Multi-agent closed site testing method and system based on cloud platform cooperative control

Abstract

The invention discloses a multi-agent closed site test method based on cloud platform cooperative control, and the method comprises the steps: building a cloud cooperative test system, building a multi-agent vehicle model and a behavior model, and carrying out the game interaction between an agent and a tested vehicle in a confrontation mode; the cloud control platform realizes real-time coordination and long-term self-optimization of a multi-agent behavior mode through a distributed scheduling and feedback learning mechanism; in the test, the real-time confrontation game and collaborative optimization adopt a master-slave game model and MINLP joint optimization framework, the priority is allocated through dynamic prediction of the time when multiple agents arrive at an interaction area, and a distributed collaborative optimization strategy of a rolling time domain is adopted, so that coordinated control and conflict avoidance of multiple workshops are realized. According to the method, the problems of insufficient dynamic interaction and limited scene coverage in a multi-agent closed site test are solved, and continuous generation of a resistance scene and adaptive optimization of agent behaviors are realized through cloud cooperative control.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle testing technology, and more specifically, to a multi-agent closed-field testing method and system based on cloud platform collaborative control. Background Technology

[0002] As autonomous driving technology advances to higher levels, intelligent connected vehicles need to cope with complex scenarios of continuous multi-vehicle interaction in real traffic (such as multi-directional vehicle competition at intersections and left-turn conflicts in oncoming traffic). Closed-field testing, as a core component of the "simulation-closed-road" three-level testing system, needs to reproduce these high-risk scenarios. However, traditional closed-field testing has the following key shortcomings:

[0003] 1) Insufficient dynamic interaction, resulting in "pseudo-interaction": Traditional trigger-based testing (such as ISO26262 and GB41798 standards) relies on predefined behaviors (such as fixed-time emergency braking and forced lane changes). The agent cannot adjust its strategy based on the real-time trajectory of the vehicle under test (VUT). For example, when the VUT slows down to avoid a collision, the agent still drives along the preset trajectory, resulting in "formal conflict but no actual behavioral coupling," failing to expose security vulnerabilities in the VUT's dynamic response.

[0004] 2) Inefficient multi-agent collaboration and prone to deadlock: Existing methods mostly support 1-2 agents in competition and use fixed trajectory control. When the number of agents is ≥3, trajectory interference (such as multi-directional vehicle congestion at intersections) or deadlock (agents blocking each other and unable to get out of trouble) is likely to occur, making it impossible to simulate the collaborative and competitive relationship of multiple vehicles in real traffic.

[0005] 3) Weak scene adversarial and diversity: Trajectory playback methods based on the Natural Driving Database (NDD) can only reproduce conventional scenes; optimization methods based on a single risk indicator (such as Time to Collision (TTC)) generate scenes that are high-risk but semantically unreasonable (such as agents accelerating suddenly for no reason), making it difficult to cover the randomness of behavior and the diversity of strategies in real traffic.

[0006] Furthermore, simulation testing cannot verify the performance of the VUT's physical layer (such as brake-by-wire response) and perception layer (such as the camera's adaptation to real lighting); public road testing suffers from high safety risks and low exposure rates in edge scenes (only 1-2 unprotected left-turn conflicts per million kilometers). Therefore, there is an urgent need for a closed-field testing method that can achieve continuous dynamic interaction among multiple agents, high adversarialness, and high realism to fill the gaps in existing technologies. Summary of the Invention

[0007] Due to the problems existing in the prior art, this invention proposes a multi-agent closed-field testing method and system based on cloud platform collaborative control, which is suitable for safety verification of autonomous driving systems in high-risk interaction scenarios. It solves the problems of insufficient dynamic interaction and limited scenario coverage in multi-agent closed-field testing, and realizes continuous generation of adversarial scenarios and adaptive optimization of agent behavior through cloud collaborative control.

[0008] To achieve the above objectives, on the one hand, the present invention provides a multi-agent closed-field testing method based on cloud platform collaborative control, comprising:

[0009] Step S11: Construct a cloud collaborative testing system to enable information interaction between the cloud platform and multiple intelligent agents; the cloud platform includes a data middleware, a time calibration server, and a cloud control platform.

[0010] Step S12: Establish a multi-agent vehicle model and behavior model: The intelligent agent vehicle model includes a vehicle kinematics model and a vehicle path and local trajectory generation model; the intelligent agent vehicle behavior model includes two states: free-flow mode and adversarial mode. The state is dynamically switched by the cloud control platform based on the spatial relationship between the tested vehicle and the interaction area; in the adversarial mode, the intelligent agent interacts with the tested vehicle in a game-like manner through instructions issued by the cloud control platform. Its control objective is to create a controlled risk situation. The control logic includes adversarial triggering and risk assessment, acceleration and steering coordinated control, safety constraints and action boundaries; the cloud control platform detects the relative speed and distance between intelligent agents in real time. When speed conflicts or path overlaps occur, the vehicle priority is adjusted through speed allocation and time delay scheduling to avoid local congestion; the cloud control platform realizes real-time coordination and long-term self-optimization of the multi-agent behavior mode through distributed scheduling and feedback learning mechanisms.

