Intelligent vehicle elastic planning and control method based on large language model guidance under network attack

Abstract

The invention discloses an intelligent vehicle elastic planning and control method based on large language model guidance under network attack, and belongs to the field of vehicle information safety and control, and the method specifically comprises the steps: firstly, building an intelligent vehicle auxiliary decision making and control system, and installing the system on a to-be-detected vehicle; the automatic driving decision-making unit generates an expected driving intention of the current to-be-detected period; the path planning unit generates a corresponding reference trajectory sequence. A vehicle reachable set estimator calculates a maximum allowable motion reachable range. The scene text editor completes scene description of the driving environment in combination with the perception information and converts the scene description into a structured text; the large language model performs comprehensive judgment on an expected driving intention and a planned reference trajectory sequence based on a maximum allowable motion reachable range and structured text description, and gives a final auxiliary decision variable. And based on the reachable set estimation value, constructing a vehicle motion controller based on data driving, and obtaining a control instruction to complete execution. According to the invention, the adaptability and robustness of the real vehicle operation environment are improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent vehicle information security and control, specifically a method for flexible planning and control of intelligent vehicles based on a large language model under network attacks. Background Technology

[0002] With the rapid development of intelligent connected vehicle technology, autonomous driving is gradually becoming more intelligent and networked. In a typical autonomous driving system, perception, decision-making, planning, and control modules work together to enable vehicles to drive autonomously and safely in complex traffic and network environments.

[0003] In recent years, the collaboration between the vehicle and the cloud has become an important development direction in the architecture of autonomous driving. Among them, OTA (Over-The-Air) remote upgrade technology, as a key means to update vehicle control strategies, decision logic, path planning parameters and model weights, has been widely integrated into various autonomous driving platforms.

[0004] Through the OTA (Over-The-Air) mechanism, vehicle manufacturers or autonomous driving technology cloud service platforms can dynamically update the software of autonomous driving systems throughout the vehicle's lifecycle, improving system performance and safety. However, the network dependence of OTA also introduces potential information security threats, especially in vehicle decision-making and path planning. If the OTA information transmission channel of an intelligent vehicle is compromised by malicious attackers, backdoors may be implanted into the control strategy, causing the vehicle's safety decision-making logic and path planning to be tampered with, or even generating trajectories that intentionally cause collisions, thereby seriously threatening the lives and property of occupants and surrounding road users.

[0005] To meet the increasingly stringent cybersecurity requirements of intelligent vehicles, publication number CN118540139B proposes a signal-level-based cloud-controlled intelligent connected vehicle lateral control system network attack detection method. This method utilizes collected raw security data and innovative attack data to detect abnormal states and identify attacks. Publication number CN119584134A integrates information from multiple sensors such as GPS, LiDAR, and IMU to achieve real-time and accurate estimation of vehicle status and effectively eliminate malicious data introduced by network attacks, thereby improving the safety and reliability of autonomous vehicles.

[0006] Existing cybersecurity protection strategies mainly focus on traditional intrusion detection, communication encryption, and data redundancy, which are relatively weak in identifying and defending against deep logic attacks at the "intelligent layer" (such as decision forgery and intent misdirection). In addition, facing complex and ever-changing road environments and potential attack patterns, traditional rule-based control systems lack real-time adaptability and interpretability, making it difficult to quickly reconstruct safe and feasible new decisions and movement trajectories after an attack occurs.

[0007] In recent years, Large Language Models (LLMs) have demonstrated powerful generalization and reasoning capabilities in natural language understanding, knowledge reasoning, and multimodal fusion. In the field of autonomous driving, LLMs have the potential to assist in identifying abnormal decisions, analyzing conflicting driving intentions, and generating reasonable behavioral suggestions. Therefore, introducing a large language model-guided mechanism, combined with vehicle motion reachability set theory, to achieve resilient planning and control under network attacks could potentially improve the system's robustness and security in the face of unknown attacks and complex scenarios. Summary of the Invention

[0008] To address the security risks that may arise in the decision-making and planning modules during the OTA upgrade process of autonomous driving systems, this invention proposes an intelligent vehicle elastic planning and control method based on a large language model under network attacks. This method integrates a large language model-assisted judgment mechanism with an elastic control reconstruction strategy to improve the vehicle's perception and judgment capabilities and motion control recovery capabilities when attacked.

