A distributed artificial intelligence system for vehicle-road-cloud collaborative automatic driving

Abstract

The application relates to the technical field of intelligent driving and intelligent traffic management, and provides a distributed artificial intelligence system for vehicle-road cloud cooperative automatic driving, which aims to provide distributed, adaptive and customized artificial intelligence functions for automatic driving vehicles and / or networked automatic driving vehicles. More specifically, the application can dynamically allocate and cooperatively process computing resources, model resources and data resources among vehicle systems, roadside systems, edge systems, traffic control centers and / or cloud platform systems according to running scenarios and resource states, support real-time or dynamic updating of models and data, improve the execution efficiency and robustness of automatic driving in functions such as perception, positioning, fusion, prediction, planning, decision-making and control, enhance the intelligent level of situation awareness, real-time decision-making and automatic driving system level, and realize safety, efficiency and personalized optimization of the automatic driving system.

Description

Technical Field

[0001] This application relates to the fields of autonomous driving and distributed artificial intelligence technology, and in particular to a distributed artificial intelligence system for autonomous driving in a vehicle-road-cloud collaborative architecture. Background Technology

[0002] As autonomous driving technology continues to evolve, the requirements for vehicle safety, real-time performance, reliability, and intelligence are constantly increasing. Existing autonomous driving systems mostly adopt a centralized AI architecture centered on the vehicle, relying primarily on the vehicle's local computing power to complete functions such as perception, localization, fusion, prediction, planning, and control. However, with the increasing complexity of algorithms and the rapid growth of multi-sensor data, a single onboard computing platform faces significant constraints in terms of computing power, energy consumption, and cost, making it difficult to simultaneously guarantee low-latency response while meeting the demands for high accuracy and high reliability. Furthermore, modern connected autonomous vehicles operate in highly dynamic traffic environments, with different road types, weather conditions, traffic flow states, and user preferences placing differentiated demands on system performance. Traditional AI architectures typically employ fixed model deployment methods, lacking the ability to dynamically adjust and personalize configurations based on ODD, trip characteristics, or vehicle characteristics, resulting in low resource utilization efficiency and limited system adaptability and scalability. Furthermore, with the development of vehicle-road cooperation and cloud computing technologies, vehicles, roadside facilities and cloud platforms have the foundation for data interaction and collaborative computing. However, existing technologies have not yet formed a systematic architecture for unified scheduling and dynamic allocation of computing power, data, models and parameters, making it difficult to fully leverage the advantages of vehicle-road-cloud collaboration.

[0003] Therefore, there is an urgent need for an autonomous driving AI system that can achieve distributed intelligent deployment, dynamic model allocation, and real-time updates and optimizations to improve the overall safety, efficiency, and personalized service capabilities of the system. Summary of the Invention

[0004] The purpose of this application is to provide a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, capable of meeting the collaborative needs of connected autonomous vehicles, intelligent connected roads, and cloud platforms, thereby providing comprehensive autonomous driving services. Specifically, in different autonomous driving scenarios, the system provides multiple autonomous driving functions, including perception, prediction, planning and decision-making, control, service, management, operation, and maintenance, through multi-platform collaboration among the vehicle, roadside units (RSU), edge systems, traffic control centers (TCC / TCU), and cloud platforms. This aims to provide enterprises with smarter and more reliable autonomous driving solutions.

[0005] To achieve the above objectives, this application adopts the following technical solution: A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the system comprising an AI allocation subsystem and an operation and maintenance management subsystem, for providing customized and / or personalized AI services to autonomous vehicles based on trip characteristics, route characteristics, road characteristics, Operation Design Domain (ODD) characteristics and / or vehicle characteristics, the system being configured to be deployed on one or more of the following platforms: (1) Vehicle system; (2) Roadside Unit (RSU) system; (3) Edge systems; (4) Traffic Control Center / Traffic Control Unit (TCC / TCU); (5) Cloud platform; The system also includes one or more of the following components: (1) AI model components; (2) Computing power components; (3) Data components; (4) Communication components; (5) Computing components; (6) Power supply components; (7) Perception component.

[0006] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the system includes: (1) The AI ​​allocation subsystem includes one or more of the following modules: a. Computing power allocation module; b. Data allocation module; c. Model allocation module; d. Parameter allocation module.

[0007] (2) An operation and maintenance management subsystem, used to support the determination and implementation of deployment strategies for one or more modules in the AI ​​allocation subsystem, and to deploy them to vehicle systems, RSU systems, edge systems, TCC / TCU and / or cloud platforms. The operation and maintenance management subsystem includes one or more of the following modules: a. Collaboration and Optimization Module; b. Execution module.

[0008] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the AI ​​model component includes one or more of micro-models, small models, medium models, and large models, wherein: (1) Micro-models and small models are deployed and allocated in the vehicle system for local real-time decision-making in the vehicle; (2) Micro-models, small models and / or medium models are deployed and allocated in RSU systems and / or edge systems to optimize traffic flow and provide edge computing support; (3) Micro-models, small models, medium models and / or large models are deployed and distributed in TCC / TCU and / or cloud platforms to provide large-scale data processing, global optimization and training support; The AI ​​model component is configured to train and deploy the vehicle via an edge system and / or cloud platform using one or more of the following training and deployment methods: (1) Method 1: Both the trained model and the data are updated using a dynamic update method; (2) Method 2: The trained model is deployed in the vehicle and updated in real time, while the data is updated using a dynamic update method; (3) Method 3: The trained model is deployed in the vehicle and updated using a dynamic update method, while the data is updated using a real-time update method; (4) Method 4: The trained model and data are both updated in real time using a real-time update method; in: (1) Real-time method updates are configured to apply to data and / or training models with a time range of 0.1 milliseconds to 1 second; (2) The dynamic update method is configured to be applicable to data and / or training models with a time range of 1 second to several days.

[0009] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the AI ​​model component is configured to select combinations and / or integrations of AI models from an AI model library to meet the safety and computing resource requirements of autonomous driving. The combinations and / or integrations of AI models include the following types: (1) A unified AI driving model, which is developed based on the AI ​​model library; (2) An end-to-end or point-to-point model for autonomous driving; (3) A world model for autonomous driving; The AI ​​model library includes one or more of the following models: (1) Deep learning models; (2) AI intelligent agents; (3) Reinforcement learning model; (4) Generative models; (5) Basic model; (6) Swarm intelligence model; The AI ​​model is used to achieve one or more of the following functions: (1) Sensory function; (2) Predictive function; (3) Planning and decision-making functions; (4) Control functions; (5) Service functions; (6) Management functions; (7) Operational functions; (8) Maintenance function.

[0010] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the data component includes one or more of the following data: (1) Vehicle data, including vehicle status, mechanical condition, sensor data, historical driving patterns and / or predictive maintenance indicators; (2) ODD data, including weather conditions, road conditions, infrastructure data and / or external hazards; (3) System data, including metadata, transaction logs and / or system configuration parameters.

[0011] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the communication component is configured as follows: (1) Provide real-time data exchange, including one or more of advanced cellular networks (4G, 5G, 5.5G, 6G), Bluetooth, Wi-Fi and / or satellite communications, and provide AI systems that can be deployed and updated in vehicles using real-time update methods; (2) Provide dynamic data exchange, including one or more of advanced cellular networks (4G, 5G, 5.5G, 6G), Bluetooth, Wi-Fi and / or satellite communications, and provide AI systems that can be deployed and updated in vehicles using dynamic update methods; (3) Use one or more low-end cellular communication networks, including 3G and / or digital broadcasting, to provide real-time and / or dynamic data exchange, ensuring that the AI ​​system can operate reliably anytime and anywhere when advanced cellular communication is unavailable.

[0012] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the computing power allocation module is used to provide customized or personalized computing resource allocation, including one or more of the following sub-modules: (1) Static computing power allocation submodule, which includes one or more of the following units: a. Static computing power identification unit, used to identify available vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources; b. Computing power demand identification unit, which estimates, predicts and plans the computing power demand of the vehicle during the trip; c. Computing power gap calculation unit: calculates the difference between the identified available computing power on the vehicle and the determined computing power requirement, as the computing power gap; d. Computing power allocation unit, which allocates computing power of expected vehicle computing resources, expected RSU computing resources, expected edge computing resources, expected cloud computing resources and / or expected TCC computing resources to compensate for computing power gaps; (2) A dynamic computing power allocation submodule, wherein the dynamic computing power allocation submodule includes: a. One or more of the supply-side internal change identification unit and the available computing power identification unit; b. One or more of the external change identification unit and computing power demand identification unit on the demand side; c. Gap calculation unit and distillation unit; d. One or more of the following computing power allocation units for vehicle trips: vehicle computing power allocation unit, path computing power allocation unit, road computing power allocation unit, and ODD computing power allocation unit.

[0013] The functions of each unit are as follows: a. An internal change recognition unit is used to monitor the current in-vehicle conditions and system parameters in order to detect and quantify changes during dynamic operation; b. Available computing power identification unit, used to determine the available computing resources of the vehicle in real time, including RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources; c. External change identification unit, used to detect and assess changes in external environmental conditions, including event-specific conditions and unexpected scenarios; d. Computing power requirement identification unit, which estimates the vehicle's computing power requirement under the current operating conditions based on the identified in-vehicle conditions and external environmental conditions; e. Gap calculation unit, which calculates the difference between the identified available computing power on the vehicle and the computing power requirement as the computing power deficiency; f. Distillation unit: Based on specific vehicle, route, road and ODD conditions, customized computing power allocation is implemented to achieve optimized, dedicated and efficient allocation of computing resources; g. Vehicle computing power allocation unit, used to allocate the computing power of the required vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources related to a specific driving route to make up for insufficient computing power; h. Path computing power allocation unit, used to allocate the computing power of the required vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources related to a specific driving route to make up for insufficient computing power; i. Road computing power allocation unit, used to allocate the required vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources according to specific road conditions, in order to make up for insufficient computing power; j. ODD computing power allocation unit, used to allocate the required vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources according to specific ODD conditions, in order to make up for insufficient computing power.

[0014] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the computing power allocation module is configured to be deployed on a vehicle computing platform, an RSU computing platform, an edge computing platform, a TCC computing platform, and / or a cloud computing platform, for allocating the required vehicle computing power, RSU computing power, edge computing power, TCC computing power, and / or cloud computing power to the vehicle-road-cloud system or the vehicle-edge-cloud system, wherein: (1) The vehicle computing platform is configured to perform real-time perception, decision-making and control using onboard units; (2) The RSU computing platform is configured to provide vehicle-to-infrastructure (V2I) communication, sensor data fusion, and localized processing; (3) The edge computing platform is configured to provide low-latency data processing and collaborative sensing through mobile edge computing (MEC) servers; (4) The TCC computing platform is configured to coordinate traffic management and optimize the operation of the entire network; (5) The cloud computing platform is configured to handle large-scale data storage, model training, and predictive analytics; The computing power allocation module optimizes and refines computing power allocation through a distillation method, providing customized or personalized computing power allocation schemes with vehicle-specific, route-specific, road-specific, and / or ODD-specific characteristics. The distillation method includes: (1) Knowledge extraction, used to identify and capture key computational patterns of specific vehicles, specific roads, specific routes and / or specific ODDs; (2) Compression steps are used to reduce redundant calculations and the allocation of computing resources.

[0015] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the data allocation module is used to allocate vehicle data and ODD data to one or more of the following scenarios: (1) Vehicle-specific scenarios; (2) Route-specific scenarios; (3) Road-specific scenarios; (4) ODD specific scenarios; (5) Specific scenarios of the trip; The data allocation module is configured for use in one or more of the following applications: (1) In-vehicle applications can distribute data from one or more vehicles to one or more targets in the vehicle itself, RSU, edge computing nodes, TCC and the cloud; (2) External applications enable multi-directional transmission between vehicles, RSUs, edge nodes, TCCs and cloud platforms, and support data transmission from any one or more of the above nodes to one or more other nodes.

