Multi-agent cooperative unmanned vehicle control system, method, and computer equipment

Abstract

This application relates to an unmanned vehicle control system, method, and computer device based on multi-agent cooperation. Through the design of the multi-agent cooperative driving system, the decision-making ability of the autonomous driving system in complex scenarios is significantly improved. The remote driver agent can effectively take over and handle special scenarios that the autonomous driving system struggles to cope with, such as construction areas and traffic accident scenes. Intelligent driving rights allocation based on scenario complexity and system state is achieved, improving the overall safety and reliability of the system. A real-time evaluation mechanism for the driver evaluation agent ensures the quality of driving operations and provides a basis for driver dispatch decisions.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to an unmanned vehicle control system, method and computer device based on multi-agent cooperation. Background Technology

[0002] With the rapid development of autonomous driving technology, the safety, reliability, and adaptability of unmanned vehicle control systems in complex traffic environments have become a focus of industry attention. Currently, autonomous driving systems mainly adopt a single agent architecture; however, the decision-making capabilities of a single agent architecture are limited, making it difficult to cope with all possible traffic situations. Summary of the Invention

[0003] Therefore, it is necessary to provide an unmanned vehicle control system, method, and computer equipment based on multi-agent collaboration to address the aforementioned technical problems.

[0004] An unmanned vehicle control system based on multi-agent cooperation, the system comprising: An autonomous driving agent is deployed locally in the vehicle and configured to perform autonomous driving operations in normal scenarios; the autonomous driving operations include path planning, speed control, and basic environmental perception. A remote driver agent, deployed at a remote control center, integrates a visual language model. The visual language model is configured to receive multimodal data collected by vehicle sensors and output scene semantic descriptions and operation instructions. Based on the output of the visual language model, the remote driver agent performs remote driving operations when the autonomous driver agent triggers a takeover request. The driver dispatch agent is configured to dynamically allocate driving rights between the autonomous driver agent and the remote driver agent based on task complexity, the load status of the remote driver agent, and the historical scenario adaptability. The driver evaluation agent is configured to evaluate the operational performance of the autonomous driver agent and the remote driver agent in real time based on preset scenario-based evaluation dimensions, and feed the evaluation results back to the driver dispatch agent.

[0005] In one embodiment, the visual language model includes: The multimodal feature fusion unit is configured to map the visual features of the image, the point cloud features of the radar, and the text commands to a unified semantic space to generate cross-modal feature vectors. The scenario reasoning unit is configured to analyze the interaction relationships and potential risk evolution trends of traffic participants based on the cross-modal feature vectors. The semantic conversion unit is configured to convert the reasoning results into structured text and natural language descriptions, which are then output to the driving control module and the remote operator interface, respectively.

[0006] In one embodiment, the dynamic allocation logic of the driver dispatch agent includes: Based on the number of traffic participants, road structure complexity, and risk level output by the visual language model, a complexity score is generated. Based on the complexity score, the remote takeover task is assigned to the appropriate remote driver agent. In addition, it monitors in real time the number of vehicles currently under management, CPU utilization, and historical response latency of each remote driver agent. When the overall load index of a remote driver agent exceeds the threshold, it automatically migrates new tasks to the same type of remote driver agent with a load index below the threshold. After the takeover is completed, update the compatibility rate of the remote driver agent in the current scenario.

[0007] In one embodiment, the data sharing bus between the agents includes: The real-time synchronization module is configured to use a timestamp alignment mechanism to ensure that the time deviation between vehicle status data and the scene semantic description of the visual language model is less than a threshold. The data compression and encryption module is configured to perform lightweight encryption and differential compression on the transmitted data; The breakpoint resume unit is configured to automatically resend critical data from the interruption period when the connection is restored after a network interruption. In one embodiment, the contextualized evaluation dimensions of the driver evaluation agent include: Urban congestion scenarios: following distance control accuracy, acceleration and deceleration smoothness, and lane departure frequency; High-speed scenarios: lane keeping centering rate, emergency lane recognition accuracy, and overtaking decision rationality; Remote takeover scenario: visual language model recognition latency, operation command response time, and vehicle stability after takeover. In one embodiment, training data for each scenario is constructed based on a preset scenario, and multiple remote driver agents adapted to different scenarios are trained using reinforcement learning based on the training data.