[0009] The specific steps of the intelligent vehicle elastic planning and control method based on a large language model under network attacks are as follows:

[0010] Step 1: Build an intelligent vehicle auxiliary decision-making and control system and install it on the vehicle under test, S;

[0011] The system includes an autonomous driving decision-making unit, a path planning unit, a vehicle reachability set estimator, a scene text encoder, a large language model judgment module, a path replanning and motion control module, and an optional actuator interface.

[0012] Among them, the autonomous driving decision unit generates the expected driving intention for the current period based on global task planning and perception information, and transmits it to the large language model judgment module.

[0013] The path planning unit generates a set of reference trajectory points in several time domains for the vehicle to be tested in the future, and transmits them to the large language model judgment module.

[0014] The vehicle reachability set estimator uses the displacement input and output of the previous t time steps to predict the vehicle displacement output at time t+1, and combines this with a matrix polyhedron data structure to generate... The estimated value of the vehicle reachability set at any given time, i.e. the maximum permissible range of movement of the vehicle, is transmitted to the large language model judgment module.

[0015] The scene text encoder converts the driving environment scene of the vehicle under test S into a structured text description and transmits it to the large language model judgment module.

[0016] The large language model judgment module integrates the maximum permissible range of motion and structured text description to make a consistent judgment on the feasibility, safety and rule compliance of the current expected driving intention and the reference trajectory, and outputs auxiliary decision values ​​to the path replanning and motion control module.

[0017] The path replanning and motion control module selects to execute the reference trajectory, local replanning, or replacement trajectory intention based on the auxiliary decision value, and generates corresponding chassis control commands to send to the actuator interface.

[0018] An optional actuator interface is used to send control commands to steering, braking and drive actuators and collect feedback to achieve a control closed loop.

[0019] Step 2: The autonomous driving decision unit generates the expected driving intention for the current test period based on the global mapping information, high-precision map and cloud instructions of the intelligent transportation system, combined with the road traffic conditions obtained by the vehicle perception.

[0020] The stated intentions include maintaining the current lane, changing lanes to the left, changing lanes to the right, accelerating, decelerating, and stopping / waiting;

[0021] Step 3: Based on the desired driving intention, the path planning unit integrates multi-source sensor data, vehicle dynamic status, and map location information, and calls the path planning algorithm to generate the corresponding reference trajectory sequence. .

[0022] Step 4: The vehicle reachability set estimator calculates the maximum permissible range of motion based on the execution input capability of the vehicle S under test.

[0023] The specific calculation is as follows:

[0024] First, the collection time is input sequence and output data sequence And construct the following signal vector:

[0025]

[0026] in Given the input sequence, For the output sequence, To output the right-shifted sequence.

[0027] Then, the vehicle displacement is calculated using vector signals to output the prediction result. And obtain the matrix polyhedron data structure. :

[0028]

[0029] symbol Represents the pseudo-inverse of a matrix;

[0030]

[0031] Matrix polyhedral data structure It contains a combination of the intelligent vehicle dynamics system matrix and the input matrix. ;

[0032] Finally, construct a system for handling disturbances. perturbation set , matrix polyhedral data structure Substitute, obtain Vehicles available at any time The estimated value;

[0033] The expression is:

[0034]

[0035] Among them, polyhedron For the multi-faceted constraints of the actuator input; initial time... Man-made regulations.

[0036] Step 5: Using a scene text editor and combining sensory information, convert the surrounding vehicle status and traffic environment into a structured text description in a fixed text encoding format to complete the scene description of the driving environment.

[0037] Step Six: Based on the maximum permissible range of motion of the vehicle under test S under the input capability, and the structured text description, use a large language model to analyze the expected driving intention and the reference trajectory sequence of the planned route. Make a comprehensive judgment and provide the final auxiliary decision variables;

[0038] Auxiliary decision variables include the combined motion sequence;

[0039] Step 7: Based on the different auxiliary decision variable values, convert them into the steering angle input that the current vehicle S chassis can execute.