[0016] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the data allocation module includes one or more of the following units: (1) Data acquisition unit, including the following sub-units: a. Onboard data acquisition subunit, used to collect onboard sensor data and vehicle status data, including speed, acceleration, steering angle, braking, engine load, battery power and / or gear data; b. External data acquisition subunit, used to collect data from other vehicles, RSU, TCC / TCU and / or the cloud; (2) Data measurement unit, including the following components: a. A detectable data recognition component used to identify whether data can be detected; b. Data size determination component, used to determine the data size required for training or inference; c. Model size determination component, used to determine the model size required for training or inference; d. Update frequency determination component, used to determine the update frequency of data; (3) Big data processing unit, including the following components: a. Data quality measurement components are used to identify data quality, including missing values, duplicates, and validity; b. Data preprocessing components are used to remove or fill missing values, remove duplicates, and reformat the dataset into a valid format; c. Data fusion component, used to fuse data after data quality measurement and data preprocessing; (4) A data allocation unit, used to update the allocation process according to real-time system requirements to ensure reliable and efficient data distribution, the data allocation unit includes the following components: a. Data management component, used to queue all data to be transmitted, ensuring proper organization and prioritization of the data; b. A data allocation strategy component, used to adjust the appropriate allocation strategy based on the measurement results of the data measurement unit; (5) A distillation preparation unit, used to prepare training and validation data for vehicle-specific models, route-specific models, road-specific models, ODD-specific models, and trip-specific models, wherein the distillation preparation unit includes the following components: a. Vehicle-specific data preparation component, used to prepare training and validation data for large-scale model distillation in vehicle-specific scenarios; b. Route-specific data preparation component, used to prepare training and validation data for large model distillation in route-specific scenarios; c. Road-specific data preparation component, used to prepare training and validation data for large model distillation in road-specific scenarios; d. ODD-specific data preparation components are used to prepare training and validation data for large model distillation in specific ODD scenarios; e. Trip-specific data preparation component, used to prepare training and validation data for large-scale model distillation in trip-specific scenarios; Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the data allocation unit is configured to dynamically adjust the data allocation strategy by comprehensively considering factors such as data urgency, network conditions, and processing capabilities to optimize transmission efficiency, wherein: (1) The urgency of the data is determined based on the criticality of real-time decision information, and higher priority is given to safety-critical data such as collision warnings, sensor anomalies and emergency braking events; (2) The network conditions are assessed based on factors such as available bandwidth, latency and signal strength, and the data transmission rate and priority are dynamically adjusted to maintain efficient communication between vehicles, infrastructure and the cloud. (3) The processing capacity is assessed based on the vehicle and cloud computing load and available resources. Highly complex tasks are intelligently downloaded to the cloud for processing, while ensuring that time-sensitive calculations are executed locally to obtain optimal performance. The data allocation unit is configured for one or more of the following applications: (1) In-vehicle applications are used to prioritize the allocation of data required for real-time vehicle response, including perception, positioning, planning and control, to ensure that safety-critical data and time-sensitive data are processed first, while non-urgent data is postponed to be processed or transmitted outside the vehicle. (2) External applications, used to distribute non-time-sensitive or large-capacity data to other vehicles, RSUs, TCCs / TCUs and the cloud for batch processing, long-term storage or large-scale analysis.

[0017] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the model allocation module is configured to allocate AI models from the AI ​​model repository according to the following formula, so as to enable the autonomous driving system to achieve a specific level of automation:

[0018] in: (1) As the output of the model allocation module, it represents any combination and / or integration of AI models from the AI ​​model repository; (2) It is a set of in-vehicle AI models selected based on one or more of the safety requirements and computing power capabilities related to in-vehicle autonomous driving systems. It is certain How to collaborate to produce The function, This refers to one of the following models: recognition model, perception model, localization model, fusion model, prediction model, planning model, decision-making model, and / or control model; (3) It is a set of external AI models that execute under conditions that satisfy vehicle computing power constraints, data constraints, power constraints, and / or communication constraints. It is certain How to collaborate to produce The function; where This refers to one of the following models: recognition model, perception model, localization model, fusion model, prediction model, planning model, decision-making model, and / or control model; (4) It is a certainty and How to collaborate to produce The function must achieve a certain degree of automation while simultaneously meeting the safety and computing resource requirements of autonomous driving. The model allocation module executes the following AI model allocation method: (1) Determine ; (2) Determine a set of in-vehicle AI models through AI model allocation method. And a set of car exterior AI models ; (3) Determine a set of AI models .

[0019] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the model allocation module is configured to allocate one or more of the following models in a specific deployment scenario: (1) Cloud-based large models, which are deployed in the cloud, include one or more of the following: sensing models, perception models, positioning models, fusion models, prediction models, planning models, decision-making models and / or control models; (2) Edge-end distillation model, deployed at the edge, includes one or more of the following: sensing model, perception model, localization model, fusion model, prediction model, planning model, decision model, and control model, and also contains knowledge distilled from large AI models; (3) Vehicle-mounted small models, deployed on the vehicle, including one or more of the following: sensing model, perception model, positioning model, fusion model, prediction model, planning model, decision-making model, and control model; (4) Vehicle-mounted micro-models, deployed on the vehicle, including one or more of the following: sensing model, perception model, positioning model, fusion model, prediction model, planning model, decision-making model, and control model.

[0020] The model allocation module is used to realize distributed model and data stream interaction between the vehicle end, road end, edge end and cloud end, and specifically implements the following methods: (1) The vehicle is equipped with a lightweight AI model to continuously obtain real-time data, update parameters and / or update the model from the road end and edge end, and send the updated data back to the road end and edge end; (2) Deploy medium-scale AI models at the roadside and edge to process data, parameters and models from the vehicle, and periodically send updated data, parameters and models to the vehicle; (3) The road end and edge end process data, parameters and models from the cloud, and periodically send updated data, parameters and models to the cloud; (4) Run large-scale AI models in the cloud, process data, parameters and models from the road end and edge end, and regularly send updated data, parameters and models to the road end and edge end; (5) Perform model training, inference correction, and backup operations in the cloud; (6) The vehicle, road, edge and cloud terminals are synchronized with each other to provide accurate processing.

[0021] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the model allocation module is configured to implement the training and deployment method for autonomous driving, including one of the following sub-modules: (1) A vehicle-edge-cloud collaborative model deployment and allocation submodule, configured to be deployable on one or more of an in-vehicle computing platform, an edge computing platform, and a cloud computing platform, wherein the vehicle-edge-cloud collaborative model deployment and allocation submodule includes one or more of the following units: a. Input analysis unit, used to analyze the input data and scenario requirements of deployment tasks; b. Hardware capability verification unit, used to verify whether the computing hardware of the vehicle platform, edge platform or cloud platform meets the requirements for model operation; c. Cloud-edge collaboration unit, used to coordinate model deployment tasks between the cloud and the edge; d. Assignment execution unit, used to perform the final assignment of the model to vehicles, edge, or cloud platforms; e. Vehicle-specific model allocation unit; f. Route-specific model allocation unit; g. Distillation unit, used to distill and compress the model to adapt it for edge or vehicle-end deployment; (2) A vehicle-edge-cloud collaborative model training and allocation submodule, configured to be deployed on an in-vehicle computing platform, an edge computing platform, and a cloud computing platform, wherein the vehicle-edge-cloud collaborative model training and allocation submodule comprises one or more of the following units: a. Input analysis unit, used to analyze the input data and scenario requirements of the training task; b. Hardware capability verification unit, used to verify whether the computing hardware of the vehicle platform, edge platform or cloud platform meets the model training requirements; c. Cloud-edge collaboration unit, used to coordinate model training tasks between the cloud and the edge; d. Assign execution units to perform the assignment of training tasks to vehicles, edge, or cloud platforms; e. Resource allocation unit; f. Scene-aware model allocation unit; g. Vehicle-specific model allocation unit; h. Route-specific model allocation unit; i. Distillation unit; Preferably, the training and deployment method is configured to employ a real-time and / or dynamic update mechanism to synchronously or stepwise train and deploy the AI ​​driving model, achieved through one of the following real-time / dynamic update methods: (1) Method 1: Synchronous real-time training and deployment are carried out within an update interval of 1 millisecond to 1 second, so that the parameters and running version of the AI ​​driving model can be continuously optimized and applied in real time; (2) Method 2: Synchronous dynamic training and deployment are carried out within update intervals ranging from 1 second to several days, allowing AI driving models to be jointly updated and retrained according to a flexible schedule that adapts to larger datasets or resource availability; (3) Method 3: A separate training and deployment mechanism is adopted. When the interval between dynamic training ranges from 1 second to several days, the deployment process is executed in real time (within the range of 1 millisecond to 1 second). In this way, the running AI driving model can be continuously updated and kept up-to-date. The principle is to use the incremental parameter updates generated by the low-frequency but wider-coverage training cycle. (4) Method 4: Real-time training and dynamic deployment are carried out simultaneously, in which the parameters of the artificial intelligence driving model are updated in real time at intervals of 1 millisecond to 1 second, and then deployed at more flexible cycles ranging from 1 second to several days.

[0022] Preferably, in the aforementioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the model allocation module is configured to be deployed in one or more of vehicles, RSUs, edge devices, TCCs / TCUs, and / or cloud platforms, and provides customized and / or personalized model allocation schemes for vehicle operation. These customized and / or personalized model allocation schemes can be optimized for specific trips, routes, roads, ODDs, and / or vehicles using distillation methods. The distillation method is configured to support hierarchical knowledge transfer across cloud, edge, and vehicle terminals in the following ways: (1) Cloud-based knowledge extraction: Large-scale AI models process massive amounts of sensor data to generate soft tags and inference backups; (2) Edge model distillation: The distilled medium-sized AI model optimizes knowledge through feature-based distillation and periodic updates; (3) Vehicle model adaptive distillation: Small AI models receive compressed and quantized knowledge to make real-time decisions; (4) Regular synchronization and model backup: The vehicle end and the edge end are updated synchronously, and the cloud acts as a failover backup when needed.

[0023] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the parameter allocation module is configured to be deployed on one or more of the vehicle, RSU, edge device, TCC / TCU and / or cloud platform, and to implement parameter allocation and updating on one or more of the vehicle, RSU, edge device, TCC / TCU and / or cloud platform. The parameter allocation module includes one or more of the following sub-modules: (1) Parameter acquisition submodule, including one or more of the following units: a. Parameter acquisition unit, used to collect AI model parameters, vehicle parameters and ODD parameters that need to be updated in one or more systems, including vehicle system, road system, edge system, TCC / TCU and cloud system; b. Parameter statistics unit, used to process and back up the types and scales of parameters that need to be updated; (2) Parameter allocation submodule, including one or more of the following units: a. Computing resource unit, used to calculate the available computing power for parameter updates in one or more systems, including vehicle systems, road systems, edge systems, TCC / TCU and cloud systems; b. Parameter allocation unit, used to provide parameter updates to one or more of the following systems based on the availability of computing resources and the type and scale of the parameters to be updated: vehicle system, road system, edge system, TCC / TCU and cloud system; (3) The parameter update submodule includes one or more of the following units: a. Vehicle parameter update unit, used to update one or more execution parameters among vehicle status parameters and operating information parameters; b. ODD parameter update unit, used to update one or more execution parameters among road parameters, object parameters, and driving environment parameters; c. AI model parameter update unit, used to update one or more execution parameters among the basic model parameters, training and optimization parameters, and deployment and adaptation parameters; d. Parameter update progress unit, used to record the progress of parameter updates; The parameter allocation module is configured to apply customized and / or personalized parameter allocation modules for one or more vehicle-specific scenarios, trip-specific scenarios, route-specific scenarios, road-specific scenarios, and operation design domain-specific scenarios during vehicle operation.

[0024] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the parameters include: (1) AI model parameters, comprising a set of learnable numerical values ​​that define how the AI ​​model processes data and generates output; configured to enable the AI ​​model to perform tasks and driving functions, including one or more of perception, recognition, localization, fusion, prediction, planning, decision-making, and control functions; the AI ​​model parameters include one or more of the following elements: a. Driving model elements; b. Vehicle-based model elements; c. Model elements based on the exterior of the vehicle; d. Model combination elements; The AI ​​model parameters include one or more of the following parameters: a. Basic model parameters; b. Training and optimizing parameters; c. Deployment and adaptation parameters; (2) Vehicle parameters, including one or more of the following: a. Vehicle status parameters; b. Operational information parameters; (3) ODD parameters, including one or more of the following: a. Road parameters; b. Object parameters; c. Driving environment parameters.

