning parameters can be understood as the actual car wash strategy. The performance of the simulated car wash strategy and the actual car wash strategy can be compared. If the simulated car wash strategy is superior to the actual car wash strategy, the sensor simulation model can be corrected; if the actual car wash strategy is superior to the simulated car wash strategy, new knowledge can be extracted. Both correcting the sensor simulation model and extracting new knowledge can be used to update the model or knowledge base, and then subsequent incremental training can be carried out after the update. After training, the trained initial car wash strategy model can be deployed to the car wash robot. After a certain period, the above steps can be repeated to continuously carry out incremental training.

[0143] Through such Figure 5 The closed-loop architecture shown establishes a rapid migration channel from simulation to reality, enabling autonomous performance evolution during operation. Since the initial car wash strategy model in the simulation engine has already handled various car models and car wash environments, the car wash robot is compatible with multiple real-world car wash environments after deployment. It can effectively control the robotic arm to perform tasks in various real-world car wash environments, achieving good car wash results. Furthermore, through continuous autonomous performance evolution, guided by reinforcement learning algorithms, each training round helps the initial car wash strategy model improve its capabilities, allowing it to continuously evolve and ultimately achieve better performance in real-world car wash environments.

[0144] In one possible implementation, the method further includes: after training the initial car wash strategy model, obtaining a new car wash strategy model by knowledge distillation based on the initial car wash strategy model; wherein the model parameter size of the initial car wash strategy model is larger than the model parameter size of the car wash strategy model.

[0145] For example, knowledge distillation can be understood as a training method that compresses or simplifies the model. Although it compresses or simplifies the model parameter size, it does not significantly reduce the model's generative ability. The model parameter size can be understood as the number of model parameters, such as the number and dimensions of weights or weight matrices.

[0146] First, an initial car wash strategy model with a relatively large model parameter scale can be trained using the training method of this application embodiment. This initial car wash strategy model can be understood as a teacher model. When the cleaning results obtained in at least one training round of this teacher model meet the preset target, it indicates that the teacher model can generate a car wash strategy that meets the expectations and has certain generation experience and ability. Then, knowledge distillation can be performed on the teacher model to train a student model with a relatively small model parameter scale. After training the student model to meet the preset target, this student model can be used in actual car wash scenarios and can be understood as a car wash strategy model.

[0147] Training using knowledge distillation can include steps such as freezing the network parameters of the teacher model, mimicking the soft-label output of the teacher model, and matching the real labels of the training samples.

[0148] For example, the same batch of training data is input into both a teacher model and a student model. The teacher model generates soft labels. The student model generates predicted outputs. The difference between the student model's predicted output and the teacher model's soft labels is calculated. The difference between the student model's predicted output and the true labels is also calculated. The two differences are weighted and summed to obtain the total loss. After obtaining the total loss, the network parameters of the student model can be updated through backpropagation, so that the subsequent predicted outputs are closer to the teacher model's soft labels. The soft labels can be understood as the car washing strategy output by the teacher model, the predicted outputs can be understood as the car washing strategy output by the student model, and the true labels can be understood as the better or optimal ground truth car washing strategy. After multiple rounds of distillation training, the knowledge of the teacher model can be distilled into the student model, thus obtaining a car washing strategy model with a smaller parameter size and higher generative ability.

[0149] In this embodiment, the initial car wash strategy model, with its larger parameter scale, more complex structure, and superior performance, can be considered a teacher model. Knowledge distillation using this teacher model not only imparts the generative experience of the teacher model but also yields a lightweight student model with high generative capabilities. This student model is the car wash strategy model, which can be used as the model for practical deployment. In actual application, it not only outputs a more suitable car wash strategy but also achieves fast computation with low computational overhead due to its smaller model parameter scale, improving the efficiency and speed of generating car wash strategies.

[0150] For example, the training method provided in this application embodiment can perform pre-adaptive training based on a simulation engine. Before deploying a new car model or a new scenario, a large number of variant environment simulation models, or different offsets and different car model parameters, can be quickly generated in a simulated environment. Using reinforcement learning or optimization algorithms, car wash strategies or control parameters of car wash robots can be pre-trained or adjusted.

[0151] In the closed loop of online incremental learning, when the real car wash robot performs its task, it continuously collects trajectory data and sensor feedback from successful car wash cases, failed car wash cases, or boundary success cases. Through a lightweight online learning module, this real-world information can be used to fine-tune the model's network parameters or optimize the car wash strategy to better adapt it to the current real-world car wash environment.