[0011] Step S13, Real-time adversarial game and collaborative optimization during testing: The real-time adversarial game and collaborative optimization algorithm adopts the Stackelberg master-slave game model and the mixed integer nonlinear programming joint optimization framework to realize real-time adversarial interaction and collaborative control between multiple agents and the tested vehicle; Multi-agent priority collaboration is a distributed scheduling framework based on the cloud control platform. By dynamically predicting and allocating priorities based on the arrival time of multiple agents in the interaction area, and adopting a rolling time-domain distributed collaborative optimization strategy, it realizes coordinated control and conflict avoidance of multiple workshops.

[0012] Step S14: After the test is completed, perform a scenario evaluation based on historical test data.

[0013] The above technical solutions can realize the dynamic behavior coupling between multiple agents and VUT, eliminate "pseudo-interactions", solve the trajectory interference and deadlock problems in multi-agent collaboration, improve the intensity of scene confrontation and behavior diversity, and provide a high-risk testing environment for autonomous driving systems that closely resembles real traffic.

[0014] Furthermore, the number of intelligent agents is no less than three; the intelligent agents include dummy humans or dummy vehicles, thus better simulating the cooperative and competitive relationships among multiple vehicles in real traffic.

[0015] Furthermore, in step S12, the vehicle path generation uses a multi-stage A* algorithm based on a high-precision map to generate a global reference route, and smooths it using a cubic Bezier curve to ensure that the trajectory continuity and curvature constraints meet the vehicle's kinematic characteristics; the local trajectory generation uses the Frenet coordinate system to project the vehicle state as longitudinal arc length and lateral offset, and applies no reversing constraint, speed and acceleration constraint, and lateral deviation constraint.

[0016] Furthermore, in step S12, when the Euclidean distance from the tested vehicle to the interaction area is not greater than a specific distance, the agent is triggered to enter adversarial mode; when the Euclidean distance from the tested vehicle to the interaction area is greater than a specific distance, the agent automatically switches to free-flow mode. This dynamic switching mechanism ensures that multiple agents can respond in advance before entering high-risk areas and resume normal driving in a timely manner after leaving the area, thereby maintaining the continuity and safety of interaction.

[0017] Furthermore, in step S12, the adversarial triggering and risk assessment control uses the post-crossover time as the risk indicator. The cloud control platform dynamically allocates conflict targets to each agent, ensuring that the adversarial behaviors of different agents within the interaction zone do not overlap or interfere with each other. In the acceleration and steering cooperative control, the agent calculates the control input based on the game strategy. ,in It is longitudinal acceleration; For the front wheel steering angle, the control law is as follows:

[0018]

[0019] in This represents the baseline longitudinal acceleration value for the agent to maintain normal driving conditions without adversarial behavior. The cloud control platform assigns the expected interaction risk level to the intelligent agent. The actual interaction risk level between the agent and the VUT at the current moment. The target vehicle speed that the intelligent agent hopes to achieve under the current adversarial conditions. The longitudinal velocity of the intelligent vehicle at time t. To control the gain coefficient;

[0020] By coordinating longitudinal acceleration and lateral steering, the intelligent agent actively performs high-risk actions, including approaching, overtaking, changing lanes, or cutting in, forming a continuous interactive scenario.

[0021] Furthermore, in step S12, the distributed scheduling and feedback learning mechanism first constructs a mode switching matrix based on the number of agents N and the testing period T. ,in This indicates that the i-th agent is at time step [i]. The cloud control platform dynamically updates the control mode and priority of each intelligent agent based on the matrix; the matrix is ​​updated in real time in each rolling time domain cycle, enabling the cloud control platform to dynamically adjust the mode, task priority and control strategy of each vehicle, ensuring the global coordination and local feasibility of the multi-vehicle system.

[0022] Further, step S13 includes:

[0023] Step S131: Master-slave game modeling and system initialization;

[0024] Step S132: In the game optimization layer, the intelligent agent, as the leader, and the tested vehicle, as the follower, each use their own utility function as the optimization objective.

[0025] Step S133: In order to explicitly embed the reaction behavior of the follower (VUT) into the optimization problem of the leader (agent), the Stackelberg master-slave game framework is adopted and the Karush-Kuhn-Tucker condition is introduced to achieve nested solution; during the solution process, collision avoidance constraint, no reversing constraint and deadlock penalty constraint are applied simultaneously.