[0008] A method for controlling unmanned vehicles based on multi-agent cooperation, the method comprising: The autonomous driving agent performs autonomous driving and monitors the complexity of the driving scenario in real time; When the complexity of the scenario exceeds a preset threshold, the autonomous driver agent sends a takeover request to the driver dispatch agent; The driver dispatch agent selects the appropriate remote driver agent based on the task complexity and the load status of the remote driver agent. The selected remote driver agent parses vehicle sensor data through a visual language model, generates scene semantic descriptions and operation instructions, and executes remote driving operations. The driver evaluation agent assesses the performance of remote driving operations in real time and feeds the results back to the driver dispatch agent.

[0009] In one embodiment, the method includes: generating a cross-modal feature vector to map the visual features of an image, the point cloud features of a radar, and text commands to a unified semantic space. Based on the cross-modal feature vectors, the interaction relationships of traffic participants and the evolution trend of potential risks are analyzed. The reasoning results are converted into structured text and natural language descriptions, and then output to the driving control module and the remote operator interface, respectively.

[0010] In one embodiment, the method includes: generating a complexity score based on the number of traffic participants, road structure complexity, and risk level output by a visual language model; and assigning remote takeover tasks to suitable remote driver agents based on the complexity score. In addition, the driver dispatch agent monitors the number of vehicles currently taken over, CPU utilization, and historical response latency of each remote driver agent in real time. When the comprehensive load index of a remote driver agent exceeds the threshold, it automatically migrates new tasks to the same type of remote driver agent with a load index below the threshold. After the takeover is completed, update the compatibility rate of the remote driver agent in the current scenario.

[0011] A computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program performing the following steps: The autonomous driving agent performs autonomous driving and monitors the complexity of the driving scenario in real time; When the complexity of the scenario exceeds a preset threshold, the autonomous driver agent sends a takeover request to the driver dispatch agent; The driver dispatch agent selects the appropriate remote driver agent based on the task complexity and the load status of the remote driver agent. The selected remote driver agent parses vehicle sensor data through a visual language model, generates scene semantic descriptions and operation instructions, and executes remote driving operations. The driver evaluation agent assesses the performance of remote driving operations in real time and feeds the results back to the driver dispatch agent.

[0012] The aforementioned unmanned vehicle control system, method, and computer equipment based on multi-agent collaboration significantly improve the decision-making ability of autonomous driving systems in complex scenarios through the design of a multi-agent collaborative driving system. The remote driver agent can effectively take over and handle special scenarios that are difficult for the autonomous driving system to cope with, such as construction areas and traffic accident scenes. It realizes intelligent driving rights allocation based on scenario complexity and system status, improving the overall safety and reliability of the system. Through the real-time evaluation mechanism of the driver evaluation agent, it ensures the quality of driving operations and provides a decision-making basis for driver dispatch. Attached Figure Description

[0013] Figure 1 This is a structural block diagram of an unmanned vehicle control system based on multi-agent cooperation in one embodiment; Figure 2 This is a flowchart illustrating an autonomous vehicle control method based on multi-agent cooperation in one embodiment; Figure 3 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0015] In one embodiment, such as Figure 1 As shown, an unmanned vehicle control system based on multi-agent collaboration is provided, including: an autonomous driver agent, a remote driver agent, a driver scheduling agent, and a driver evaluation agent.