[0040] Specifically, by introducing a quadratic motion planning module, the original planned reference trajectory sequence is processed. Perform dynamic corrections and mark the new lane-change trajectory as... .

[0041] Step 8: Estimated value based on reachability set Construct a data-driven vehicle motion controller to obtain control commands. :

[0042]

[0043] Among them, the function The cost function designed for the controller, These are the upper and lower bounds of the slack variable, respectively. For the present Always looking towards the future The predicted value within a given time period. Input The set of constraints that need to be satisfied by the input The constraint set can be defined by humans or be the actual constraints imposed on the system.

[0044] Step 9: Send control commands The current vehicle's S-chassis execution system takes the driving intention as its reference and executes the control commands through the actuators.

[0045] The present invention has the following advantages:

[0046] 1. This invention addresses the security risks that may arise in the decision-making and planning modules during the OTA upgrade process in autonomous driving systems, and designs a safety control technology to prevent malicious decision-making and unreasonable motion planning.

[0047] 2. This invention utilizes a large language model to judge the rationality of existing decisions and planning trajectories for autonomous driving, resulting in better interactivity and traffic scenario understanding capabilities;

[0048] 3. In this invention, the local path planning and motion control unit do not rely on an accurate vehicle dynamics model. Based on data-driven operation, it can effectively improve the adaptability and robustness to the actual vehicle operating environment. Attached Figure Description

[0049] Figure 1 This is a flowchart of the intelligent vehicle elastic planning and control method based on a large language model under network attacks according to the present invention;

[0050] Figure 2 This is a schematic diagram of the intelligent vehicle auxiliary decision-making and control system built according to the present invention; Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention are within the scope of protection of the present invention.

[0052] The aforementioned method for resilient planning and control of intelligent vehicles based on large language models under network attacks, such as... Figure 1 As shown, the specific steps are as follows:

[0053] Step 1: Build an intelligent vehicle auxiliary decision-making and control system and install it on the vehicle under test, S;

[0054] like Figure 2 As shown, the system includes an autonomous driving decision-making unit, a path planning unit, a vehicle reachability set estimator, a scene text encoder, a large language model judgment module, a path replanning and motion control module, and an optional actuator interface module.

[0055] Among them, the autonomous driving decision unit, as module 1, generates the expected driving intention for the current period based on global task planning and perception information; this module is the original decision module of the autonomous driving system, and its main purpose is to provide intelligent vehicle driving decisions.

[0056] The path planning unit is vehicle path planning module 2, which receives the driving decision signals and global mapping information from module 1. Based on the desired driving intention, combined with the high-precision map and the current state of the vehicle, it executes an intelligent planning algorithm to generate a set of reference trajectory points to be tracked in several future time domains.

[0057] Modules 1 and 2 are the vehicle's existing decision-making and planning control units, which often require dynamic updates to algorithms and training databases, making them highly vulnerable to hacker attacks during OTA (Over-The-Air) updates.

[0058] The vehicle reachability set estimator, as module 4, calculates the maximum permissible range of motion of the vehicle as the reachability set boundary information under given input capabilities and constraints. The scene text encoder, as module 5, transforms perception information, map information, expected driving intention, reference trajectory point set and reachability set boundary information into structured text descriptions so that the large language model can understand and reason.

[0059] By combining the text encoder in Module 4 and Module 5, vehicle and traffic perception information is fused together. The reachability set information is then fused in the form of a text paradigm and input into the large language model to determine whether the existing decision and planning trajectories are reasonable, and to generate different auxiliary decision variable values.

[0060] The large language model judgment module, as module 3, is an auxiliary judgment unit for the large language model. This module relies on the large language intelligent agent deployed in the cloud, which integrates the boundary information of the reachable set and the structured text description to make a consistent judgment on the feasibility, safety and rule compliance of the current expected driving intention and the reference trajectory, and outputs auxiliary decision values.

[0061] Module 6 is the path replanning and motion control module. Based on the auxiliary decision value, it selects to execute the original trajectory, local replanning, or replacement intention, and generates corresponding chassis control commands. This unit is mainly responsible for modifying the original reference trajectory and generating the final vehicle chassis control execution input. Considering that it is difficult to obtain the actual vehicle dynamics model, this part of the algorithm is designed based on real vehicle data to reduce the dependence on the accurate dynamics model.