[0025] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the parameter allocation submodule uses the following allocation method to allocate one or more of the following: vehicle parameters, ODD parameters, and AI model parameters: (1) The computing power resource unit calculates the available computing power for parameter updates in each system; (2) The parameter allocation unit obtains the parameter type and scale through the communication component; (3) The parameter allocation unit allocates parameters to system groups according to parameter type and scale; (4) The parameter allocation unit allocates parameters to each system according to the available computing power of each system; The parameter update submodule uses the following update methods to update one or more of the vehicle parameters, ODD parameters, and AI model parameters: (1) The vehicle parameter update unit, ODD parameter update unit and / or AI model parameter update unit collect parameter update requirements; (2) The parameter update progress unit records the parameter update status and performs parameter update after stopping when the vehicle is in an unsafe operating state; (3) The vehicle parameter update unit, ODD parameter update unit and / or AI model parameter update unit update parameters according to the computing power availability of each system; (4) The parameter update progress unit records the parameter update status; (5) When there is an update requirement in the next interval, the parameter update submodule will execute the update method again; Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the coordination and optimization module is used to manage one or more of the AI ​​model components, computing power components, and data components, and includes the following sub-modules: (1) A collaboration submodule, used to organize, manage, and prioritize the execution of tasks based on allocated resources, wherein the collaboration submodule includes one or more of the following units: a. Task Management Unit, used to manage the order and timing of task execution on one or more of the following platforms: vehicle system, RSU system, edge system, TCC / TCU, and cloud platform; b. Resource matching unit, used to allocate tasks to computing resources according to task requirements; c. Priority processing unit, used to balance tasks according to priority level, to ensure that critical tasks receive immediate attention, and to optimize the overall resource utilization of the system, wherein the priority level is classified according to factors such as task urgency, time sensitivity, system impact and resource availability; d. Load balancing unit, used to distribute workloads across one or more of the following platforms: vehicle systems, RSU systems, edge systems, TCC / TCU and cloud platforms, to avoid bottlenecks and underutilization of resources; e. Cluster management unit, used to oversee coordination between distributed systems, ensuring effective communication between one or more of the following platforms: vehicle systems, RSU systems, edge systems, TCC / TCU, and cloud platforms; (2) An optimization submodule, used to ensure that allocated resources are used most efficiently and tasks are executed with optimal performance, wherein the optimization submodule includes one or more of the following units: a. Resource expansion unit, used to adjust computing resources according to real-time demand; b. Energy consumption optimization unit, used to minimize energy consumption while meeting performance requirements; c. Cost optimization unit, used to allocate cloud and local resources in the most cost-effective manner; d. Performance tuning unit, used to fine-tune task execution parameters to achieve optimal resource utilization.

[0026] Preferably, in the above-mentioned distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, the coordination and optimization module operates in the following manner: (1) The collaborative submodule dynamically adjusts the execution priority of the vehicle's AI driving model based on vehicle operating status data (including acceleration, braking and navigation requirements); (2) The collaborative submodule prioritizes artificial intelligence decisions in autonomous driving navigation by dynamically allocating computing resources among the perception, planning and control systems; (3) Optimize the sub-module to dynamically reallocate computing power between safety-critical AI driving tasks (such as collision avoidance) and non-essential tasks to maximize efficiency; (4) The optimization sub-module minimizes AI decision-making delay by intelligently allocating tasks between on-board computing resources and external computing processing.

[0027] This application provides a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, which has at least the following beneficial effects: This application discloses a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, which optimizes the allocation of computing resources, AI models, and data in the autonomous driving system, significantly improving the system's performance, reliability, and adaptability. Through dynamic computing power allocation and personalized AI services, the system can flexibly adjust resources according to vehicle, path, road, and ODD conditions, ensuring the priority execution of critical tasks and improving decision-making efficiency and safety. Simultaneously, cross-platform collaboration and distributed optimization ensure efficient data and resource flow between the vehicle, road, and cloud, ensuring stable system operation in changing traffic environments and meeting the autonomous driving needs of different scenarios. Attached Figure Description

[0028] 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.

[0029] Figure 1 A key component of a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving was demonstrated.

[0030] Figure 2 An exemplary subsystem architecture for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving is demonstrated.

[0031] Figure 3 An exemplary modular structure for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving is demonstrated.

[0032] Figure 4 This paper demonstrates a system architecture for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving.

[0033] Figure 5 A schematic diagram of the model allocation module is shown.

[0034] Figure 6 A diagram illustrating the limitations of a vehicle's computing power.

[0035] Figure 7 An exemplary structure of the communication component of an AI system is shown.

[0036] Figure 8 The architecture of the static computing power allocation submodule in the vehicle-road / edge-cloud system is demonstrated.

[0037] Figure 9 The architecture of the vehicle-road / edge-cloud dynamic computing power allocation submodule was demonstrated.

[0038] Figure 10 The workflow for allocating static computing resources for planning purposes is demonstrated.

[0039] Figure 11The workflow for allocating real-time vehicle operation dynamic computing capabilities was demonstrated.

[0040] Figure 12 The architecture of the vehicle external data allocation submodule is shown.

[0041] Figure 13 The architecture of the vehicle data distribution submodule is shown.

[0042] Figure 14 The architecture of the data acquisition unit is shown.

[0043] Figure 15 The architecture of the data measurement unit is shown.

[0044] Figure 16 The architecture of the big data preprocessing unit is shown.

[0045] Figure 17 The architecture of the data allocation unit is shown.

[0046] Figure 18 The structure of the distillation preparation unit is shown; Figure 19 The hardware structure of the data allocation module was demonstrated; Figure 20 The workflow of the big data reprocessing unit was demonstrated; Figure 21 The workflow of the data allocation unit used for training data allocation in the vehicle-road-cloud system is demonstrated. Figure 22 The workflow of the data allocation unit for inference data allocation in the vehicle-road-cloud system is demonstrated; Figure 23 It demonstrates the structure of a large-scale AI model; Figure 24 The structure of a medium-sized AI model was demonstrated. Figure 25 The structure of a small-scale AI model was demonstrated. Figure 26 The structure of the miniature AI model was demonstrated; Figure 27 (a) shows the structure of the AI ​​model repository; Figure 27(b) illustrates the mathematical framework for the model assignment module used in autonomous driving systems; Figure 28 The deployment process of the model allocation module is demonstrated; Figure 29 The configuration process of the model allocation module is demonstrated; Figure 30 This demonstrates an example of deploying and configuring the model allocation module; Figure 31This demonstrates an example of the deployment and configuration of the model allocation module at S1 level (Level 1 according to the SAE autonomous driving level definition); Figure 32 A small AI model was demonstrated; Figure 33 A large-scale AI model was demonstrated; Figure 34 A distillation-type intermediate AI model is demonstrated; Figure 35 It demonstrates the interaction process between a small AI model in a vehicle, a medium-sized AI model on the edge, and a large AI model in the cloud. Figure 36 It demonstrates the L1 level architecture; Figure 37 It demonstrates an L2-level architecture; Figure 38 It showcased the L3-level architecture; Figure 39 It showcased an L4-level architecture; Figure 40 The workflow of a dynamic vehicle control system that uses onboard, roadside, and cloud computing to operate in tandem was demonstrated. Figure 41 It demonstrates a model deployment and allocation submodule for vehicle-edge-cloud collaboration; Figure 42 It demonstrates a model training allocation submodule for vehicle-edge-cloud collaboration; Figure 43 The component structure of the parameter allocation module is shown; Figure 44 The process of using the parameter allocation module to provide parameter allocation and updates during vehicle operation is demonstrated; Figure 45 This demonstrates the process of using the parameter acquisition submodule to collect and / or statistically evaluate parameters; Figure 46 An exemplary parameter allocation process is shown; Figure 47 This demonstrates an exemplary parameter allocation process for a parameter scale exceeding 1 billion. Figure 48 An exemplary parameter allocation process is shown for parameter sizes between 100 million and 1 billion. Figure 49 An exemplary parameter allocation process is shown for parameter sizes between 5 million and 100 million. Figure 50 An exemplary parameter allocation process is shown for a parameter size of less than 5 million. Figure 51 An exemplary parameter update flowchart is shown; Figure 52 The component structure of the collaboration and optimization module is shown; Figure 53 The component structure of the collaborative submodule is shown; Figure 54 It demonstrates the component structure of the optimized submodule; Figure 55 The flowchart for the collaboration and optimization module is shown.