[0016] The autonomous driving agent is deployed locally in the vehicle and performs autonomous driving operations in normal scenarios, including path planning, speed control, and basic environmental perception. This agent is responsible for maintaining the vehicle's safe operation under normal driving conditions, acquiring information about the surrounding environment through onboard sensors and making corresponding decisions.

[0017] The remote driver agent is deployed at a remote control center and integrates a visual language model. This model receives multimodal data collected by vehicle sensors and outputs scene semantic descriptions and operation commands. Based on the output of the visual language model, the remote driver agent executes remote driving operations when the autonomous driver agent triggers a takeover request.

[0018] The visual language model comprises three core units. The multimodal feature fusion unit maps visual features from images, point cloud features from radar, and text commands to a unified semantic space, generating cross-modal feature vectors. The scene reasoning unit, based on these cross-modal feature vectors, analyzes the interaction relationships among traffic participants and the evolution trends of potential risks. The semantic conversion unit converts the reasoning results into structured text and natural language descriptions, outputting them to the driving control module and the remote operator interface, respectively. The structured text contains precise numerical information such as the location, speed, and behavioral intentions of scene elements, which the driving control algorithm can directly process. The natural language description summarizes the current traffic conditions, potential risks, and suggested actions in a human-understandable way, such as "There is a vehicle slowing down 50 meters ahead; a bicycle on the right may change lanes. It is recommended to slow down and maintain a safe distance on the right." The remote operator can quickly understand the scene and make decisions based on this information, sending steering, acceleration, and deceleration commands to the vehicle through the remote control interface.

[0019] The driver dispatch agent dynamically allocates driving rights between the autonomous driver agent and the remote driver agent based on task complexity, the load status of the remote driver agent, and historical scenario adaptability. This agent continuously monitors the complexity of the current driving scenario, the processing capacity load of each agent, and past performance data in similar scenarios to determine the most suitable driving control mode for the current situation.

[0020] The driver evaluation agent evaluates the operational performance of the autonomous driver agent and the remote driver agent in real time based on preset scenario-based evaluation dimensions, and feeds back the evaluation results to the driver dispatch agent.

[0021] In one embodiment, the evaluation dimensions include key indicators for three typical scenarios: in urban congestion scenarios, the evaluation assesses following distance control accuracy, acceleration and deceleration smoothness, and lane departure frequency; in highway scenarios, the evaluation assesses lane centering rate, emergency lane recognition accuracy, and overtaking decision rationality; and in remote takeover scenarios, the evaluation assesses visual language model recognition latency, operation command response time, and vehicle stability after takeover. These multi-dimensional evaluation results provide decision-making basis for the driver dispatch agent.

[0022] In one embodiment, data sharing between agents is achieved through a dedicated data bus comprising three key modules. The real-time synchronization module employs a timestamp alignment mechanism to ensure that the time deviation between vehicle status data and the scene semantic description in the visual language model is less than a preset threshold, typically controlled at the millisecond level, to guarantee the timeliness of control decisions. The data compression and encryption module performs lightweight encryption and differential compression on transmitted data, improving transmission efficiency while ensuring data security. The breakpoint resume unit automatically retransmits critical data from the interrupted period when the connection is restored after a network interruption, ensuring that the system can maintain basic functionality even in unstable network environments.

[0023] In one embodiment, the system constructs training data for various scenarios based on preset scenarios, and uses reinforcement learning methods to train multiple remote driver agents adapted to different scenarios based on this training data. These specialized agents can provide more accurate remote driving support for different driving environments. During training, the system simulates various complex traffic scenarios, allowing the agents to repeatedly learn the best driving strategies, and guides them to form safe and efficient driving behavior patterns through a reward function.

[0024] During system operation, the autonomous driver agent initially attempts to handle all driving tasks. When encountering complex scenarios or uncertainties, it sends a takeover request to the driver dispatch agent. After assessing the situation, the driver dispatch agent may transfer control to the remote driver agent. The remote driver agent analyzes vehicle sensor data using a visual language model, generates scene understanding and driving decisions, and sends control commands. The driver evaluation agent monitors the performance of each agent throughout the process and uses the evaluation results to continuously optimize the dispatch strategy. Through this multi-agent collaborative mechanism, the system can achieve stable operation of autonomous vehicles in complex and changing environments while ensuring safety.