[0152] In addition, by building and updating the knowledge base, prior knowledge obtained from different vehicle models and different environmental scenarios can be successfully configured and stored as reusable templates or feature vectors, which can be used to accelerate the adaptation to similar scenarios in the future.

[0153] The training method in this application embodiment benefits from the high-fidelity simulation engine's advantage of broad scene coverage, supporting combinations of various vehicle models and environments, far exceeding the scope of real-vehicle debugging. Furthermore, the simulation model built within the simulation engine possesses high physical accuracy, such as low simulation error in brush head contact force and high trajectory prediction accuracy. In addition, various car wash environments can be simulated by adjusting highly realistic disturbances; for example, rain and fog disturbances can be calibrated using meteorological data, and optical disturbances can be added based on measured reflectivity.

[0154] The pre-adaptive training module boasts high training efficiency, reducing the tens of hours of manual debugging to just a few hours of automated training combined with simulation, thus significantly improving training efficiency. Furthermore, the generated car wash strategies exhibit high cleaning coverage and robustness even under large parking offsets. This training method also conserves environmental resources through simulation modeling.

[0155] The online incremental learning module can continuously evolve, constantly improving its generation capabilities within a preset period. Furthermore, through simulation and training on real-world data, it can reduce the recurrence rate of historical faults, achieving a fault immunity effect. Incremental learning can accumulate prior knowledge obtained from real-world scenarios; through knowledge base updates and matching, it enables rapid knowledge reuse and rapid deployment.

[0156] Figure 6 This is a flowchart illustrating a car washing method using a car wash robot, provided as an embodiment of this application. The executor of this method can be a car wash robot. The car wash robot includes an initial car wash strategy model or a car wash strategy model obtained after training using the method described in any of the above embodiments, such as... Figure 6 As shown, the car wash method includes:

[0157] S601, in any real car wash scenario, collects the actual cleaning parameters required to clean the target vehicle through the car wash robot's sensors.

[0158] For example, any real-world car wash scenario can be the scenario of a car wash facility where the car wash robot is deployed. The car wash robot may include sensors such as LiDAR and vision cameras, which can collect data from the target vehicle to be cleaned to obtain real-world cleaning parameters. These real-world cleaning parameters can be referred to the relevant descriptions in the above embodiments.

[0159] S602, input the actual cleaning parameters into the initial car wash strategy model or the car wash strategy model to generate the car wash strategy for the target vehicle.

[0160] The car wash strategy for the target vehicle can be understood as a car wash strategy obtained by calculating, analyzing, and predicting real cleaning parameters based on the current real car wash scenario and the current target vehicle, using an initial car wash strategy model or a car wash strategy model. The car wash strategy can be referred to the relevant descriptions in the above embodiments.

[0161] S603 uses a car wash strategy for the target vehicle to clean the target vehicle in any real car wash scenario.

[0162] For example, after determining the car washing strategy for the target vehicle, the car washing robot can control its walking parts, robotic arms, and other moving parts to perform cleaning actions such as spraying, wiping, or vacuuming on the target vehicle according to the car washing strategy, thereby cleaning the target vehicle.

[0163] Since the initial car wash strategy model or car wash strategy model in the embodiments of this application is obtained by training using the training method of any of the above embodiments, the generated car wash strategy can be adapted to any real car wash scenario and the target vehicle to a high extent, so that the car wash robot can achieve better cleaning results after performing cleaning actions according to the current car wash strategy, such as achieving higher cleanliness, shorter total cleaning time and / or fewer collisions.

[0164] The car washing method of the car washing robot provided in this application embodiment can achieve the same technical effect as the model training method of any of the above embodiments of this application. The implementation principle and technical effect are similar, and will not be repeated here.

[0165] Figure 7 This is a schematic diagram of the structure of the model training device provided in the embodiments of this application, as shown below. Figure 7 As shown, this application embodiment provides a model training apparatus, the apparatus comprising:

[0166] Loading module 701 is used to load the car wash environment simulation model, the vehicle simulation model, and the car wash robot simulation model.

[0167] The data acquisition module 702 is used to acquire the simulation cleaning parameters required by the vehicle washing simulation model by simulating the data acquisition operation through the car washing robot simulation model in the environment simulated by the car washing environment simulation model.