[0026] Step S134: The cloud control platform adopts a rolling time-domain collaborative control mechanism. In each time step, the initial state is updated according to the solution results of the previous cycle to predict the trajectory for several future steps and adjust the control commands in real time. The cloud control platform synchronizes the state information of each intelligent agent and the vehicle under test through the V2X network. If the model prediction error is detected to exceed the threshold, it will automatically trigger re-optimization.

[0027] Furthermore, in step S13, the multi-agent priority cooperative control includes:

[0028] The cloud control platform first predicts the arrival time of each vehicle at the center of the interaction area based on the real-time location, speed, and reference path information of multiple agents; then, it assigns priorities to each agent according to the order of arrival time. Vehicles arriving earlier have higher priority, thus determining the interaction order; after the priority is determined, the cloud control platform establishes a priority-based scheduling matrix. This matrix, in conjunction with the aforementioned game optimization output, serves as the upper-level constraint input for multi-vehicle local programming.

[0029] During the execution phase of the multi-agent system, the cloud control platform adopts a rolling time-domain distributed collaborative optimization architecture to achieve coordinated control among the agents, realize conflict avoidance between multiple vehicles, and thus improve the stability and traffic efficiency of continuous interaction between multiple vehicles in closed field scenarios. When the cloud control platform detects that multiple high-priority vehicles are approaching simultaneously in the interaction area, it will automatically divide the area into sub-regions and implement block optimization to control the scale of the optimization problem, thereby reducing the computational load and ensuring real-time performance. A soft constraint mechanism is introduced during the optimization process. When a low-priority vehicle cannot meet the avoidance constraint in the current prediction time domain, its target trajectory is dynamically adjusted through a penalty term, so that it automatically yields or slows down in the next time domain.

[0030] After priority scheduling and distributed collaborative optimization, the cloud control platform outputs the optimal state trajectory set and control sequence of all intelligent agents, and sends them to the site execution module in real time in the next control cycle. The cloud control platform also monitors the interaction status and driving trajectory consistency of each vehicle. When a local blockage or execution deviation is detected, the local re-optimization process is immediately initiated to restore system coordination.

[0031] Furthermore, in step S14, after the test is completed, the test vehicle data, multi-agent data, and control data are stored in the database and analyzed based on historical data and vehicle-mounted sensing device data.

[0032] On the other hand, the present invention provides a multi-agent closed-field testing system based on cloud platform collaborative control, for implementing the multi-agent closed-field testing method described above; including:

[0033] The cloud-based collaborative testing system enables information interaction between the cloud platform and multiple intelligent agents; the cloud platform includes a data middleware, a time calibration server, and a cloud control platform.

[0034] The site execution module includes the vehicle under test and multiple agents based on the intelligent vehicle model and behavior patterns described above;

[0035] The cloud coordination framework system ensures real-time adversarial game and collaborative optimization as described above during testing.

[0036] Compared with the prior art, the present invention has the following technical effects:

[0037] (1) This invention provides a cloud collaborative closed field test method for continuous interaction scenarios of multiple agents (3 or more), which is applicable to the safety verification and performance evaluation of L3~L4 level autonomous driving systems in high-risk traffic scenarios (such as unprotected left turns and signalless intersections).

[0038] (2) This invention can realize the dynamic behavior coupling of multiple agents and VUT, eliminate "pseudo-interaction", solve the trajectory interference and deadlock problem in multi-agent collaboration, improve the scene confrontation intensity and behavior diversity, and provide a high-risk test environment for autonomous driving system that is close to real traffic.

[0039] (3) This invention improves the adversarial intensity of closed-field testing and enhances the diversity of scenarios through cloud collaborative control and dynamic interaction of multiple agents. It provides a risk assessment environment for autonomous driving systems that is closer to real traffic and is suitable for safety certification, algorithm iteration and edge scenario testing of intelligent connected vehicles. Attached Figure Description

[0040] The invention, its features and advantages will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings.

[0041] Figure 1 This is a flowchart of a multi-agent closed-field testing method based on cloud platform collaborative control provided in an embodiment of the present invention;

[0042] Figure 2 This is a structural diagram of the network communication between the cloud platform and the multi-agent system provided in the embodiments of the present invention;

[0043] Figure 3 This is a flowchart of the design of a multi-agent vehicle model and behavior model in cloud control testing provided in this embodiment of the invention;

[0044] Figure 4 This is a flowchart of adversarial game and collaborative optimization in cloud control testing provided in the embodiments of the present invention. Detailed Implementation

[0045] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this is not intended to limit the invention.