[0025] In one embodiment, the dynamic allocation logic of the driver dispatch agent includes the following steps: The first step is complexity score generation and task allocation. The driver dispatch agent generates a scenario complexity score based on the number of traffic participants, road structure complexity, and risk level output by the visual language model. The complexity score uses a 100-point scale, with the number of traffic participants accounting for 40 points, road structure complexity for 30 points, and risk level for 30 points. A perfect score is awarded when the number of traffic participants exceeds 10; a perfect score is awarded for road structure complexity when the road structure includes four or more lanes or complex intersections; and a perfect score is awarded for risk level when the visual language model detects high-risk behaviors (such as sudden lane changes or running red lights). Based on the final calculated complexity score, the driver dispatch agent assigns remote takeover tasks to the appropriate remote driver agent. For example, high-complexity scenarios with a complexity score of 80-100 are assigned to Class A remote driver agents, medium-complexity scenarios with a score of 50-79 are assigned to Class B remote driver agents, and low-complexity scenarios with a score of 0-49 are assigned to Class C remote driver agents.

[0026] The second stage involves real-time load monitoring and task migration for remote driver agents. The driver dispatch agent monitors three metrics in real-time for each remote driver agent: the number of vehicles currently under takeover, CPU utilization, and historical response latency. The system calculates a comprehensive load index with the following weights: 40% for the number of vehicles currently under takeover, 30% for CPU utilization, and 30% for historical response latency. When the comprehensive load index of a remote driver agent exceeds a preset threshold (default 85%), the driver dispatch agent automatically migrates the new task to a similar remote driver agent with a load index below the threshold. During task migration, the system ensures a seamless switchover, avoiding control gaps during takeover. For example, when the load index of remote driver agent A1 reaches 87%, a new, highly complex task will be assigned to remote driver agent A2, whose load index is 65%, instead of continuing to be assigned to agent A1.

[0027] The third step is the adaptation rate update after takeover. After a remote takeover task is completed, the driver dispatch agent updates the adaptation rate of the remote driver agent in the current scenario based on performance metrics during the takeover process (including takeover success rate, average response time, and operational accuracy). The adaptation rate is expressed as a percentage, with an initial value of 80%. It is increased by 1-5% after each successful takeover and safe completion of the task, and decreased by 5-10% after each problematic takeover. The adaptation rate data is stored in the system database as an important reference for the allocation of similar scenario tasks in the future. For example, if a Class A remote driver agent has successfully taken over multiple times in a complex highway scenario, and its adaptation rate in this scenario increases from 80% to 95%, the system will prioritize assigning takeover tasks in complex highway scenarios to this agent.

[0028] Through the dynamic allocation logic of the above three stages, the driver dispatch agent can intelligently allocate and dynamically adjust remote takeover tasks based on the complexity of the scenario, the real-time load status of the remote driver agent, and the historical adaptability rate, thereby improving the safety, reliability, and efficiency of the entire remote driving system.

[0029] In one embodiment, such as Figure 2 As shown, a method for controlling unmanned vehicles based on multi-agent cooperation is provided, including: Step 202: The autonomous driver agent performs autonomous driving and monitors the complexity of the driving scenario in real time.

[0030] During autonomous driving, the autonomous driver agent collects environmental data through onboard sensors, including camera images, radar detection results, and vehicle status information. Simultaneously, the autonomous driver agent calculates the complexity of the current driving scenario in real time based on a pre-set algorithm. Complexity assessment metrics include multiple dimensions such as traffic density, weather conditions, road difficulty, and the uncertainty of traffic participant behavior.