[0168] The generation module 703 is used to input the simulation cleaning parameters into the initial car wash strategy model and generate the car wash strategy of the vehicle simulation model.

[0169] The cleaning module 704 is used to control the car wash robot simulation model to perform a car wash operation on the vehicle simulation model in a simulated environment using a car wash strategy, and to obtain the cleaning result.

[0170] The optimization module 705 is used to adjust the parameters of the initial car wash strategy model based on the cleaning results.

[0171] In one possible implementation, the environmental simulation model includes a basic environmental simulation model and at least one variant environmental simulation model; the loading module 701 is specifically used for:

[0172] During the first training round of multiple training rounds, the basic environment simulation model is loaded;

[0173] During training cycles other than the first round, a variant environment simulation model is loaded.

[0174] In one possible implementation, the device further includes a building module for:

[0175] Based on the real car wash environment, an environmental simulation model corresponding to the real car wash environment is constructed to obtain the basic environmental simulation model;

[0176] By adjusting the environmental factor parameters of the basic environmental simulation model, at least one variant environmental simulation model is obtained. The environmental factor parameters include at least one of the following: weather disturbance parameters, illumination disturbance parameters, and obstacle disturbance parameters.

[0177] In one possible implementation, the device further includes an acquisition module, used to: acquire target cleaning parameters corresponding to a target car wash case in a real car wash scenario, wherein the target car wash case is a car wash case in which the cleaning result of the car wash robot does not meet the preset cleaning result among multiple car wash cases in a real car wash scenario.

[0178] The generation module 703 is also used to: generate a car wash strategy corresponding to the target cleaning parameters in the initial car wash strategy model after the target cleaning parameters are input and adjusted;

[0179] The cleaning module 704 is also used to: control the car wash robot simulation model to simulate cleaning the vehicle simulation model corresponding to the target car wash case using the car wash strategy corresponding to the target cleaning parameters, and obtain the cleaning result;

[0180] The optimization module 705 is also used to: adjust the parameters of the initial car wash strategy model after parameter adjustment based on the cleaning results.

[0181] In one possible implementation, the apparatus further includes a distillation module for:

[0182] After the initial car wash strategy model is trained, a new car wash strategy model is obtained by knowledge distillation based on the initial car wash strategy model; wherein the model parameter size of the initial car wash strategy model is larger than that of the car wash strategy model. The model training device provided in this application embodiment can be used to execute the technical solution of the model training method in any of the above embodiments of this application, and its implementation principle and technical effect are similar, and will not be repeated here.

[0183] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 8 As shown, the electronic device of this embodiment may include: at least one processor 801; and a memory 802 communicatively connected to the at least one processor; wherein the memory 802 stores instructions that can be executed by the at least one processor 801, and the instructions are executed by the at least one processor 801 to cause the electronic device to perform the method as described in any of the above embodiments.

[0184] Optionally, the memory 802 can be either standalone or integrated with the processor 801.

[0185] The implementation principle and technical effects of the electronic device provided in this embodiment can be found in the foregoing embodiments, and will not be repeated here.

[0186] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the method of any of the foregoing embodiments.

[0187] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the method of any of the foregoing embodiments.

[0188] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules is only a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules may be combined or integrated into another system, or some features may be ignored or not executed.

[0189] The integrated modules described above, implemented as software functional modules, can be stored in a computer-readable storage medium. These software functional modules, stored in a storage medium, include several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) or processor to execute some steps of the methods of the various embodiments of this application.

[0190] It should be understood that the aforementioned processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. A general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the application can be directly manifested as being executed by a hardware processor, or executed by a combination of hardware and software modules within the processor. The memory may include random access memory (RAM), and may also include non-volatile memory (NVM), such as at least one disk storage device, and may also be a USB flash drive, external hard drive, read-only memory, disk, or optical disc, etc.

[0191] The aforementioned storage media can be implemented from any type of volatile or non-volatile storage device or a combination thereof, such as Static Random-Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read-Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The storage media can be any available medium accessible to general-purpose or special-purpose computers.

[0192] An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Alternatively, the storage medium can be an integral part of the processor. The processor and storage medium can reside within an application-specific integrated circuit (ASIC). Alternatively, the processor and storage medium can exist as discrete components within an electronic device or host device.

[0193] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0194] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0195] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as ROM / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, air conditioner, or network device, etc.) to execute the methods of the various embodiments of this application.