[0046] In the following detailed description, numerous specific details are set forth to provide a more thorough understanding of the invention. However, it will be apparent to those skilled in the art that well-known algorithms and models are not shown in detail to avoid obscuring the gist of the invention. The device structure and accompanying drawings of the present invention primarily describe the principles of the invention. While the technical details of the design principles are not fully described, the specific details of the power mechanism, power supply system, and control system of the device or system are readily apparent to those skilled in the art who understand the principles of the invention. Furthermore, the technologies not detailed in the following effect embodiments are readily available prior art.

[0047] Furthermore, the execution order of actions, steps, etc. in the systems and methods shown in the claims, specification, and drawings can be implemented in any order, unless a specific order is explicitly specified, and as long as the output of the preceding processing is not used in the subsequent processing.

[0048] Example

[0049] See Figure 1 This embodiment provides a multi-agent closed-field testing method based on cloud platform collaborative control, including:

[0050] Step S11: Construct a cloud collaborative testing system architecture and establish information interaction between the cloud and multiple intelligent agents;

[0051] Step S12: Establish the multi-agent vehicle model and behavior model design: The intelligent agent vehicle model includes a vehicle kinematics model and a vehicle path and local trajectory generation model; the intelligent agent vehicle behavior model includes two states: free-flow mode and adversarial mode. The state is dynamically switched by the cloud control platform based on the spatial relationship between the tested vehicle and the interaction area; In the adversarial mode, the intelligent agent interacts with the tested vehicle in a game-like manner through instructions issued by the cloud control platform. Its control objective is to create a controlled risk situation. The control logic includes adversarial triggering and risk assessment, acceleration and steering coordinated control, safety constraints and action boundaries; The cloud control platform detects the relative speed and distance between intelligent agents in real time. When speed conflicts or path overlaps occur, the vehicle priority is adjusted through speed allocation and time delay scheduling to avoid local congestion; The cloud control platform realizes real-time coordination and long-term self-optimization of the multi-agent behavior mode through distributed scheduling and feedback learning mechanisms;

[0052] Step S13, Real-time adversarial game and collaborative optimization: The real-time adversarial game and collaborative optimization algorithm adopts the Stackelberg master-slave game model and the mixed integer nonlinear programming joint optimization framework to realize real-time adversarial interaction and collaborative control between multiple agents and the tested vehicle; the multi-agent priority collaboration is a distributed scheduling framework based on the cloud control platform. By dynamically predicting and allocating priorities based on the arrival time of multiple agents in the interaction area, and adopting a rolling time domain distributed collaborative optimization strategy, the coordinated control and conflict avoidance of multiple workshops are realized.

[0053] Step S14: After the test is completed, a scenario evaluation is performed based on historical test data.

[0054] In this embodiment, the multi-agent closed-field testing method based on cloud platform collaborative control first establishes information interaction between the cloud module, the 5G communication network, and the field-side module to achieve real-time data closed loop between the vehicle under test and multiple intelligent target vehicles. Then, relying on the establishment of vehicle kinematics and behavior pattern models of multiple agents, the adversarial mode and free flow mode are dynamically switched based on the distance between the VUT and the interaction area. During the test, real-time adversarial game and collaborative optimization are performed, integrating adversarial objectives, smoothness objectives, rule compliance objectives, and safety constraints. The Stackelberg game model is used, combined with the mixed integer nonlinear programming (MINLP) model to optimize the agent adversarial trajectory. Finally, after the test is completed, the test vehicle data, multi-agent data, and control data are stored in the database, and detailed analysis is performed based on historical data and vehicle-mounted sensing device data.

[0055] The multi-agent closed-field testing method based on cloud platform collaborative control proposed in this invention realizes dynamic behavior coupling between multiple agents and VUT, eliminating "pseudo-interaction"; solves trajectory interference and deadlock problems in multi-agent collaboration; enhances scene adversarial intensity and behavioral diversity, and provides a high-risk testing environment for autonomous driving systems that closely resembles real traffic.

[0056] See Figure 2 In the embodiments proposed by the present invention, step S11 may include:

[0057] Step S111: Build a cloud platform and a multi-agent system; the cloud platform includes a data middleware, a time calibration server and a cloud control platform, and may also include a digital twin system to display the actual mapping.

[0058] Step S112: Connect the multi-agent system to the cloud platform and interact with the cloud control platform in real time through the data middleware.

[0059] In this embodiment, the data uploaded and transmitted in the cloud platform is monitored to obtain real-time multi-agent motion state data and cloud collaborative control data.

[0060] See Figure 3 In this embodiment, step S12 may include:

[0061] Step 121: The vehicle kinematics model, combined with the agent's kinematic characteristics and the vehicle's longitudinal and lateral dynamic constraints, constructs the state equations. The state vector of the agent vehicle is defined as follows: ,in These are the vehicle's position coordinates in the site coordinate system; For heading angle; For longitudinal velocity. Define the control input vector as... ,in It is longitudinal acceleration; This refers to the steering angle of the front wheels.