[0047] 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 concepts of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0048] First, the reference numerals in the attached figures are explained as follows: 101. Artificial intelligence systems for autonomous driving; 102. Computing power components; 103. AI model components; 104. Data components; 201. A distributed artificial intelligence system for vehicle-road-cloud collaborative autonomous driving; 202. AI allocation subsystem; 203. Operation and maintenance management subsystem; 301. A distributed artificial intelligence system for vehicle-road-cloud collaborative autonomous driving; 302. AI allocation subsystem; 303. Operation and maintenance management subsystem; 304. Computing power allocation module; 305. Data allocation module; 306. Model allocation module; 307. Parameter allocation module; 308. Collaboration and optimization module; 309. Execution module; 401. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving; 402. Computing power allocation module; 403. Data allocation module; 404. Model allocation module; 405. Parameter allocation module; 406. Deployment; 407. Vehicle system; 408. RSU system; 409. Edge system; 410. TCC / TCU; 411. Cloud platform; 501. Model Allocation Module; 502. Large Model; 503. Medium Model; 504. Small Model; 505. Micro Model; 506. Deployment; 507. Vehicle System; 508. RSU System and Edge System; 509. TCC / TCU and Cloud Platform; 601. Limitations of vehicle computing power; 602. Power supply; 603. Cooling system; 604. Cost; 605. Computing processing power; 701. Communication components of AI systems (advanced cellular networks available); 702. Communication components of AI systems (advanced cellular networks not available); 703. 3G; 704. 4G; 705. 5G; 706. 6G; 707. Digital broadcasting; 708. Wi-Fi; 709. Bluetooth; 710. Satellite; 801. Static computing power allocation submodule; 802. Static computing power identification unit; 803. Computing power demand identification unit; 804. Gap calculation unit; 805. Computing power allocation unit; 806. Vehicle computing platform; 807. RSU computing platform; 808. Edge computing platform; 809. Cloud platform; 810. TCC computing platform; 811. Required vehicle computing resources; 812. Required RSU computing resources; 813. Required edge computing resources; 814. Required cloud computing resources; 815. Required TCC computing resources; 816. Autonomous vehicle computing resources; 817. Idle vehicle computing resources; 818. Deployment; 819. Allocation; 901. Dynamic computing power allocation submodule; 902. Supply; 903. Internal change identification unit; 904. Available computing power identification unit; 905. Demand; 906. External change identification unit; 907. Computing power demand identification unit; 908. Gap calculation unit; 909. Distillation unit; 910. Vehicle computing power allocation unit; 911. Path computing power allocation unit; 912. Road computing power allocation unit; 913. Onboard computing platform; 914. RSU computing platform; 915. Edge computing platform; 916. Cloud platform; 917. TCC-based computing platform; 918. Required vehicle computing resources; 919. Required RSU computing resources; 920. Required edge computing resources; 921. Required cloud computing resources; 922. Required TCC computing resources; 923. Autonomous vehicle computing resources; 924. Idle vehicle computing resources; 925. Deployment; 926. Allocation; 1201. External Data Distribution Submodule; 1202. Data Acquisition Unit; 1203. Data Measurement Unit; 1204. Big Data Preprocessing Unit; 1205. Data Distribution Unit; 1206. Distillation Preparation Unit; 1207. Vehicle; 1208. RSU; 1209. Edge; 1210. Cloud; 1211. TCC; 1212. Vehicle; 1213. RSU; 1214. Edge; 1215. Cloud; 1216. TCC; 1217. Distribution; 1218. Deployment; 1301. Vehicle Data Distribution Submodule; 1302. Data Acquisition Unit; 1303. Data Measurement Unit; 1304. Big Data Preprocessing Unit; 1305. Data Distribution Unit; 1306. Distillation Preparation Unit; 1307. Vehicle; 1308. Vehicle; 1309. RSU; 1310. Edge; 1311. Cloud; 1312. TCC; 1313. Distribution; 1314. Deployment; 1401, Data Acquisition Unit; 1402, External Data Acquisition Subunit; 1403, In-vehicle Data Acquisition Subunit; 1501. Data Measurement Unit; 1502. Detectable Data Recognition Component; 1503. Data Size Determination Component; 1504. Model Size Determination Component; 1505. Update Frequency Determination Component; 1601. Data preprocessing unit; 1602. Data quality measurement component; 1603. Data preprocessing component; 1604. Data fusion component; 1701. Data allocation unit; 1702. Data management component; 1703. Data allocation strategy component; 1801. Distillation preparation unit; 1802. Vehicle-specific data preparation component; 1803. Route-specific data preparation component; 1804. Road-specific data preparation component; 1805. ODD-specific data preparation component; 1806. Trip-specific data preparation component; 1901. Data distribution system hardware; 1902. Data storage; 1903. Memory; 1904. Central processing unit (CPU); 1905. Graphics processor (GPU); 2301. Large-scale AI model; 2302. Recognition; 2303. Perception; 2304. Localization; 2305. Fusion; 3206. Prediction; 2307. Decision-making; 2308. Planning; 2309. Control; 2310. Specific vehicle model allocation unit; 2311. Specific route model allocation unit; 2312. Distillation unit; 2401. Medium-sized AI model; 2402. Recognition; 2403. Perception; 2404. Localization; 2405. Fusion; 2406. Prediction; 2407. Decision-making; 2408. Planning; 2409. Control; 2410. Specific vehicle model allocation unit; 2411. Specific route model allocation unit; 2412. Distillation unit; 2501. Small AI Model; 2502. Recognition; 2503. Perception; 2504. Localization; 2505. Fusion; 2506. Prediction; 2507. Decision Making; 2508. Planning; 2509. Control; 2510. Vehicle-Specific Model Allocation Unit; 2511. Route-Specific Model Allocation Unit; 2512. Distillation Unit; 2601. Miniature AI Model; 2602. Recognition; 2603. Perception; 2604. Localization; 2605. Fusion; 2606. Prediction; 2607. Decision Making; 2608. Planning; 2609. Control; 2610. Specific Vehicle Model Allocation Unit; 2611. Specific Route Model Allocation Unit; 2612. Distillation Unit; 2701. AI Model Repository; 2702. Deep Learning Models; 2703. AI Agent Models; 2704. Reinforcement Learning Models; 2705. Generative Models; 2706. Basic Models; 2707. Swarm Intelligence Models; 3001. Vehicle system; 3002. Traffic control center; 3003. Edge node; 3004. Roadside unit system; 3005. Cloud platform; 3006. Sensing; 3007. Perception; 3008. Positioning; 3009. Fusion; 3010. Prediction; 3011. Planning; 3012. Decision-making; 3013. Control; 3101. Vehicle system; 3102. Sensing; 3103. Perception; 3104. Localization; 3105. Fusion; 3106. Prediction; 3107. Planning; 3108. Decision-making; 3109. Control; 3110. Roadside unit system; 3111. Cloud platform; 3201. Small-scale AI model; 3202. Application; 3203. Perception; 3204. Localization; 3205. Prediction; 3206. Decision-making; 3207. Control; 3208. Fusion; 3209. Deployment location; 3210. Vehicle system; 3211. Roadside unit system; 3212. Cloud platform; 3213. Model parameters; 3214. Less than 1 billion; 3215. Safety-critical task; 3216. Priority; 3217. Hardware requirements; 3218. Vehicle computing platform; 3219. Input; 3220. Vehicle perception data; 3221. External data; 3222. CAN bus vehicle status data; 3223. Cloud communication data; 3301. Large-scale AI models; 3302. Applications; 3303. Basic applications of AI models; 3304. Perception; 3305. Localization; 3306. Prediction; 3307. Decision-making; 3308. Control; 3309. Fusion; 3310. Enhanced applications of AI models; 3311. Macro-level prediction; 3312. Parallel decision-making; 3313. Multi-source data fusion; 3314. Big data analysis; 3315. Traffic signal control; 3316. Deployment location; 3317. Cloud platform; 3318. Model parameters; 3319. Greater than 7 billion; 3320. Priority; 3321. Based on intelligent agents; 3322. Hardware requirements; 3323. Cloud cluster; 3324. Input; 3325. Multi-source detection data; 3326. Traffic memory data; 3327. Human commands; 3328. Cloud platform built-in data; 3401. Distilled Medium-Level AI Models; 3402. Applications; 3403. Basic Applications of AI Models; 3404. Perception; 3405. Localization; 3406. Prediction; 3407. Decision Making; 3408. Control; 3409. Fusion; 3410. Enhanced Applications of Medium-Level AI Models; 3411. Meso-Level Prediction; 3412. Parallel Decision Making; 3413. Cooperative Control; 3414. Cooperative Perception; 3415. Applications of Large-Scale AI Model Inheritance; 3416. Parallel Decision Making; 3417. Multi-source data fusion; 3418. Traffic signal control; 3419. Deployment location; 3420. Cloud platform; 3421. Roadside unit system; 3422. Model parameters; 3423. 1 billion to 7 billion; 3424. Priority; 3425. Agent-based; 3426. Hardware requirements; 3427. Edge platform; 3428. Input; 3429. Multi-source detection data; 3430. Large model knowledge; 3431. Human instructions; 3432. Edge-embedded data; 3501. Cloud; 3502. Large-scale AI model; 3503. Inference result feedback; 3504. Inference correction knowledge training; 3505. Edge; 3506. Distilled medium-sized AI model; 3507. Real-time data acquisition; 3508. Periodic synchronous update, retraining weights, model inference results; 3509. Vehicle; 3510. Small-scale AI model; 3511. Model backup, in case of medium-sized model failure; 3512. Vehicle perception data, vehicle status data; 3601. In-vehicle driving model; 3602. Perception; 3603. Recognition; 3604. Localization; 3605. Fusion; 3606. Prediction; 3607. Decision-making; 3608. Planning; 3609. Control; 3701. Vehicle-mounted driving model; 3702. Perception; 3703. Recognition; 3704. Localization; 3705. Fusion; 3706. Prediction; 3707. Decision-making; 3708. Planning; 3709. Control; 3710. Beyond-line-of-sight perception; 3711. Data fusion; 3712. Prediction; 3713. Decision-making; 3714. Planning; 3715. Control; 3716. Raw data; 3717. Fuded data; 3718. Processed data; 3801. In-vehicle driving model; 3802. Perception; 3803. Recognition; 3804. Localization; 3805. Fusion; 3806. Prediction; 3807. Decision-making; 3808. Planning; 3809. Control; 3810. Beyond-line-of-sight perception; 3811. Data fusion; 3812. Prediction; 3813. Decision-making; 3814. Planning; 3815. Control; 3816. Idle vehicle; 3817. Cloud / edge server; 3818. Cloud platform; 3901. Vehicle-mounted driving model; 3902. Perception; 3903. Recognition; 3904. Localization; 3905. Fusion; 3906. Prediction; 3907. Decision-making; 3908. Planning; 3909. Control; 3910. Beyond-line-of-sight perception; 3911. Data fusion; 3912. Prediction; 3913. Decision-making; 3914. Planning; 3915. Control; 3916. Raw data; 3917. Fuded data; 3918. Processed data; 3919. Idle vehicle; 3920. Cloud / edge server; 3921. Cloud platform; 4101. Vehicle-Edge-Cloud Collaborative Model Deployment and Allocation Submodule; 4102. Input Analysis Unit; 4103. Hardware Capability Verification Unit; 4104. Distillation Unit; 4105. Cloud-Edge Collaboration Unit; 4106. Allocation Execution Unit; 4107. Vehicle-Specific Model Allocation Unit; 4108. Route-Specific Model Allocation Unit; 4109. Vehicle Platform; 4110. Edge Platform; 4111. Cloud Platform; 4112. Allocation; 4113. Deployment; 4114. Vehicle; 4115. Edge; 4116. Cloud; 4201. Vehicle-Edge-Cloud Collaborative Model Training and Allocation Submodule; 4202. Input Analysis Unit; 4203. Hardware Capability Verification Unit; 4204. Cloud-Edge Collaboration Unit; 4205. Resource Invocation Unit; 4206. Scene Awareness Model Allocation Unit; 4207. Allocation Execution Unit; 4208. Vehicle-Specific Model Allocation Unit; 4209. Route-Specific Model Allocation Unit; 4210. Distillation Unit; 4211. Vehicle Platform; 4212. Edge Platform; 4213. Cloud Platform; 4214. Beyond Line of Sight Input; 4215. Allocation; 4216. Deployment; 4217. Vehicle-Based Training; 4218. Edge-Based Training; 4219. Cloud-Based Training; 4301. Parameter Allocation Module; 4302. Parameter Acquisition Submodule; 4303. Parameter Allocation Submodule; 4304. Parameter Update Submodule; 4305. Parameter Acquisition Unit; 4306. Parameter Statistics Unit; 4307. Computing Resource Unit; 4308. Parameter Allocation Unit; 4309. Vehicle Parameter Update Unit; 4310. ODD Parameter Update Unit; 4311. AI Model Parameter Update Unit; 4312. Parameter Update Progress Unit; 5201. Collaboration and Optimization Module; 5202. Collaboration Submodule; 5203. Optimization Submodule; 5301. Collaboration Submodule; 5302. Task Management Unit; 5303. Resource Matching Unit; 5304. Priority Processing Unit; 5305. Load Balancing Unit; 5306. Cluster Management Unit; 5401. Optimization Submodule; 5402. Resource Expansion Unit; 5403. Energy Consumption Optimization Unit; 5404. Cost Optimization Unit; 5405. Performance Tuning Unit.

[0049] Example: Figure 1 This paper demonstrates key components of a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving. The AI ​​system 101 for autonomous driving includes one or more of a computing power component 102, an AI model component 103, and / or a data component 104. The system is configured to be deployed on one or more platforms, including a vehicle system, an RSU system, an edge system, a TCC / TCU system, and / or a cloud platform. Specifically, the computing power component 102 is configured to distribute computing power among the vehicle system, RSU system, edge system, TCC / TCU system, and cloud platform; the AI ​​model component 103 includes micro-models, small models, medium models, and / or large models; and the data component 104 is configured to perform distributed data exchange among the vehicle system, RSU system, edge system, TCC / TCU system, and cloud platform.

[0050] Figure 2 An exemplary architecture for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving is demonstrated. A distributed artificial intelligence system 201 for vehicle-road-cloud cooperative autonomous driving includes an AI allocation subsystem 202 and an operation and maintenance management subsystem 203. The operation and maintenance management subsystem 203 supports the determination and implementation of deployment strategies for one or more modules in the AI ​​allocation subsystem 202, deploying them on one or more platforms including vehicle systems, RSU systems, edge systems, TCC / TCUs, and / or cloud platforms.

[0051] Figure 3 An exemplary architecture for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving is demonstrated. The AI ​​system 301 for autonomous driving includes an AI allocation subsystem 302 and an operation and maintenance management subsystem 303. The AI ​​allocation subsystem 302 includes one or more of a computing power allocation module 304, a data allocation module 305, a model allocation module 306, and / or a parameter allocation module 307. The operation and maintenance management subsystem 303 includes one or more of a coordination and optimization module 308 and an execution module 309.

[0052] Figure 4A system architecture for a distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving is demonstrated. The AI ​​allocation system 401 for autonomous driving includes one or more of the following: a computing power allocation module 402, a data allocation module 403, a model allocation module 404, and / or a parameter allocation module 405. The AI ​​allocation system 401 for autonomous driving is hosted 406 on one or more of the following: a vehicle system 407, an RSU system 408, an edge system 409, a TCC / TCU 410, and / or a cloud platform 411.

[0053] Figure 5 This is a schematic diagram of the model allocation module. The model allocation module 501 includes one or more of the following: large model 502, medium model 503, small model 504, and / or micro model 505. Micro model 505 and small model 504 are deployed 506 in vehicle system 507; micro model 505, small model 504, and / or medium model 503 are deployed 506 in RSU system and / or edge system 508; micro model 505, small model 504, medium model 503, and large model 502 are deployed 506 in TCC / TCU and / or cloud platform 509.

[0054] Figure 6 This is a schematic diagram illustrating the limitations of a vehicle's computing power. The limitations of a vehicle's computing power 601 include one or more of the following: power supply 602, cooling system 603, cost 604, and / or computing processing capacity 605.

[0055] Figure 7 An exemplary structure of the communication components of an AI system is shown. When an advanced cellular communication network is available, the communication component 701 of the AI ​​system includes one or more of 3G 703, 4G 704, 5G 705, 6G 706, digital broadcasting 707, Wi-Fi 708, Bluetooth 709, and satellite 710. When an advanced cellular communication network is unavailable, the communication component 702 of the AI ​​system includes one or more of 3G 703, 4G 704, and digital broadcasting 707.

[0056] Figure 8The architecture of a static computing power allocation submodule 801 in a vehicle-to-road / edge-cloud system is illustrated. This module includes one or more of the following: a static computing power identification unit 802, a computing power demand identification unit 803, a gap computing unit 804, and a computing power allocation unit 805. This submodule 801 can be deployed 818 on any of the following platforms: a vehicle computing platform 806, an RSU computing platform 807, an edge computing platform 808, a cloud computing platform 809, and a TCC computing platform 810, and configured to allocate the required vehicle computing resources 811, RSU computing resources 812, edge computing resources 813, cloud computing resources 814, and / or TCC computing resources 815. This submodule 801 can also utilize vehicle computing resources 816 and / or idle vehicle computing resources 817 to compensate for identified deficiencies, thereby achieving efficient static computing power allocation in the vehicle-to-road / edge-cloud system.