[0062] An optional actuator interface module is used to send control commands to steering, braking and drive actuators and collect feedback to achieve a closed-loop control system.

[0063] Step 2: The autonomous driving decision unit generates the expected driving intention for the current test period based on the global mapping information, high-precision map and cloud instructions of the intelligent transportation system, combined with the road traffic conditions obtained by the vehicle perception.

[0064] The stated intentions include maintaining the current lane, changing lanes to the left, changing lanes to the right, accelerating, decelerating, and stopping / waiting;

[0065] Step 3: Based on the desired driving intention, the path planning unit integrates multi-source sensor data, vehicle dynamic status, and map location information, and calls the path planning algorithm to generate corresponding reference trajectory horizontal and vertical position point sets, forming a sequence. .

[0066] Among them, the longitudinal displacement satisfies The lateral lane-changing trajectory uses a fifth-order polynomial. Furthermore, the coefficients are solved by the start and end boundary conditions, and the steering angle reference is... The calculations show that the trajectory satisfies vehicle dynamics and reachability set constraints.

[0067] Step 4: The vehicle reachability set estimator calculates the maximum permissible range of motion based on the execution input capability of the vehicle S under test, and uses it as the prior physical knowledge input in the large language model.

[0068] The specific calculation is as follows:

[0069] First, in order to obtain vehicle reachability information, the collection time is [duration missing]. input sequence and output data sequence And construct the following signal vector:

[0070]

[0071] in Given the input sequence, For the output sequence, To output the right-shifted sequence.

[0072] Then, the vehicle displacement is calculated using vector signals to output the prediction result. And obtain the matrix polyhedron data structure. :

[0073]

[0074] symbol This represents the pseudo-inverse of the matrix; as can be seen, the expression obtains the system expression in the form of data stacking, which means that the combination of the system matrix and the input matrix at the current moment... Implied in matrix polyhedra middle.

[0075]

[0076] Matrix polyhedral data structure It contains a combination of the intelligent vehicle dynamics system matrix and the input matrix. ;

[0077] Finally, construct a system for handling disturbances. perturbation set , matrix polyhedral data structure Substitute, obtain Vehicles available at any time The estimated value;

[0078] The expression is:

[0079]

[0080] Among them, polyhedron For the multi-faceted constraints input to the actuator. Initial time... Man-made regulations.

[0081] Step 5: Using a scene text editor and combining sensory information, convert the surrounding vehicle status and traffic environment into a structured text description in a fixed text encoding format to complete the scene description of the driving environment.

[0082] Step Six: Based on the maximum permissible range of motion of the vehicle under test S under the input capability, and the structured text description, use a large language model to analyze the expected driving intention and the reference trajectory sequence of the planned route. Make a comprehensive judgment and provide the final auxiliary decision variables;

[0083] Auxiliary decision variables include the combined motion sequence;

[0084] Step 7: Based on the different auxiliary decision variable values, convert them into the steering angle input that the current vehicle S chassis can execute.

[0085] Specifically, to address unreasonable reference trajectories, a secondary motion planning module is introduced to modify the original planned reference trajectory sequence. Perform dynamic corrections and mark the new lane-change trajectory as... .

[0086] Step 8: Estimated value based on reachability set Construct a data-driven vehicle motion controller to obtain control commands. :

[0087]

[0088] Among them, the function The cost function designed for the controller, These are the upper and lower bounds of the slack variable, respectively. For the present Always looking towards the future The predicted value within a given time period. Input The set of input constraints needs to be satisfied. The constraint set can be defined by humans or be the actual constraints imposed on the system.

[0089] Step 9: Send control commands The current vehicle's S-chassis execution system takes the driving intention as its reference and executes the control commands through the actuators.

[0090] Example:

[0091] like Figure 2 As shown, this embodiment illustrates the application of this method to a terminal. It is understood that the vehicle flexible motion planning and control method provided in this embodiment can be applied not only to the field of information security control of vehicles in complex network environments, but also to other fields. The method includes the following steps:

[0092] Step 1: In Module 1, the autonomous driving decision-making unit generates the desired driving intention based on the global mapping information and cloud-based instructions from the intelligent transportation system.