[0062] Constructing the vehicle's kinematic equations:

[0063] ,

[0064] Where L is the vehicle wheelbase.

[0065] Step 122: At the path generation level, a global reference route is generated using a multi-stage A* algorithm based on a high-resolution map (HD Map), and smoothed using a cubic Bezier curve to ensure trajectory continuity and that curvature constraints satisfy the vehicle's kinematic characteristics. Local trajectory generation uses the Frenet coordinate system, projecting the vehicle state as longitudinal arc length and lateral offset, and applying the following constraints:

[0066] ① No reversing constraint:

[0067] ,in It represents the instantaneous longitudinal velocity of the vehicle along the reference path, ensuring that the vehicle's longitudinal movement is without backward movement.

[0068] ② Velocity and acceleration constraints:

[0069]

[0070] ③ Lateral deviation constraint:

[0071] ,in This represents the lateral offset of the vehicle relative to the reference path. The maximum lateral offset allowed within the lane boundary range.

[0072] Step 123: Define behavioral pattern variables for each agent. , among which, when When, it indicates that the agent is in free-flow mode; when This indicates that the agent is in adversarial mode. The cloud control platform, based on a high-precision map (HD Map) and real-time positioning data, identifies the spatial relationship between the VUT and the interaction zone. The interaction zone can be an unsignalized intersection, merging zone, or unprotected left-turn area. The system calculates the Euclidean distance from the VUT to the interaction zone in real time.

[0073]

[0074] when (Preferably 5 meters) triggers the agent to enter adversarial mode; when When necessary, the system automatically switches to free-flow mode. This dynamic switching mechanism ensures that multiple agents can respond in advance before entering high-risk areas and resume normal operation promptly after leaving the area, thus maintaining continuity of interaction and safety. In adversarial mode, the agents engage in game-like interaction with the VUT through commands issued by the cloud control platform. The control objective is to create controlled risk scenarios to examine the VUT's decision-making and safety in sudden interactions.

[0075] In adversarial mode, the agent interacts with the VUT in a game-like manner through instructions issued by the cloud control platform. Its control objective is to create controlled risk situations in order to examine the VUT's decision-making and security in sudden interactions.

[0076] The control logic specifically includes:

[0077] (1) Countermeasures Triggering and Risk Assessment

[0078] Post Encroachment Time (PET) is used as a risk indicator. When PET ≤ 5 seconds, the system considers the interaction to have entered a high-risk state. The cloud control platform dynamically assigns conflict targets to each agent to ensure that the adversarial behaviors of different agents in the interaction zone do not overlap or interfere with each other.

[0079] (2) Acceleration and steering coordinated control

[0080] The agent calculates control inputs based on game strategy. The control law is adopted:

[0081]

[0082] in This represents the baseline longitudinal acceleration value for the agent to maintain normal driving conditions without adversarial behavior. The cloud control platform assigns the expected interaction risk level to the intelligent agent. The actual interaction risk level between the agent and the VUT at the current moment. The target vehicle speed that the intelligent agent hopes to achieve under the current adversarial conditions. The longitudinal velocity of the intelligent vehicle at time t. This is to control the gain coefficient.

[0083] By coordinating longitudinal acceleration and lateral steering, the intelligent agent can proactively perform high-risk actions such as approaching, overtaking, changing lanes, or cutting in, thereby forming a continuous interactive scenario.

[0084] (3) Safety constraints and action boundaries

[0085] Security constraints must still be met during the confrontation process. If the detected distance is less than the threshold, the cloud control center immediately triggers an emergency braking command. Furthermore, to prevent overly aggressive control actions, upper limits are set for both lateral offset rate and acceleration to ensure that the countermeasures are controllable and reproducible.

[0086] Step 124: After the VUT leaves the interaction area, the agent switches to free-flow mode and resumes normal cruising to maintain traffic order and driving continuity. The agent maintains its cruising speed, such as... The system maintains smooth acceleration and deceleration through longitudinal PID control. Lateral deviation is controlled within ±0.3m to ensure lane stability. The cloud control platform monitors the relative speed and distance between intelligent agents in real time. When speed conflicts or path overlaps occur, vehicle priorities are adjusted through speed allocation and time delay scheduling to avoid local congestion. To avoid sudden changes in movement during transitions, the system smoothly interpolates control inputs within a 2-second window when entering or exiting the interaction zone, ensuring a continuous longitudinal acceleration rate of change, thereby improving vehicle dynamic comfort and scene reproduction stability.