[0057] Figure 9 The architecture of the vehicle-road / edge-cloud dynamic computing power allocation submodule 901 is demonstrated. This module includes: an internal change identification unit 903 and / or an available computing power identification unit 904 from the supply side 902, an external change identification unit 906 and / or a computing power demand identification unit 907 from the demand side 905, a gap calculation unit 908, a distillation unit 909, a vehicle computing power allocation unit 910, a route computing power allocation unit 911, and / or a road computing power allocation unit 912. This submodule can be deployed 925 on any of the following platforms: on-board computing platform 913, RSU computing platform 914, edge computing platform 915, cloud computing platform 916, and TCC computing platform 917, and can allocate 926 the required computing resources, including vehicle computing resources 918, RSU computing resources 919, edge computing resources 920, cloud computing resources 921, and / or TCC computing resources 922. In some embodiments, submodule 901 may also utilize vehicle computing resources 923 and idle vehicle computing resources 924 to cope with dynamic fluctuations, ensuring that the system can meet real-time requirements and adapt to continuously changing supply conditions.

[0058] Figure 10This document demonstrates a static computing resource allocation workflow for planned applications, illustrating the allocation schemes before real-time operation or under external conditions through a typical sequence of steps. The system first collects information on all available computing resources, including onboard processing hardware (such as CPUs, GPUs, and memory) and potential external resources (such as RSU edge servers or cloud clusters). This resource list is collected at a set time or before the trip begins, ensuring that known capacity is fully recorded. Subsequently, the system calculates the expected computing requirements for vehicle operation based on factors such as route complexity, anticipated traffic density, weather conditions, and the AI ​​model being operated onboard. During this planning phase, the system analyzes whether standard sensor data streams (such as LiDAR, radar, and cameras) need to be offloaded, or whether existing onboard hardware can independently complete the task. The workflow checks the difference between resource availability and predicted demand. If there is no shortage, the system executes the predetermined allocation scheme; if a resource shortage is detected (e.g., the onboard CPU / GPU or memory cannot meet the expected workload), the calculation step precisely determines the required additional capacity. The system then selects one or more external resources (such as idle vehicle resources, RSU computing units, and / or remote cloud platforms) to compensate for the deficiency. During this phase, the static computing capacity allocation method employs a preset or rule-based allocation strategy, comprehensively considering cost, bandwidth, or latency constraints set during the planning phase to determine the task allocation scheme. Finally, after resource reallocation or acquisition is complete, the system will conclude the workflow by confirming whether the revised static plan meets the expected needs of the upcoming trip or operating period. If the trip or operating plan changes (e.g., due to changes in road conditions or updated user needs), the system can rerun the process to re-verify whether all identified needs can still be met.

[0059] Figure 11This demonstration showcases the workflow of dynamic computing power allocation for real-time vehicle operation, illustrating how a vehicle can cope with insufficient computing power under constantly changing conditions. In this process, the system first continuously monitors onboard parameters, including CPU / GPU utilization, battery status, and sensor health (such as LiDAR, radar, and camera streams), while simultaneously observing changes in the external environment, such as rapidly changing weather patterns, sudden traffic congestion, or unforeseen road hazards. When internal modules detect a potential increase in load (e.g., the addition of new sensors or advanced perception tasks triggered by adverse environmental factors), they automatically compare the total computing requirements with current onboard resources. If the onboard CPU / GPU capacity is sufficient, the vehicle will continue operating without adjustment. However, when the system identifies insufficient computing power (e.g., real-time sensor fusion requires more GPUs and memory than currently available), it immediately assesses available external resources to fill the gap. These resources may include dedicated RSU servers near the road, idle GPUs from surrounding vehicles, edge data centers with additional CPU / GPU capacity, or cloud servers capable of near real-time processing of large-scale inference. When the vehicle determines that accessing customized RSU computing resources (or similar external resources) is the optimal solution, the system will offload specific tasks such as intensive perception, map updates, or advanced path planning via wireless communication. This resource allocation is dynamic and has minimal impact on the overall autonomous driving function. (The final step is shown in the diagram.) Figure 11 This demonstrates how the system allocates customized RSU computing resources to compensate for insufficient onboard computing power, ensuring that the vehicle maintains safe and responsive handling performance under constantly changing operating or environmental scenarios. After resource allocation, the system re-enters a continuous monitoring loop, ready to detect new deficiencies or release redundant external resources. By coordinating real-time collaboration between the vehicle's own hardware, road infrastructure, and even other vehicles, this dynamic workflow provides robust autonomous driving performance even when computing demands fluctuate.

[0060] Figure 12 The architecture of the vehicle-external data distribution submodule 1201 is illustrated. This submodule includes one or more of the following: a data acquisition unit 1202, a data measurement unit 1203, a big data preprocessing unit 1204, a data distribution unit 1205, and a distillation preparation unit 1206. The distribution submodule 1201 can be deployed on vehicles 1207, RSUs 1208, edge devices 1209, cloud devices 1210, and / or TCCs 1211. This submodule can distribute and integrate data from any of the following devices: vehicles 1207, RSUs 1208, edge devices 1209, cloud devices 1210, and TCCs 1211, to one or more of the following devices: vehicles 1212, RSUs 1213, edge devices 1214, cloud devices 1215, and TCCs 1216.

[0061] Figure 13The architecture of the vehicle-mounted data distribution submodule 1301 is shown, comprising one or more of the following: a data acquisition unit 1302, a data measurement unit 1303, a big data preprocessing unit 1304, a data distribution unit 1305, and a distillation preparation unit 1306. This vehicle-mounted distribution submodule 1301 is configured to be deployed on vehicle 1307 and is responsible for distributing data from one or more vehicles 1307 and integrating it into one or more of vehicles 1308, RSU 1309, edge computing 1310, cloud computing 1311, and TCC 1312.

[0062] Figure 14 The architecture of the data acquisition unit 1401 is shown, which includes one or more external data acquisition subunits 1402 and on-board data acquisition subunits 1403. The external data acquisition subunit 1402 is responsible for acquiring on-board sensor data and vehicle status data, while the on-board data acquisition subunit 1403 is responsible for acquiring data from other vehicles, RSU, TCC, TCU, and the cloud.

[0063] Figure 15 The architecture of the data measurement unit 1501 is illustrated. This unit includes one or more of the following components: a detectable data identification component 1502, a data size determination component 1503, a model size determination component 1504, and an update frequency determination component 1505. Specifically, the detectable data identification component 1502 determines whether the data is detectable; the data size determination component 1503 determines the size of the data; the model size determination component 1504 determines the model size corresponding to the data during training or inference; and the update frequency determination component 1505 is specifically responsible for determining the data update frequency.

[0064] Figure 16 The architecture of a big data preprocessing unit 1601 is illustrated, which includes one or more of a data quality measurement component 1602, a data preprocessing component 1603, and a data fusion component 1604. The data quality measurement component 1602 is responsible for identifying data quality, including missing values, duplicate values, and validity checks. The data preprocessing component 1603 receives messages or notifications from the data quality measurement component 1602, performs missing value imputation and duplicate value removal, and reformats the dataset to a valid format. The data fusion component 1604, after completing data quality measurement and preprocessing, is responsible for fusing the data.

[0065] Figure 17The architecture of the data allocation unit 1701 is illustrated, which includes one or more of a data management component 1702 and a data allocation strategy component 1703. The data management component 1702 is responsible for queuing all data to be transmitted, ensuring proper organization and prioritization of the data. The data allocation strategy component 1703 dynamically adjusts the allocation strategy based on the measurement data from the data measurement unit 1701 and the priority of the data sent by the data management component 1702, thereby ensuring the granularity of data allocation.

[0066] Figure 18 The structure of the distillation preparation unit 1801 is shown, which includes one or more of the following components: a vehicle-specific data preparation component 1802, a route-specific data preparation component 1803, a road-specific data preparation component 1804, an ODD-specific data preparation component 1805, and a trip-specific data preparation component 1806. Specifically, the vehicle-specific data preparation component 1802 is configured to prepare training and validation data for a specific vehicle used in large-scale model distillation. The route-specific data preparation component 1803 is configured to prepare training and validation data for a specific route used in large-scale model distillation. The road-specific data preparation component 1804 is configured to prepare training and validation data for a specific road used in large-scale model distillation. The ODD-specific data preparation component 1805 is configured to prepare training and validation data for a specific operating design domain (ODD) used in large-scale model distillation. The trip-specific data preparation component 1806 is configured to prepare training and validation data for a specific trip used in large-scale model distillation.

[0067] Figure 19 The structure of the data allocation module hardware 1901 is shown, which includes one or more of the following components: data storage 1902, random access memory (RAM) 1903, central processing unit (CPU) 1904, and graphics processing unit (GPU) 1905. For example, in some embodiments, the data storage 1902 may be a Samsung 990 Pro solid-state drive (SSD); the RAM 1903 may be a Corsair Vengeance RGB DDR5 (6000MHz–8000MHz); the CPU 1904 may be an Intel Core i9-14900K; and the GPU 1905 may be an NVIDIA GeForce RTX 4090.

[0068] Figure 20This demonstrates the workflow of a big data reprocessing unit used for data cleaning and fusion. The process first determines whether the data is clean or unclean. If the data is unclean, it is preprocessed and then returned for reassessment. If the data is clean, its quality is further evaluated. If the data quality is good, it proceeds directly to the fusion stage; if the data quality is poor, it is preprocessed to improve its integrity. After preprocessing, the data is again assessed for cleanliness and good quality. If it still does not meet the requirements, it returns to continue processing; if it meets the criteria, it enters the fusion stage, where different datasets are merged. Finally, the process ends, ensuring that only high-quality, fully processed data is used for subsequent analysis.

[0069] Figure 21 This document demonstrates the workflow of a data allocation unit for training data distribution in a vehicle-road-cloud system. The data allocation unit receives notifications from a data measurement unit configured to determine data detectability, measure data size, evaluate the training model size, and assess update frequency. Based on these measurements, the data allocation unit distributes data to the vehicle, roadside, or cloud for training, achieving efficient processing between computing power and real-time requirements. The process first determines data detectability. If the data is not detectable, the update frequency is evaluated: if the update frequency is less than 10 ms, the data is allocated to the vehicle; if the update frequency is between 10 ms and 100 ms, the data is allocated to the roadside; otherwise, the data is allocated to the cloud. If the data is detectable, the data size is evaluated. If the data size is less than 10 GB, it is determined whether the training model size is less than 0.1B parameters; if it is less than this threshold, the update frequency is further evaluated, and the data is allocated to the vehicle, roadside, or cloud according to the update frequency. If the training model size is greater than 0.1B parameters but less than 1B parameters, it is allocated to either the roadside or the cloud based on the update frequency; if the training model size exceeds 1B parameters, it is directly allocated to the cloud. If the data size is between 10 GB and 100 GB, it is determined whether the training model size is between 0.1B and 1B parameters; if this condition is met, it is allocated to either the roadside or the cloud based on the update frequency; otherwise, it is directly allocated to the cloud. If the data size exceeds 100 GB, it is directly allocated to the cloud. Through this structured process, optimized data allocation based on real-time requirements, computational load, and data constraints is achieved between the vehicle, roadside, and cloud.

[0070] Figure 22This document demonstrates the workflow of a data allocation unit for inference data distribution in a vehicle-road-cloud system. The data allocation unit receives notifications from a data measurement unit configured to determine data detectability, measure data size, evaluate model size, and assess update frequency. Based on these measurements, the data allocation unit distributes data to the vehicle, roadside, or cloud for inference processing, achieving efficient processing according to computing power and real-time requirements. The process first determines data detectability. If the data is not detectable, the update frequency is evaluated: if the update frequency is less than 10 ms, the data is allocated to the vehicle; if the update frequency is between 10 ms and 100 ms, the data is allocated to the roadside; otherwise, the data is allocated to the cloud. If the data is detectable, the data size is further evaluated. If the data size is less than 10 GB, it is determined whether the model size is less than 0.1 B parameters. If the model size is below the threshold, the update frequency is further evaluated: if the update frequency is less than 10 ms, data is allocated to the vehicle; if the update frequency is between 10 ms and 100 ms, data is allocated to the roadside; otherwise, data is allocated to the cloud. If the model size is greater than 0.1B parameters and less than 1B parameters, the update frequency is evaluated: if the update frequency is between 10 ms and 100 ms, data is allocated to the roadside; otherwise, data is allocated to the cloud. If the model size exceeds 1B parameters, data is directly allocated to the cloud. If the data size is between 10 GB and 100 GB, it is determined whether the model size is between 0.1B and 1B parameters. If this condition is met, the update frequency is used to determine: if the update frequency is between 10 ms and 100 ms, data is allocated to the roadside; otherwise, data is allocated to the cloud. If the model size is not within the range of 0.1B to 1B parameters, data is allocated to the cloud. If the data size exceeds 100 GB, data is directly allocated to the cloud. Through the above structured process, optimized inference data allocation is achieved between the vehicle, roadside, and cloud based on real-time processing requirements, computational load, and data constraints.