[0196] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0197] Other embodiments of the present application will readily occur to those skilled in the art upon consideration of the specification and practice of the application disclosed herein. This application is intended to cover any variations, uses, or adaptations of the embodiments of this application that follow the general principles of the embodiments of this application and include common knowledge or customary techniques in the art not disclosed in the embodiments of this application.

[0198] It should be understood that the embodiments of this application are not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from their scope. The scope of the embodiments of this application is limited only by the appended claims.

Claims

1. A model training method, characterized in that, The method comprises: loading a car washing environment simulation model, a vehicle simulation model and a car washing robot simulation model; under an environment simulated by the car washing environment simulation model, simulating a collection operation by the car washing robot simulation model to obtain a simulation cleaning parameter required for cleaning the vehicle simulation model; inputting the simulation cleaning parameter into an initial car washing strategy model to generate a car washing strategy for the vehicle simulation model; controlling the car washing robot simulation model to simulate a cleaning operation on the vehicle simulation model under the simulated environment by using the car washing strategy to obtain a cleaning result; adjusting parameters of the initial car washing strategy model according to the cleaning result.

2. The method of claim 1, wherein, The environment simulation model comprises a basic environment simulation model and at least one variant environment simulation model; and the loading of the car washing environment simulation model comprises: in a first round of training in a plurality of rounds of training, loading the basic environment simulation model; in a non-first round of training, loading the variant environment simulation model.

3. The method of claim 2, wherein, The method further comprises: constructing an environment simulation model corresponding to a real car washing environment according to the real car washing environment to obtain the basic environment simulation model; obtaining at least one variant environment simulation model by adjusting environment factor parameters of the basic environment simulation model, the environment factor parameters comprising at least one of weather disturbance parameters, illumination disturbance parameters and obstacle disturbance parameters.

4. The method according to any one of claims 1 to 3, characterized in that, The method further comprises: obtaining a target cleaning parameter corresponding to a target car washing case in a real car washing scene, the target car washing case being a car washing case in which a cleaning result of the car washing robot in the real car washing scene does not meet a preset cleaning result; inputting the target cleaning parameter into the initial car washing strategy model after parameter adjustment to generate a car washing strategy corresponding to the target cleaning parameter; controlling the car washing robot simulation model to simulate cleaning of a vehicle simulation model corresponding to the target car washing case by using the car washing strategy corresponding to the target cleaning parameter, and obtaining a cleaning result; adjusting parameters of the initial car washing strategy model after parameter adjustment according to the cleaning result.

5. The method according to any one of claims 1 to 3, characterized in that, The method further comprises: after training of the initial car washing strategy model is completed, obtaining a car washing strategy model by using a knowledge distillation method according to the initial car washing strategy model, wherein a model parameter size of the initial car washing strategy model is greater than a model parameter size of the car washing strategy model.

6. A car washing method of a car washing robot, characterized by, The car washing robot comprises an initial car washing strategy model or a car washing strategy model obtained by training by using the method according to any one of claims 1 to 5, and the car washing method comprises: in any real car washing scene, collecting a real cleaning parameter required for cleaning a target vehicle by using a sensor of the car washing robot; inputting the real cleaning parameter into the initial car washing strategy model or the car washing strategy model to generate a car washing strategy for the target vehicle; cleaning the target vehicle in the any real car washing scene by using the car washing strategy for the target vehicle.

7. A model training apparatus characterized by comprising: The device comprises: a loading module configured to load a car washing environment simulation model, a vehicle simulation model and a car washing robot simulation model; The collection module is configured to simulate a collection operation by the washing robot simulation model in the environment simulated by the car washing environment simulation model, and obtain simulation cleaning parameters required for cleaning the car simulation model. The generation module is configured to input the simulation cleaning parameters into an initial car washing strategy model, and generate a car washing strategy for the car simulation model. The cleaning module is configured to control the washing robot simulation model to simulate a cleaning operation on the car simulation model in the simulated environment by using the car washing strategy, and obtain a cleaning result. The optimization module is configured to adjust parameters of the initial car washing strategy model according to the cleaning result.

8. An electronic device, comprising: The computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to cause the processor to execute the method in any one of claims 1-6. The computer readable storage medium stores computer execution instructions, and the computer execution instructions are executed by the processor to cause the processor to execute the method in any one of claims 1-6. The computer program is executed by the processor to implement the method in any one of claims 1-6. ​ 9. A computer-readable storage medium, characterized in that, ​ 10. A computer program product, characterised in that, ​

Images