[0087] Step 125: Construct a mode switching matrix based on the number of agents N and the test period T. ,in This indicates that the i-th agent is at time step [i]. The cloud control platform dynamically updates the control mode and priority of each intelligent agent based on the matrix. This matrix is ​​updated in real time within each rolling time domain cycle, enabling the cloud control platform to dynamically adjust the mode, task priority, and control strategy of each vehicle, ensuring global coordination and local feasibility of the multi-vehicle system.

[0088] See Figure 4 In this embodiment, step S13 may include:

[0089] Step 131: At the start of the closed-field test, the cloud control platform establishes a master-slave hierarchical decision-making model for the multi-agent system, defining the agents as leaders and the VUTs as followers. The agents predict the VUT's response behavior through a game-theoretic strategy and adjust their acceleration and steering inputs in real time based on the prediction results to achieve the expected interaction scenario. The system first initializes the state vectors and control variables of each vehicle, where the agent's state vector is:

[0090]

[0091] The control input is:

[0092]

[0093] Discretization update of the kinematic model:

[0094]

[0095] Where L is the wheelbase. It's a long walk away.

[0096] In the game optimization layer, the leader and followers each use their own utility function as the optimization objective. The optimization objective on the VUT side is to maintain smooth and safe driving, and its cost function is defined as:

[0097]

[0098] in, These are the reference velocity and the reference trajectory, respectively. These are weighting coefficients. This is the longitudinal acceleration control variable of the VUT at time t, used to describe the acceleration or deceleration behavior of the follower vehicle in the longitudinal direction. Positive values ​​indicate vehicle acceleration, and negative values ​​indicate vehicle braking. It is an important control variable affecting the ride comfort and safety of the VUT.

[0099] The optimization objective of the intelligent agent takes into account the creation of interaction risks, smooth driving, and rule compliance. The cost function is defined as follows:

[0100]

[0101] in As a risk mitigation indicator, Let be the post-crossover time between the agent and the VUT at time t. This is a safety threshold; For trajectory smoothing index, Let be the longitudinal acceleration of the intelligent vehicle at time t. Let be the lateral acceleration of the intelligent vehicle at time t; For compliance indicators, , The actual position coordinates of the intelligent vehicle in the site coordinate system; , , These are the weight parameters.

[0102] Step 132: In order to explicitly embed the reaction behavior of the follower (VUT) into the optimization problem of the leader (agent), the system adopts the Stackelberg master-slave game framework and introduces Karush–Kuhn–Tucker (KKT) conditions to achieve nested solution.

[0103] The optimization problem of VUT can be formalized as:

[0104]

[0105] Using KKT conditions

[0106]

[0107] in Let Lagrange multiplier vectors be the constraints. Let VUT be the set of inequality constraint functions.

[0108] By mapping the optimal response to the upper-level problem of the agent, the two-level game problem can be transformed into a single-level mixed-integer nonlinear programming problem. An optimization model is constructed using the CasaADi symbolic modeling tool, and nonlinear programming iterations are performed using the Ipopt solver. Within each rolling time window, the cloud control platform re-solves the optimization problem, updates the optimal control inputs for the agent and the VUT, and achieves continuous dynamic evolution of the adversarial scenario.

[0109] The following constraints are applied simultaneously during the solution process:

[0110] ① Collision avoidance constraint: The distance between any two workshops satisfies ,in ;

[0111] ② No reversing constraint: ;

[0112] ③ Deadlock penalty constraint: When the dwell time in the interaction area exceeds At that time, the penalty item is activated. This is to prevent vehicles from being stationary for extended periods.

[0113] Step 133: The cloud control platform employs a rolling time-domain collaborative control mechanism. Within each time step, the system updates the initial state based on the solution results of the previous cycle to predict the trajectory for the next few steps and adjusts the control commands in real time. The cloud control platform synchronizes the state information of each agent and the VUT via a V2X network. If the model prediction error exceeds a threshold, it automatically triggers re-optimization to ensure consistency between the global trajectory and local behavior.

[0114] Step 134: The multi-agent priority coordination mechanism is based on the distributed scheduling framework of the cloud control platform. By dynamically predicting and allocating priorities based on the arrival time of multiple agents in the interaction area, and adopting a rolling time-domain distributed collaborative optimization strategy, it realizes coordinated control and conflict avoidance of multiple workshops, thereby improving the stability and traffic efficiency of continuous interaction of multiple vehicles in closed field scenarios.

[0115] Based on the real-time location, speed, and reference path information of multiple agents, the cloud control platform predicts the arrival time of each vehicle at the center of the interaction area. Priorities are assigned to each agent based on their arrival time. Vehicles arriving earlier have higher priority, thus determining the interaction order. After the priority is determined, the cloud control platform establishes a priority-based scheduling matrix. This matrix, in conjunction with the aforementioned game optimization output, serves as the upper-level constraint input for multi-vehicle local programming.