[0071] Figure 23 The structure of a large-scale AI model 2301 is shown, which includes an application layer comprising recognition 2302, perception 2303, localization 2304, fusion 2305, prediction 2306, decision-making 2307, planning 2308, and control 2309. The large-scale AI model 2301 also includes a vehicle-specific model allocation unit 2310 for customizing the model according to the hardware conditions or operational constraints of a specific vehicle; a route-specific model allocation unit 2311 for adapting AI functions based on route conditions; and a distillation unit 2312 for refining or compressing the model as needed to generate smaller-scale model variants.

[0072] Figure 24 The structure of a medium-scale AI model 2401 is shown, which also includes an application layer comprising recognition 2402, perception 2403, localization 2404, fusion 2405, prediction 2406, decision-making 2407, planning 2408, and control 2409. Furthermore, the medium-scale AI model 2401 also includes a vehicle-specific model allocation unit 2410, a route-specific model allocation unit 2411, and a distillation unit 2412. Compared to large-scale AI models, this medium-scale AI model can employ a smaller network size or fewer computational resources to achieve a balance between performance and efficiency.

[0073] Figure 25 The structure of a small-scale AI model 2501 is shown, which includes an application layer comprising recognition 2502, perception 2503, localization 2504, fusion 2505, prediction 2506, decision-making 2507, planning 2508, and control 2509. The small-scale AI model 2501 also includes a vehicle-specific model allocation unit 2510, a route-specific model allocation unit 2511, and a distillation unit 2512. Compared to medium-scale AI models, this model typically has stricter resource constraints or a more streamlined structure, making it suitable for vehicles with limited onboard computing power.

[0074] Figure 26 The structure of a miniature AI model 2601 is shown, which includes an application layer comprising recognition 2602, perception 2603, localization 2604, fusion 2605, prediction 2606, decision-making 2607, planning 2608, and control 2609. The miniature AI model 2601 also includes a vehicle-specific model allocation unit 2610, a route-specific model allocation unit 2611, and a distillation unit 2612. This minimal-scale model prioritizes minimizing the use of onboard resources while still providing basic autonomous driving functions, and can support advanced functions through model distillation or external computation when needed.

[0075] Figure 27A The structure of an AI model repository 2703 for AI model components is demonstrated. This repository contains various types of AI technologies and is organized into six categories: deep learning models 2704, AI agent models 2705, reinforcement learning models 2706, generative models 2707, basic models 2708, and swarm intelligence models 2709. This model repository structure, based on technology implementation paths, systematically organizes, classifies, and retrieves AI models to support the implementation of autonomous driving functions.

[0076] In one implementation, the AI ​​model component is configured to select combinations and / or integrate AI models from an AI model library. For example, the system may employ a unified AI driving model that integrates perception, prediction, planning, decision-making, and control functions into a single model, achieving direct mapping from sensor inputs to control commands through end-to-end training. In another implementation, the system may employ a world model capable of learning dynamic changes in the environment and simulating future states, thereby supporting better decision-making and planning. Through these model combination methods, the system can meet the computational resource constraints of different scenarios while ensuring safety.

[0077] Figure 27B The mathematical framework for the model allocation module in an autonomous driving system is illustrated. As shown in the figure, the framework presents the core equation controlling AI model allocation: M = F(V, OV), where M represents the combined output of AI models selected from the model repository; V 2708 represents the set of AI models based on the vehicle; and OV 2709 represents the set of AI models based on external components. The function F determines the coordination between V and OV to generate a model combination M that meets a predetermined level of automation while simultaneously satisfying safety and computational resource constraints. Furthermore, the figure also shows the dependent equation: V = f1(x1, x2, ..., x...). ) and OV = f2(x1, x2, …, x ), where f1 and f2 represent functions used to determine the vehicle-side AI model and the vehicle-side AI model, respectively, and the variables x1 to x It represents a single model from different functional domains, including functional modules such as perception, recognition, localization, fusion, prediction, planning, decision-making, and control.

[0078] Figure 28 The deployment process of the AI ​​model allocation module is demonstrated. The process first determines a function F as the criterion for model allocation. Based on function F, a set of AI models M is determined. After the set of AI models M is determined, the process ends, ensuring that the AI ​​models are allocated for subsequent use.

[0079] Figure 29 The configuration process for the AI ​​model allocation module is demonstrated. This process begins by defining the model selection target, followed by input data type validation, driving scenario analysis, and hardware capability assessment. Next, based on the assessment results, the model size is determined, and finally, the optimized model is deployed for use.

[0080] Figure 30This demonstrates an example of the deployment and configuration of the AI ​​model allocation module. This module selects vehicle system 3001 as the deployment platform from among vehicle systems, traffic control center (TCC) 3002, edge nodes 3003, roadside unit systems 3004, and cloud platforms 3005. The external vehicle (OV) submodule provides additional data support for this process. Within vehicle system 3001, eight driving models are activated to complete the deployment: sensing 3006, perception 3007, localization 3008, fusion 3009, prediction 3010, planning 3011, decision-making 3012, and control 3013 models.

[0081] Figure 31 This demonstrates an example of AI model allocation module deployment and configuration at Level S1 (Level 1 according to the SAE definition of autonomous driving). For Level S1 autonomous driving, this module primarily allocates and coordinates vehicle-based sub-modules (Vehicle System 3101) to perform eight driving functions: sensing 3102, perception 3103, localization 3104, fusion 3105, prediction 3106, planning 3107, decision-making 3108, and control 3109. At this stage, external sub-modules such as roadside units 3110 and cloud platforms 3111 are not yet enabled.

[0082] Figure 32 A small AI model 3201 is demonstrated, designed for deployment in vehicle systems, roadside units, or cloud platforms. This model includes core modules at the application layer 3202: perception 3203, localization 3204, prediction 3205, decision-making 3206, control 3207, and fusion 3208. Its model parameters 3213 are less than 1 billion parameters 3214, relying on a standard onboard computing platform 3218 as its hardware requirement 3217, and focusing on performing safety-critical tasks 3215 in the priority module 3216. Model inputs include vehicle perception data 3220, external data 3221, state data 3222, and cloud data 3223, thereby achieving efficient autonomous driving functions with limited resources.

[0083] Figure 33A large-scale AI model 3301 is demonstrated, whose application layer 3302 includes a basic module and an enhancement module 3310. The basic module includes: perception 3304, localization 3305, prediction 3306, decision-making 3307, control 3308, and fusion 3309. The enhancement module includes: macro-prediction 3311, parallel decision-making 3312, multi-source data fusion 3313, big data analysis 3314, and traffic signal control 3315. This model is deployed on a cloud platform 3317 as part of the deployment location 3316. Under the model parameters 3318, the model has over 7 billion parameters 3319, and prioritizes the execution of agent-related tasks 3321 in the priority module 3320. The hardware requirements 3322 specify that it needs to be deployed in a cloud cluster 3323. The input module 3324 supports multi-source detection data 3325, traffic memory data 3326, human commands 3327, and cloud platform built-in data 3328.

[0084] Figure 34 A distilled, intermediate-scale AI model 3401 is demonstrated, whose application layer 3402 includes basic modules—perception 3404, localization 3405, prediction 3406, decision-making 3407, control 3408, and fusion 3409. Enhanced applications 3410 include meso-level prediction 3411, parallel decision-making 3412, cooperative control 3413, and cooperative perception 3414. Applications inherited from a large-scale AI model 3415 include parallel decision-making 3416, multi-source data fusion 3417, and traffic signal control 3418. This model is deployed as part of a deployment location 3419 on a cloud platform 3420 and a roadside unit system 3421. Model parameters 3422 are between 1 billion and 7 billion 3423. A priority module 3424 handles agent-related tasks 3425. Hardware requirements 3426 are an edge platform 3427. The input module 3428 includes multi-source detection data 3429, large model knowledge 3430, manual instructions 3431, and edge-embedded data 3432.

[0085] Figure 35 The interaction process between the small AI model 3510 on the vehicle side 3509, the medium AI model 3506 on the edge 3505, and the large AI model 3502 on the cloud 3501 is demonstrated. The vehicle sends perception and status data 3512, the edge collects data in real time 3507 and periodically synchronizes and updates content or model inference results 3508; the cloud 3501 receives the inference result feedback 3503 from the edge 3505 and provides inference correction and knowledge training to the edge 3505; when the medium AI model 3506 on the edge 3505 crashes, the vehicle-side model is backed up 3511 to the cloud 3501.

[0086] In one specific implementation, the real-time update method is applicable to data and / or model updates within a time range of 0.1 milliseconds to 1 second. For example, for safety-critical tasks such as collision warning and emergency braking, the system employs a real-time update method to ensure that model parameters are synchronized within milliseconds. The dynamic update method is applicable to data and / or model updates within a time range of 1 second to several days. For example, for non-urgent tasks such as driving behavior analysis, map updates, and model retraining, the system employs a dynamic update method to perform batch updates when vehicles are idle or network conditions permit. Through the coordinated operation of the above update mechanisms, the system can effectively manage computing resources and network bandwidth while ensuring real-time response.

[0087] Figure 36 The L1-level architecture was showcased: a vehicle-centric onboard driving model 3601. This architecture corresponds to the scenario where micro-models and small models are deployed in the vehicle system. It includes multiple driving functions, covering perception 3602, recognition 3603, localization 3604, fusion 3605, prediction 3606, decision-making 3607, planning 3608, and control 3609. In this vehicle-centric configuration, all eight driving functions are deployed on the onboard platform, relying on local sensor data and computing power to achieve autonomous driving operation without relying on external data or computing resources.

[0088] Figure 37 The L2-level architecture was showcased: an onboard model + an external model / data. This architecture corresponds to scenarios where micro-models and small models are deployed in the vehicle system, while medium-sized models are deployed in the RSU system or edge system. The onboard driving model 3701 is equipped with multiple onboard driving functions (perception 3702, recognition 3703, localization 3704, fusion 3705, prediction 3706, decision-making 3707, planning 3708, and control 3709), while simultaneously receiving and fusing external driving model capabilities (including beyond-line-of-sight perception 3710, data fusion 3711, prediction 3712, decision-making 3713, planning 3714, and control 3715). This layer also includes external data sources, providing raw data 3716, fused data 3717, and processed data 3718. Thus, the vehicle can leverage external data to enhance its local perception and decision-making processes, achieving more comprehensive environmental perception and more refined autonomous driving functions.

[0089] Figure 38The L3-level architecture was showcased: an in-vehicle model + an external model / computing system. This architecture further utilizes the computing resources of edge computing platforms and cloud platforms. The in-vehicle driving model 3801 is equipped with multiple in-vehicle driving functions (perception 3802, recognition 3803, localization 3804, fusion 3805, prediction 3806, decision-making 3807, planning 3808, and control 3809). Additional or high-computing-power-requirement tasks are performed by the external driving model (beyond line-of-sight perception 3810, data fusion 3811, prediction 3812, decision-making 3813, planning 3814, and control 3815). This configuration relies on external computing resources such as idle vehicles 3816, cloud / edge servers 3817, and cloud platforms 3818 to handle complex or high-load computing tasks, thereby reducing the processing pressure on the in-vehicle platform and improving real-time performance under different operating conditions.