[0093] Step 2: In Module 2, the path planning unit uses the on-board path planning intelligent algorithm unit to generate lane change reference trajectory points based on sensor information, the vehicle's current status, and map location information.

[0094] Step 3, Module 3, is the auxiliary decision-making unit based on a large language model. Considering that the vehicle's original intelligent driving system is vulnerable to malicious attacks in complex network environments, this module relies on the scene text encoder and perception information in Module 5 to describe the driving environment scene using a fixed text encoding format, representing the surrounding vehicle status and traffic environment. Simultaneously, considering that malicious attackers might lurk at the decision-making end and maliciously modify the planned reference trajectory at the planning end, turning it into a completely unexecutable trajectory, the large language model judgment module utilizes the vehicle reachability set estimator output in Module 4 to obtain the maximum permissible range of motion based on the current vehicle's execution input capabilities, using this information to assist the large model in its judgment.

[0095] Next, to describe the traffic environment in which the vehicle is currently operating, the text encoder 5 completes a description of the current traffic scene in the form of a text segment. This scene description can be input as exemplified below. Simultaneously, vehicle behavior is encoded as auxiliary decision variable values ​​in the large language model; for example, changing lanes to the left is encoded as 1, maintaining the current lane and decelerating is encoded as 2, and so on.

[0096] Demonstration of traffic scenarios

[0097] ## Scene Description

[0098] You are driving on the road... There are x vehicles around you, of which vehicle 1 is traveling on line..., at position (...,...) meters and speed (...,...) meters / second...

[0099] ## risk assessment

[0100] Risk quantification value of surrounding vehicle 1: xxxx

[0101] Risk quantification value of surrounding vehicle 2: xxxx

[0102] ## Previous Decisions

[0103] Previous decision: #### x

[0104] ## Planned Driving Route Information

[0105] The planned trajectory points of the vehicle in the next x seconds:

[0106] - t=1s: Position: (..,…), Velocity: (…,…) meters per second

[0107] - t=2s: Position: (…, …), Velocity: (…, …) meters per second

[0108] - …

[0109] ## Reachable Set Information

[0110] Information on the reachable set in the next x seconds:

[0111] - t=1s: Reachable set: [...] to [...]

[0112] - t=2s: Reachable set: [...] to [...]

[0113] …

[0114] ## Additional Notes

[0115] When making a decision, in addition to considering previous decisions, all vehicles traveling in the current lane and the target lane should also be considered... If your decision is the same as the previous decision, consider the planned route and reachability set information... and give the final auxiliary decision value.

[0116] Step 4: Based on the input information from modules 4 and 5, and relying on the expert world knowledge of the large language model, comprehensively judge the rationality of the original driving decisions and planned trajectories, and give the final auxiliary decision value.

[0117] Step 5: After the large model completes the comprehensive judgment on the rationality of the original driving decision and the planned trajectory and generates the final auxiliary decision value, according to different auxiliary decision values, the present invention needs to convert them into the final steering angle input that can be executed by the car chassis.

[0118] Step 6: Based on the above data-driven reachability set estimate Construct a motion controller;

[0119] Step 7: Send control commands The vehicle chassis actuator system executes the aforementioned control commands through a sophisticated internal control algorithm.

[0120] In summary, this invention targets malicious attacks on autonomous driving decision-making and planning systems. It constructs a motion reachability set estimator to obtain the permissible movement range of the vehicle in the future. Using a scene text editor, a description of the current traffic scene is constructed in the form of an input text paradigm and input into a large language model. Subsequently, the large language model integrates the traffic scene, the vehicle's motion reachability set range, and the decision-making and planning results received by the system to comprehensively judge the rationality of the existing planning and control results and provide auxiliary decision-making. Based on the auxiliary decision-making variables, a data-driven local motion planning and control module is used to achieve local adjustments to the reference trajectory and final execution.