[0031] Step 204: When the scene complexity exceeds a preset threshold, the autonomous driver agent sends a takeover request to the driver scheduling agent.

[0032] The preset threshold can be dynamically adjusted based on vehicle performance, driving environment characteristics, and safety requirements, and is typically set to 80% to 90% of the reliable processing capacity of the autonomous driving system. The takeover request includes information such as the current scenario complexity score, key risk point analysis, and vehicle status parameters, so that the driver can dispatch the agent to make decisions.

[0033] Step 206: The driver dispatch agent selects a suitable remote driver agent based on the task complexity and the load status of the remote driver agent.

[0034] The driver dispatch agent maintains a remote driver resource pool, recording information such as the professional skill rating, current load status, and historical performance data for each remote driver agent. The selection process employs a multi-factor weighted scoring mechanism, comprehensively considering factors such as the match between scenario complexity and driver capabilities, the driver's current load level, and response speed, to select the most suitable remote driver agent for handling the current situation.

[0035] Step 208: The selected remote driver agent parses the vehicle sensor data through a visual language model, generates scene semantic descriptions and operation instructions, and executes remote driving operations.

[0036] Step 210: The driver evaluation agent evaluates the remote driving operation performance in real time and feeds the results back to the driver dispatch agent.

[0037] It should be understood that, although Figure 2 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 2 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.

[0038] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 3 As shown, the computer device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database stores adaptation data for remote driver agents. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a multi-agent cooperative unmanned vehicle control method.

[0039] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0040] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0041] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0042] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0043] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. An unmanned vehicle control system based on multi-agent cooperation, characterized in that, The system includes: An autonomous driving agent is deployed locally in the vehicle and configured to perform autonomous driving operations in normal scenarios; the autonomous driving operations include path planning, speed control, and basic environmental perception. A remote driver agent, deployed at a remote control center, integrates a visual language model. This model is configured to receive multimodal data collected by vehicle sensors and output scene semantic descriptions and operation commands. Based on the output of the visual language model, the remote driver agent performs remote driving operations when the autonomous driver agent triggers a takeover request. The visual language model includes a multimodal feature fusion unit, a scene reasoning unit, and a semantic conversion unit. The multimodal feature fusion unit receives multimodal data collected by vehicle sensors and generates cross-modal feature vectors. The scene reasoning unit outputs the interaction relationships and potential risk evolution trends of traffic participants based on the cross-modal feature vectors. The semantic conversion unit converts the reasoning results into structured text and natural language descriptions. The driver dispatch agent is configured to dynamically allocate driving rights between the autonomous driver agent and the remote driver agent based on task complexity, the load status of the remote driver agent, and the historical scenario adaptation rate. The dynamic allocation logic includes: generating a complexity score based on task complexity, performing task migration based on the load status of the remote driver agent, and updating the adaptation rate based on the historical scenario adaptation rate. The adaptation rate is used to determine the priority of task allocation. The driver evaluation agent is configured to evaluate the operational performance of the autonomous driver agent and the remote driver agent in real time based on preset scenario-based evaluation dimensions, and feed the evaluation results back to the driver dispatch agent.

2. The system according to claim 1, characterized in that, The visual language model includes: The multimodal feature fusion unit is configured to map the visual features of the image, the point cloud features of the radar, and the text commands to a unified semantic space to generate cross-modal feature vectors. The scenario reasoning unit is configured to analyze the interaction relationships and potential risk evolution trends of traffic participants based on the cross-modal feature vectors. The semantic conversion unit is configured to convert the reasoning results into structured text and natural language descriptions, which are then output to the driving control module and the remote operator interface, respectively.