[0116] Step 135: During the execution phase of the multi-agent system, the cloud control platform employs a rolling time-domain distributed collaborative optimization architecture to achieve coordinated control among the agents. When the cloud control platform detects multiple high-priority vehicles approaching simultaneously within the interaction area, it automatically divides the area into sub-regions and implements block optimization, keeping the optimization problem size under control, thereby reducing computational load and ensuring real-time performance. A soft constraint mechanism is introduced during the optimization process. When a low-priority vehicle cannot meet the avoidance constraint in the current prediction time domain, the system dynamically adjusts its target trajectory through a penalty term, causing it to automatically yield or slow down in the next time domain.

[0117] Step 136: After priority scheduling and distributed collaborative optimization, the cloud control platform outputs the set of optimal state trajectories for all agents. With control sequence The system then sends the data to the site execution module in real time during the next control cycle. Simultaneously, the platform monitors the interaction status and trajectory consistency of each vehicle. Upon detecting local congestion or execution deviation, it immediately initiates a local re-optimization process to restore system coordination.

[0118] Compared to conventional closed-field testing, this embodiment improves the adversarial intensity of closed-field testing by 66.7% and the scenario diversity by 56.7% for tested conditions such as left turns at unsignalized intersections, based on the above testing methods and through cloud collaborative control and dynamic interaction of multiple intelligent agents. This provides a risk assessment environment for autonomous driving systems that is closer to real traffic conditions and is suitable for safety certification, algorithm iteration, and edge scenario testing of intelligent connected vehicles.

[0119] Those skilled in the art should understand that variations can be implemented by combining existing technology with the above embodiments, which will not be elaborated here. Such variations do not affect the essence of the present invention, and will not be elaborated here either.

[0120] The preferred embodiments of the present invention have been described above. It should be understood that the present invention is not limited to the specific embodiments described above. Systems and structures not described in detail should be understood as being implemented in a conventional manner in the art. Any person skilled in the art can make many possible variations and modifications to the technical solutions of the present invention using the methods and techniques disclosed above, or modify them into equivalent embodiments with equivalent changes, without departing from the scope of the present invention. This does not affect the essential content of the present invention. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the content of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-agent closed-field testing method based on cloud platform collaborative control, characterized in that, include: Step S11: Construct a cloud collaborative testing system to enable information interaction between the cloud platform and multiple intelligent agents; the cloud platform includes a data middleware, a time calibration server, and a cloud control platform. Step S12: Establish a multi-agent vehicle model and behavior model: The intelligent agent vehicle model includes a vehicle kinematics model and a vehicle path and local trajectory generation model; the intelligent agent vehicle behavior model includes two states: free-flow mode and adversarial mode. The state is dynamically switched by the cloud control platform based on the spatial relationship between the tested vehicle and the interaction area; in the adversarial mode, the intelligent agent interacts with the tested vehicle in a game-like manner through instructions issued by the cloud control platform. Its control objective is to create a controlled risk situation. The control logic includes adversarial triggering and risk assessment, acceleration and steering coordinated control, safety constraints and action boundaries; the cloud control platform detects the relative speed and distance between intelligent agents in real time. When speed conflicts or path overlaps occur, the vehicle priority is adjusted through speed allocation and time delay scheduling to avoid local congestion; the cloud control platform realizes real-time coordination and long-term self-optimization of the multi-agent behavior mode through distributed scheduling and feedback learning mechanisms. Step S13, Real-time adversarial game and collaborative optimization during testing: The real-time adversarial game and collaborative optimization algorithm adopts the Stackelberg master-slave game model and the mixed integer nonlinear programming joint optimization framework to realize real-time adversarial interaction and collaborative control between multiple agents and the tested vehicle; Multi-agent priority collaboration is a distributed scheduling framework based on the cloud control platform. By dynamically predicting and allocating priorities based on the arrival time of multiple agents in the interaction area, and adopting a rolling time-domain distributed collaborative optimization strategy, it realizes coordinated control and conflict avoidance of multiple workshops. Step S14: After the test is completed, perform a scenario evaluation based on historical test data.

2. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, The multi-agent system consists of no fewer than three agents; the agents include dummy humans or dummy vehicles.

3. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, In step S12, the vehicle path generation uses a multi-stage A* algorithm based on a high-precision map to generate a global reference route, and smooths it using a cubic Bezier curve; the local trajectory generation uses the Frenet coordinate system, projects the vehicle state as longitudinal arc length and lateral offset, and applies no reversing constraint, speed and acceleration constraint, and lateral deviation constraint.

4. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, In step S12, when the Euclidean distance from the tested vehicle to the interaction area is not greater than a specific distance, the agent is triggered to enter the adversarial mode; when the Euclidean distance from the tested vehicle to the interaction area is greater than a specific distance, the agent automatically switches to the free flow mode.

5. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, In step S12, the adversarial triggering and risk assessment control uses the post-crossing time as a risk indicator. The cloud control platform dynamically allocates the conflict targets of each intelligent agent so that the adversarial behaviors of different intelligent agents in the interaction area do not overlap or interfere with each other. In acceleration and steering coordinated control, the agent calculates the control input based on a game-theoretic strategy. ,in It is longitudinal acceleration; For the front wheel steering angle, the control law is as follows: in This represents the baseline longitudinal acceleration value for the agent to maintain normal driving conditions without adversarial behavior. The cloud control platform assigns the expected interaction risk level to the intelligent agent. The actual interaction risk level between the agent and the VUT at the current moment. The target vehicle speed that the intelligent agent hopes to achieve under the current adversarial conditions. The longitudinal velocity value of the intelligent vehicle at time t. To control the gain coefficient; By coordinating longitudinal acceleration and lateral steering, the intelligent agent actively performs high-risk actions, including approaching, overtaking, changing lanes, or cutting in, forming a continuous interactive scenario.

6. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, In step S12, the distributed scheduling and feedback learning mechanism first constructs a mode switching matrix based on the number of agents N and the test period T. ,in This indicates that the i-th agent is at time step [i]. The cloud control platform dynamically updates the control mode and priority of each intelligent agent based on the matrix; the matrix is ​​updated in real time in each rolling time domain cycle, enabling the cloud control platform to dynamically adjust the mode, task priority and control strategy of each vehicle.

7. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, Step S13 includes: Step S131: Master-slave game modeling and system initialization; Step S132: In the game optimization layer, the intelligent agent, as the leader, and the tested vehicle, as the follower, each use their own utility function as the optimization objective. Step S133: Use the Stackelberg master-slave game framework and introduce Karush-Kuhn-Tucker conditions to achieve nested solution; during the solution process, apply collision avoidance constraints, no-reverse constraints, and deadlock penalty constraints simultaneously. Step S134: The cloud control platform adopts a rolling time-domain collaborative control mechanism. In each time step, the initial state is updated according to the solution results of the previous cycle to predict the trajectory for several future steps and adjust the control commands in real time. The cloud control platform synchronizes the state information of each intelligent agent and the vehicle under test through the V2X network. If the model prediction error is detected to exceed the threshold, it will automatically trigger re-optimization.

8. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 7, characterized in that, In step S13, the multi-agent priority cooperative control includes: The cloud control platform first predicts the arrival time of each vehicle at the center of the interaction area based on the real-time location, speed, and reference path information of multiple agents; then, it assigns priorities to each agent according to the order of arrival time. Vehicles arriving earlier have higher priority, thus determining the interaction order; after the priority is determined, the cloud control platform establishes a priority-based scheduling matrix. This matrix, in conjunction with the aforementioned game optimization output, serves as the upper-level constraint input for multi-vehicle local programming. During the execution phase of the multi-agent system, the cloud control platform adopts a distributed collaborative optimization architecture with a rolling time domain to achieve coordinated control among the agents. When the cloud control platform detects that multiple high-priority vehicles are approaching simultaneously in the interaction area, it will automatically divide the area into sub-regions and implement block optimization to keep the scale of the optimization problem under control. A soft constraint mechanism is introduced during the optimization process. When a low-priority vehicle is unable to meet the avoidance constraint in the current prediction time domain, its target trajectory is dynamically adjusted through a penalty term so that it can automatically give way or slow down in the next time domain. After priority scheduling and distributed collaborative optimization, the cloud control platform outputs the optimal state trajectory set and control sequence of all intelligent agents, and sends them to the site execution module in real time in the next control cycle. The cloud control platform also monitors the interaction status and driving trajectory consistency of each vehicle. When a local blockage or execution deviation is detected, the local re-optimization process is immediately initiated to restore system coordination.

9. The multi-agent closed-field testing method based on cloud platform collaborative control according to claim 1, characterized in that, In step S14, after the test is completed, the test vehicle data, multi-agent data, and control data are stored in the database and analyzed based on historical data and vehicle-mounted sensing device data.

10. A multi-agent closed-site testing system based on cloud platform collaborative control, characterized in that, A method for implementing the multi-agent closed-field testing method as described in any one of claims 1 to 9; comprising: The cloud-based collaborative testing system enables information interaction between the cloud platform and multiple intelligent agents; the cloud platform includes a data middleware, a time calibration server, and a cloud control platform. The site execution module includes the vehicle under test and a multi-agent system based on the intelligent agent vehicle model and behavior patterns as described in any one of claims 1 to 9; The cloud coordination framework system ensures real-time adversarial game and collaborative optimization as described in any one of claims 1 to 9 during testing.

Images