[0090] Figure 39 The L4-level architecture was showcased: an in-vehicle model + an external model / data / computing. This architecture fully utilizes all resources at the vehicle, road, edge, and cloud levels to achieve comprehensive vehicle-road-cloud collaboration. The in-vehicle driving model 3901 is equipped with multiple in-vehicle driving functions (perception 3902, recognition 3903, localization 3904, fusion 3905, prediction 3906, decision-making 3907, planning 3908, and / or control 3909), while additional driving functions are offloaded to the external driving model (beyond line-of-sight perception 3910, data fusion 3911, prediction 3912, decision-making 3913, planning 3914, and control 3915) or shared with external resources. At this level, external data (raw data 3916, fused data 3917, processed data 3918) and external computing (idle vehicles 3919, cloud / edge servers 3920, cloud platform 3921) together provide strong support for the vehicle's autonomous driving tasks, ensuring that the system can dynamically expand its perception, computing, and model update capabilities according to network availability, safety requirements, and route complexity.

[0091] Figure 40 This demonstrates the workflow of a dynamic vehicle control system that utilizes onboard, roadside, and cloud computing in a collaborative manner. The process begins with onboard computation and data input analysis using external data. The workflow then branches into three paths: applying vehicle system computation, applying roadside system computation, and applying cloud platform computation. The vehicle system computation path proceeds by deploying sensing and localization models, performing sensor data fusion, predicting traffic flow and trajectories for decision-making, generating and executing control commands, and performing vehicle control tasks. The roadside system computation path performs sensor data fusion and predicts traffic flow and trajectories for decision-making, then inputs the results into the stage of generating and executing control commands. The cloud platform computation path retrains and fine-tunes the model based on new extreme scenarios, then deploys the updated model to the vehicle, and then continues sensor data fusion. The workflow terminates after the vehicle control tasks are completed.

[0092] Figure 41 A model deployment and allocation submodule for vehicle-edge-cloud collaboration is demonstrated. This submodule 4101 includes an input analysis unit 4102, a hardware capability verification unit 4103, a distillation unit 4104, a cloud-edge collaboration unit 4105, an allocation execution unit 4106, a vehicle-specific model allocation unit 4107, and a route-specific model allocation unit 4108. The submodule 4101 receives allocation instructions 4112 from the vehicle platform 4109, the edge platform 4110, and / or the cloud platform 4111, and performs deployment 4113 across the vehicle 4114, edge 4115, and cloud 4116. By distributing AI model components and tasks across the vehicle, edge, and cloud layers, the submodule 4101 can dynamically adjust deployment based on network availability, processing load, and route complexity.

[0093] Figure 42 A model training allocation submodule for vehicle-edge-cloud collaboration is demonstrated. The vehicle-edge-cloud collaborative model training allocation submodule 4201 includes an input analysis unit 4202, a hardware capability verification unit 4203, a cloud-edge collaboration unit 4204, a resource allocation unit 4205, a scene-aware model allocation unit 4206, an allocation execution unit 4207, a vehicle-specific model allocation unit 4208, a route-specific model allocation unit 4209, and a distillation unit 4210. The vehicle-edge-cloud collaborative model training allocation submodule 4201 receives allocation instructions 4215 from the vehicle platform 4211, the edge platform 4212, the cloud platform 4213, or beyond-line-of-sight input 4214, and deploys 4216 to support vehicle-based training 4217, edge-based training 4218, and / or cloud-based training 4219. By dynamically distributing training tasks across vehicles, edge nodes, and the cloud, the vehicle-edge-cloud collaborative model training allocation submodule 4201 ensures efficient and scalable model updates based on real-time traffic conditions, bandwidth limitations, and evolving security requirements.

[0094] Figure 43 The component structure of the parameter allocation module is shown. The parameter allocation module 4301 consists of three sub-modules: a parameter acquisition sub-module 4302, a parameter allocation sub-module 4303, and a parameter update sub-module 4304. Specifically, the parameter acquisition sub-module 4302 includes a parameter acquisition unit 4305 and a parameter statistics unit 4306; the parameter allocation sub-module 4303 includes a computing resource unit 4307 and a parameter allocation unit 4308; and the parameter update sub-module 4304 consists of a vehicle parameter update unit 4309, an ODD parameter update unit 4310, an AI model parameter update unit 4311, and a parameter update progress unit 4312.

[0095] Figure 44 This demonstrates the process of parameter allocation and updating using a parameter allocation module during vehicle operation. The parameter acquisition unit collects all parameters requiring updates during vehicle operation, while the parameter statistics unit identifies the parameter types and calculates their scale. The computing power resource unit calculates the available computing power for each host system, and the parameter allocation unit distributes parameters to different host systems based on their type, scale, and computing power availability. Finally, the parameter update submodule employs a sequential update strategy to update all types of parameters.

[0096] Figure 45 This demonstrates the process of collecting and / or statistically evaluating parameters using the parameter acquisition submodule. The parameter acquisition unit collects all parameters that need to be updated during vehicle operation via a communication component. The parameter statistics unit identifies the parameter types, including vehicle parameters, ODD parameters, and AI model parameters, and calculates the scale of each type of parameter.

[0097] Figure 46 An exemplary parameter allocation process is illustrated. First, the computing power resource unit calculates the available computing power of each host system during the parameter update process. Then, the parameter allocation unit obtains the type and scale of the parameters from the parameter statistics unit of the parameter acquisition submodule, and allocates the parameters to different host system groups based on the availability of computing power resources. When the parameter scale is greater than 1 billion, it is allocated to the cloud system and / or TCC / TCU; when the scale is between 100 million and 1 billion, it is allocated to the cloud system, TCC / TCU, and / or edge system; when the scale is between 5 million and 100 million, it is allocated to the cloud system, TCC / TCU, edge system, and / or road system; if the parameter scale is below the above range, it is allocated to the cloud system, TCC / TCU, edge system, road system, and / or vehicle system.

[0098] Figure 47 An exemplary parameter allocation process is demonstrated for a parameter scale exceeding 1 billion. When the computing power of the TCC / TCU is sufficient to support the update of parameters of this scale, the parameter allocation submodule allocates the parameters to the TCC / TCU; if the computing power of the TCC / TCU is insufficient to complete the update, the parameter allocation submodule will allocate the parameters to the cloud system for processing.

[0099] Figure 48 An exemplary parameter allocation process is demonstrated for parameter sizes between 100 million and 1 billion. If the computing power of the edge system can support the update of parameters at this scale, the parameter allocation submodule will allocate the parameters to the edge system; subsequently, if the computing power of the TCC / TCU can support the parameter update, the parameter allocation submodule will continue to allocate the parameters to the TCC / TCU; if the computing power of the TCC / TCU is insufficient, the parameter allocation submodule will allocate the parameters to the cloud system.

[0100] Figure 49 An exemplary parameter allocation process is demonstrated for parameter scales between 5 million and 100 million. If the computing power of the road system can support the update of parameters at this scale, the parameter allocation submodule will prioritize allocating parameters to the road system; subsequently, if the edge system has sufficient computing power, the parameters will be allocated to the edge system; then, if the computing power of the TCC / TCU can meet the update requirements, the parameters will continue to be allocated to the TCC / TCU; if the computing power of the TCC / TCU is still insufficient to support the update, the parameters will finally be allocated to the cloud system for processing.

[0101] Figure 50 An exemplary parameter allocation process for parameters smaller than 5 million is demonstrated. If the computing power of the onboard system can support the update of parameters of this scale, the parameter allocation submodule will prioritize allocating parameters to the onboard system; subsequently, if the road system has sufficient computing power, the parameters will be allocated to the road system; next, if the edge system can support the parameter update, the parameters will continue to be allocated to the edge system; then, if the computing power of the TCC / TCU meets the update requirements, the parameters will be allocated to the TCC / TCU; if none of the above systems can meet the computing power requirements, the parameters will finally be allocated to the cloud system for processing.

[0102] Figure 51 An exemplary parameter update flowchart is shown. The vehicle parameter update unit, ODD parameter update unit, and / or AI model parameter update unit first obtain parameter update requests from the parameter allocation submodule and determine whether an update request exists within the current time interval. If no update request exists, the parameter update process ends; if an update request exists but the vehicle is in an unsafe operating state, the parameter update progress unit records the vehicle parameter update status and executes the update after the vehicle stops. If an update request exists and the vehicle is in a safe operating state, the parameter update begins. During the parameter update process, if the computing power of each host system can support the current parameter update, the vehicle parameter update unit, ODD parameter update unit, and / or AI model parameter update unit will execute a complete parameter update, and the parameter update progress unit will record the update status; if the computing power is insufficient, the vehicle parameter update unit, ODD parameter update unit, and / or AI model parameter update unit will execute a partial parameter update, and the parameter update progress unit will record the update status. In the next update cycle, these update units will repeat the same judgment and update process based on new update requests and computing power availability.

[0103] Figure 52 The component structure of the collaboration and optimization module is shown. The collaboration and optimization module 5201 consists of a collaboration submodule 5202 and an optimization submodule 5203, and is used to uniformly manage AI models, computing resources and data.

[0104] Figure 53 The component structure of the collaboration submodule is shown. The collaboration submodule 5301 includes a task management unit 5302, a resource matching unit 5303, a priority processing unit 5304, a load balancing unit 5305, and a cluster management unit 5306, which are used to organize, manage, and arrange the execution order of various tasks according to the allocated resources.

[0105] Figure 54 The component structure of the optimization submodule is shown. The optimization submodule 5401 consists of a resource expansion unit 5402, an energy consumption optimization unit 5403, a cost optimization unit 5404, and a performance tuning unit 5405, and aims to ensure that the allocated resources are used most efficiently and that all tasks are executed at optimal performance.

[0106] Figure 55 The flowchart illustrates the collaboration and optimization modules. The process begins with the collaboration submodule, which manages the order and timing of task execution. Next, the collaboration submodule matches tasks with allocated resources to ensure correct task-resource mapping. The optimization submodule then monitors resource usage and system performance in real time, dynamically adjusting task execution methods based on workload fluctuations to maintain overall operational efficiency within resource constraints. Finally, the optimization submodule generates and fine-tunes the optimization results and feeds them back to the collaboration submodule, initiating the next task execution cycle.

[0107] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

Claims

1. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving, characterized in that, The system includes an AI allocation subsystem and an operation and maintenance management subsystem, used to provide customized and / or personalized AI services to autonomous vehicles based on trip characteristics, route characteristics, road characteristics, operational design domain (ODD) characteristics, and / or vehicle characteristics. The system is configured to be deployed on one or more of the following platforms: (1) Vehicle system; (2) Roadside Unit (RSU) system; (3) Edge systems; (4) Traffic Control Center (TCC) / Traffic Control Unit (TCU); (5) Cloud platform; The system also includes one or more of the following components: (1) AI model components; (2) Computing power components; (3) Data components; (4) Communication components.

2. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 1, characterized in that: In the system: (1) The AI ​​allocation subsystem includes one or more of the following modules: a. Computing power allocation module; b. Data allocation module; c. Model allocation module; d. Parameter allocation module, (2) The operation and maintenance management subsystem includes one or more of the following modules: a. Collaboration and Optimization Module; b. Execution module.

3. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 1, characterized in that: The AI ​​model components include one or more of the following: micro-models, small models, medium models, and large models, wherein: (1) Micro-models and small models are deployed and allocated in the vehicle system; (2) Micro-models, small models and / or medium models are deployed and distributed in RSU systems and / or edge systems; (3) Micro-models, small models, medium models and / or large models are deployed and distributed in TCC / TCU and / or cloud platforms.

4. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 1, characterized in that: The AI ​​model component is configured to train and deploy the vehicle via an edge system and / or cloud platform using one or more of the following training and deployment methods: (1) Method 1: Both the trained model and the data are updated using a dynamic update method; (2) Method 2: The trained model is deployed in the vehicle and updated in real time, while the data is updated using a dynamic update method; (3) Method 3: The trained model is deployed in the vehicle and updated using a dynamic update method, while the data is updated using a real-time update method; (4) Method 4: The trained model and data are both updated in real time using a real-time update method; in: (1) Real-time method updates are configured to apply to data and / or training models with a time range of 0.1 milliseconds to 1 second; (2) The dynamic update method is configured to be applicable to data and / or training models with a time range of 1 second to several days.

5. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 1, characterized in that: The AI ​​model component is configured to select combinations and / or integrations of AI models from an AI model library to meet the safety and computing resource requirements of autonomous driving. The combinations and / or integrations of AI models include the following types: (1) A unified AI driving model, which is developed based on the AI ​​model library; (2) An end-to-end or point-to-point model for autonomous driving; (3) A world model for autonomous driving.

6. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 1, characterized in that: The data component includes one or more of the following data: (1) Vehicle data; (2) ODD data; (3) System data.

7. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The computing power allocation module is used to provide customized or personalized computing resource allocation, and includes one or more of the following sub-modules: (1) A static computing power allocation submodule, which includes one or more of the following units: a. Static computing power identification unit, used to identify available vehicle computing resources, RSU computing resources, edge computing resources, cloud computing resources and / or TCC computing resources; b. Computing power demand identification unit, which estimates, predicts and plans the computing power demand of the vehicle during the trip; c. Computing power gap calculation unit: calculates the difference between the identified available computing power on the vehicle and the determined computing power requirement, and uses it as the computing power gap; d. Computing power allocation unit, which allocates computing power of expected vehicle computing resources, expected RSU computing resources, expected edge computing resources, expected cloud computing resources and / or expected TCC computing resources to compensate for computing power gaps; (2) A dynamic computing power allocation submodule, comprising: a. One or more of the supply-side internal change identification unit and the available computing power identification unit; b. One or more of the external change identification unit and computing power demand identification unit on the demand side; c. Gap calculation unit and distillation unit; d. One or more of the following computing power allocation units for vehicle trips: vehicle computing power allocation unit, path computing power allocation unit, road computing power allocation unit, and ODD computing power allocation unit.

8. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The computing power allocation module is configured to allocate the required vehicle computing power, RSU computing power, edge computing power, TCC computing power, and / or cloud computing power to the vehicle-road-cloud system or vehicle-edge-cloud system, wherein: (1) The vehicle computing platform is configured to perform real-time perception, decision-making and control using onboard units; (2) The RSU computing platform is configured to provide vehicle-to-infrastructure (V2I) communication, sensor data fusion, and localized processing; (3) The edge computing platform is configured to provide low-latency data processing and collaborative sensing through mobile edge computing MEC servers; (4) The TCC computing platform is configured to coordinate traffic management and optimize the operation of the entire network; (5) The cloud computing platform is configured to handle large-scale data storage, model training and predictive analytics.

9. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The computing power allocation module optimizes and refines computing power allocation through a distillation method, providing customized or personalized computing power allocation schemes with vehicle-specific, route-specific, road-specific, and / or ODD-specific characteristics. The distillation method includes: (1) Knowledge extraction, used to identify and capture key computational patterns of specific vehicles, specific roads, specific routes and / or specific ODDs; (2) Compression steps are used to reduce redundant calculations and the allocation of computing resources.

10. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The data allocation module is used to allocate vehicle data and ODD data to one or more of the following scenarios: (1) Vehicle-specific scenarios; (2) Route-specific scenarios; (3) Road-specific scenarios; (4) ODD specific scenarios; (5) Specific scenarios of the trip; The data allocation module is configured for use in one or more of the following applications: (1) In-vehicle applications; (2) External application of vehicles.

11. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The data allocation module includes one or more of the following units: (1) Data acquisition unit, Includes the following sub-units: a. Vehicle-mounted data acquisition subunit; b. External data acquisition subunit; (2) A data measurement unit, comprising the following components: a. A detectable data recognition component used to identify whether data can be detected; b. Data size determination component, used to determine the data size required for training or inference; c. Model size determination component, used to determine the model size required for training or inference; d. Update frequency determination component, used to determine the update frequency of data; (3) Big data processing unit, including the following components: a. Data quality measurement components are used to identify data quality, including missing values, duplicates, and validity; b. Data preprocessing components are used to remove or fill missing values, remove duplicates, and reformat the dataset into a valid format; c. Data fusion component, used to fuse data after data quality measurement and data preprocessing; (4) A data allocation unit, comprising the following components: a. Data management component, used to queue all data to be transmitted, ensuring proper organization and prioritization of the data; b. A data allocation strategy component, used to adjust the appropriate allocation strategy based on the measurement results of the data measurement unit; (5) Distillation preparation unit, including the following components: a. Vehicle-specific data preparation component, used to prepare training and validation data for large-scale model distillation in vehicle-specific scenarios; b. Route-specific data preparation component, used to prepare training and validation data for large model distillation in route-specific scenarios; c. Road-specific data preparation component, used to prepare training and validation data for large model distillation in road-specific scenarios; d. ODD-specific data preparation components are used to prepare training and validation data for large model distillation in specific ODD scenarios; e. Trip-specific data preparation component, used to prepare training and validation data for large model distillation in trip-specific scenarios.

12. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 10, characterized in that: The data allocation unit is configured to dynamically adjust the data allocation strategy by comprehensively considering the urgency of the data, network conditions, and processing capabilities, in order to optimize transmission efficiency.

13. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The model allocation module is configured to allocate AI models from the AI ​​model repository according to the following formula, so as to enable the autonomous driving system to achieve a specific level of automation: in: (1) As the output of the model allocation module, it represents any combination and / or integration of AI models from the AI ​​model repository; (2) It is a set of in-vehicle AI models selected based on one or more of the safety requirements and computing power capabilities related to in-vehicle autonomous driving systems. It is certain How to collaborate to produce The function, This refers to one of the following models: recognition model, perception model, localization model, fusion model, prediction model, planning model, decision-making model, and / or control model; (3) It is a set of external AI models that execute under conditions that satisfy vehicle computing power constraints, data constraints, power constraints, and / or communication constraints. It is certain How to collaborate to produce The function; where This refers to one of the following models: recognition model, perception model, localization model, fusion model, prediction model, planning model, decision-making model, and / or control model; (4) It is a certainty and How to collaborate to produce The function.

14. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The model allocation module is configured to allocate one or more of the following models in a specific deployment scenario: (1) Large-scale cloud model, deployed in the cloud; (2) Edge distillation model, deployed at the edge; (3) Vehicle-mounted miniature model, deployed on the vehicle; (4) Vehicle-mounted micro-model, deployed on the vehicle; The model allocation module is used to realize distributed model and data flow interaction between the vehicle end, road end, edge end and cloud.

15. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The model allocation module includes one of the following sub-modules: (1) Vehicle-Edge-Cloud Collaborative Model Deployment and Allocation Submodule, which includes one or more of the following units: a. Input analysis unit, used to analyze the input data and scenario requirements of deployment tasks; b. Hardware capability verification unit, used to verify whether the computing hardware of the vehicle platform, edge platform or cloud platform meets the requirements for model operation; c. Cloud-edge collaboration unit, used to coordinate model deployment tasks between the cloud and the edge; d. Assignment execution unit, used to perform the final assignment of the model to vehicles, edge, or cloud platforms; e. Vehicle-specific model allocation unit; f. Route-specific model allocation unit; g. Distillation unit, used to distill and compress the model to adapt it for edge or vehicle-end deployment; (2) Vehicle-Edge-Cloud Collaborative Model Training Allocation Submodule, which includes one or more of the following units: a. Input analysis unit, used to analyze the input data and scenario requirements of the training task; b. Hardware capability verification unit, used to verify whether the computing hardware of the vehicle platform, edge platform or cloud platform meets the model training requirements; c. Cloud-edge collaboration unit, used to coordinate model training tasks between the cloud and the edge; d. Assign execution units to perform the assignment of training tasks to vehicles, edge, or cloud platforms; e. Resource allocation unit; f. Scene-aware model allocation unit; g. Vehicle-specific model allocation unit; h. Route-specific model allocation unit; i. Distillation unit.

16. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The model allocation module is configured to provide customized and / or personalized model allocation schemes for vehicle operation, wherein the customized and / or personalized models include one or more of the following models: (1) User-customized model; (2) Route-specific model; (3) Road-specific model; (4) Vehicle-specific models; (5) ODD-specific model; (6) Journey-specific model; The customized and / or personalized model allocation scheme is optimized for specific trips, routes, roads, ODDs, and / or vehicles using a distillation method. This distillation method is configured to support hierarchical knowledge transfer across cloud, edge, and vehicle endpoints in the following ways: (1) Knowledge extraction from the cloud; (2) Edge model distillation; (3) Adaptive distillation of vehicle model; (4) Regularly synchronize and back up the model.

17. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 4, characterized in that: The training and deployment method is configured to employ a real-time and / or dynamic update mechanism to synchronously or stepwise train and deploy the AI ​​driving model, achieved through one of the following real-time / dynamic update methods: (1) Method 1: Synchronous real-time training and deployment are carried out within an update interval of 1 millisecond to 1 second, so that the parameters and running version of the AI ​​driving model can be continuously optimized and applied in real time; (2) Method 2: Synchronous dynamic training and deployment are carried out within update intervals ranging from 1 second to several days, allowing AI driving models to be jointly updated and retrained according to a flexible schedule that adapts to larger datasets or resource availability; (3) Method 3: A separate training and deployment mechanism is adopted. When the interval between dynamic training ranges from 1 second to several days, the deployment process is executed in real time within the range of 1 millisecond to 1 second. In this way, the running AI driving model can be continuously updated and kept up-to-date. The principle is to use the incremental parameter updates generated by the low-frequency but wider-coverage training cycle. (4) Method 4: Real-time training and dynamic deployment are carried out simultaneously, in which the parameters of the artificial intelligence driving model are updated in real time at intervals of 1 millisecond to 1 second, and then deployed at more flexible cycles ranging from 1 second to several days.

18. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The parameter allocation module is configured to implement parameter allocation and updating on one or more of the following: vehicles, RSUs, edge devices, TCCs / TCUs, and / or cloud platforms, and includes one or more of the following sub-modules: (1) Parameter acquisition submodule, including one or more of the following units: a. Parameter acquisition unit, used to collect AI model parameters, vehicle parameters and ODD parameters that need to be updated in one or more systems, including vehicle system, road system, edge system, TCC / TCU and cloud system; b. Parameter statistics unit, used to process and back up the types and scales of parameters that need to be updated; (2) Parameter allocation submodule, including one or more of the following units: a. Computing resource unit, used to calculate the available computing power for parameter updates in one or more systems, including vehicle systems, road systems, edge systems, TCC / TCU and cloud systems; b. Parameter allocation unit, used to provide parameter updates to one or more of the following systems based on the availability of computing resources and the type and scale of the parameters to be updated: vehicle system, road system, edge system, TCC / TCU and cloud system; (3) Parameter update submodule, including one or more of the following units: a. Vehicle parameter update unit, used to update one or more execution parameters among vehicle status parameters and operating information parameters; b. ODD parameter update unit, used to update one or more execution parameters among road parameters, object parameters, and driving environment parameters; c. AI model parameter update unit, used to update one or more execution parameters among the basic model parameters, training and optimization parameters, and deployment and adaptation parameters; d. Parameter update progress unit, used to record the progress of parameter updates.

19. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The collaboration and optimization module is used to manage one or more of the AI ​​model components, computing power components, and data components, and includes the following sub-modules: (1) A collaborative submodule, which includes one or more of the following units: a. Task Management Unit, used to manage the order and timing of task execution on one or more of the following platforms: vehicle system, RSU system, edge system, TCC / TCU, and cloud platform; b. Resource matching unit, used to allocate tasks to computing resources according to task requirements; c. Priority processing unit, used to balance tasks according to priority level, to ensure that critical tasks receive immediate attention, and to optimize the overall resource utilization of the system, wherein the priority level is classified according to factors such as task urgency, time sensitivity, system impact and resource availability; d. Load balancing unit, used to distribute workloads across one or more of the following platforms: vehicle systems, RSU systems, edge systems, TCC / TCU and cloud platforms, to avoid bottlenecks and underutilization of resources; e. Cluster management unit, used to oversee coordination between distributed systems, ensuring effective communication between one or more of the following platforms: vehicle systems, RSU systems, edge systems, TCC / TCU, and cloud platforms; (2) An optimization submodule, which includes one or more of the following units: a. Resource expansion unit, used to adjust computing resources according to real-time demand; b. Energy consumption optimization unit, used to minimize energy consumption while meeting performance requirements; c. Cost optimization unit, used to allocate cloud and local resources in the most cost-effective manner; d. Performance tuning unit, used to fine-tune task execution parameters to achieve optimal resource utilization.

20. A distributed artificial intelligence system for vehicle-road-cloud cooperative autonomous driving according to claim 2, characterized in that: The collaboration and optimization module operates as follows: (1) The collaborative submodule dynamically adjusts the execution priority of the vehicle's AI driving model based on vehicle operating status data, including acceleration, braking and navigation requirements; (2) The collaborative submodule prioritizes artificial intelligence decisions in autonomous driving navigation by dynamically allocating computing resources among the perception, planning and control systems; (3) Optimize the sub-module to dynamically reallocate computing power between safety-critical AI driving tasks and non-essential tasks to maximize efficiency; (4) The optimization sub-module minimizes AI decision-making delay by intelligently allocating tasks between on-board computing resources and external computing processing.

Images