3. The system according to claim 2, characterized in that, The dynamic allocation logic of the driver dispatch agent includes: Based on the number of traffic participants, road structure complexity, and risk level output by the visual language model, a complexity score is generated. Based on the complexity score, the remote takeover task is assigned to the appropriate remote driver agent. In addition, it monitors in real time the number of vehicles currently taken over, CPU utilization, and historical response latency of each remote driver agent. When the overall load index of a remote driver agent exceeds the threshold, it automatically migrates new tasks to the same type of remote driver agent with a load index below the threshold. After the takeover is completed, update the compatibility rate of the remote driver agent in the current scenario.

4. The system according to claim 1, characterized in that, The data sharing bus between the various agents includes: The real-time synchronization module is configured to use a timestamp alignment mechanism to ensure that the time deviation between vehicle status data and the scene semantic description of the visual language model is less than a threshold. The data compression and encryption module is configured to perform lightweight encryption and differential compression on the transmitted data; The breakpoint resume unit is configured to automatically resend critical data from the interruption period when the connection is restored after a network interruption.

5. The system according to claim 1, characterized in that, The contextualized evaluation dimensions of the driver evaluation agent include: Urban congestion scenarios: following distance control accuracy, acceleration and deceleration smoothness, and lane departure frequency; High-speed scenarios: lane keeping centering rate, emergency lane recognition accuracy, and overtaking decision rationality; Remote takeover scenario: visual language model recognition latency, operation command response time, and vehicle stability after takeover.

6. The system according to any one of claims 1 to 5, characterized in that, Based on preset scenarios, training data for each scenario is constructed, and reinforcement learning is used to train multiple remote driver agents adapted to different scenarios based on the training data.

7. A method for controlling unmanned vehicles based on multi-agent cooperation, characterized in that, The method includes: The autonomous driving agent performs autonomous driving and monitors the complexity of the driving scenario in real time; When the complexity of the scenario exceeds a preset threshold, the autonomous driver agent sends a takeover request to the driver dispatch agent; The driver dispatch agent selects a suitable remote driver agent based on task complexity and the load status of the remote driver agent; wherein, a complexity score is generated based on task complexity, task migration is performed based on the load status of the remote driver agent, and the adaptability is updated based on the historical scenario adaptability rate; the adaptability is used to determine the priority of task allocation. The selected remote driver agent parses vehicle sensor data through a visual language model, generates scene semantic descriptions and operation instructions, and executes remote driving operations. The visual language model includes a multimodal feature fusion unit, a scene reasoning unit, and a semantic conversion unit. The multimodal feature fusion unit receives multimodal data collected by vehicle sensors and generates cross-modal feature vectors. The scene reasoning unit outputs the interaction relationships of traffic participants and the evolution trend of potential risks based on the cross-modal feature vectors. The semantic conversion unit converts the reasoning results into structured text and natural language descriptions. The driver evaluation agent assesses the performance of remote driving operations in real time and feeds the results back to the driver dispatch agent.

8. The method according to claim 7, characterized in that, The visual language model parses vehicle sensor data, including: To map the visual features of images, the point cloud features of radar, and text commands to a unified semantic space, a cross-modal feature vector is generated. Based on the cross-modal feature vectors, the interaction relationships of traffic participants and the evolution trend of potential risks are analyzed. The reasoning results are converted into structured text and natural language descriptions, and then output to the driving control module and the remote operator interface, respectively.

9. The method according to claim 7, characterized in that, The driver dispatch agent selects a suitable remote driver agent based on task complexity and the load status of the remote driver agent, including: Based on the number of traffic participants, road structure complexity, and risk level output by the visual language model, a complexity score is generated. Based on the complexity score, the driver dispatch agent assigns remote takeover tasks to the appropriate remote driver agent. In addition, the driver dispatch agent monitors the number of vehicles currently taken over, CPU utilization, and historical response latency of each remote driver agent in real time. When the comprehensive load index of a remote driver agent exceeds the threshold, it automatically migrates new tasks to the same type of remote driver agent with a load index below the threshold. After the takeover is completed, update the compatibility rate of the remote driver agent in the current scenario.

10. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 7 to 9.

Images