[0121] This invention primarily addresses the security planning and control tasks of vehicles under malicious cyberattacks. The system mainly includes the following functional modules: Module 1 (Autonomous Driving Decision Unit) and Module 2 (Path Motion Planning Module) are followed by Module 3 (Large Language Model Judgment Module). This module considers the current traffic scenario in Module 5 and the estimated vehicle mobility set information obtained in Module 4, integrates existing decision and planning results, utilizes world knowledge and a human driving expert database to comprehensively judge the rationality of the driving decisions and paths generated by the autonomous driving system, and provides auxiliary decision variables. Module 6 is the Path Replanning and Motion Control Module. Guided by the auxiliary decision variables, local path replanning and motion control algorithms are used to modify and execute unreasonable reference trajectories.

[0122] This invention effectively enhances the information security of autonomous driving systems by comprehensively judging the rationality of generated driving decisions and planned trajectories through expert knowledge from a large language model. It also achieves redundant backup of the autonomous driving system. Furthermore, this invention considers the time sensitivity of the intelligent vehicle's decision-making and control units, fully leveraging the advantages of data-driven motion control algorithms in complex environments and their high computational speed. The collaboration between data-driven path replanning and motion control units and the knowledge-driven large language model decision-making unit effectively improves the feasibility and adaptability of this invention in real-world vehicle applications.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for intelligent vehicle resilience planning and control based on large language model guidance under network attack, characterized in that, The specific steps are as follows: Step one, build an intelligent vehicle auxiliary decision and control system and install it on the vehicle S to be tested; The system includes an automatic driving decision unit, a path planning unit, a vehicle reachable set estimator, a scene text encoder, a large language model judgment module, a path re-planning and motion control module, and an optional actuator interface; Step two, the automatic driving decision unit generates the expected driving intention for the current test period according to the global mapping information, high-precision map and cloud instructions of intelligent transportation system, combined with the road traffic state obtained by vehicle-mounted perception; Step three, the path planning unit fuses multi-source sensor data, vehicle dynamic state and map location information based on the expected driving intention, calls the path planning algorithm to generate the corresponding reference trajectory sequence ; Step four, the vehicle reachable set estimator calculates the maximum allowable motion reachable range based on the execution input capability of the vehicle S to be tested; The specific calculation is: First, an input sequence and an output data sequence of a length of are collected, and a signal vector is constructed as follows: wherein is an input sequence, is an output sequence, is an output right-shifted sequence; Then, the vehicle displacement output prediction result is calculated using the vector signal and the matrix polyhedron data structure is obtained : Notation denotes the pseudo-inverse of a matrix; Matrix polyhedron data structure comprises a combination of a matrix of a dynamics system of an intelligent vehicle and an input matrix ; Finally, construct a system for handling disturbances. perturbation set , matrix polyhedral data structure Substitute, obtain Vehicles available at any time The estimated value; The expression is: wherein the polyhedron is a polyhedral surface constraint for actuator inputs; Step five, use the scene text editor to convert the vehicle state and traffic environment into structured text description in a fixed text encoding form, complete the scene description of the driving environment combined with the perception information; Step six, based on the vehicle S to be tested, the maximum allowable motion reachable range under the input capacity, and the structured text description, use a large language model to obtain the expected driving intention and the planned reference trajectory sequence Make a comprehensive judgment and give the final auxiliary decision variable; Step seven, according to different auxiliary decision variable values, it is converted into the steering angle input executable by the chassis of the current vehicle S; Step eight, estimating based on reachable set , constructing a data-driven vehicle motion controller to obtain control instructions : where the function is a cost function designed for the controller, are the upper and lower bounds of the slack variable, respectively, is the predicted value at the future time instant for the current time instant; Step nine, sending control instructions The control instructions are executed by actuators in the vehicle S chassis execution system as a driving intent.

2. The method of claim 1, wherein, In step one, the automatic driving decision unit generates the expected driving intention for the current period based on global task planning and perception information, and transmits it to the large language model judgment module; The path planning unit generates a set of reference trajectory points for the future vehicle to be tested in several time domains, and transmits it to the large language model judgment module; The vehicle reachable set estimator uses the displacement input and output of the previous t time points to predict the vehicle displacement output at t+1 time point, and combines the matrix polyhedron data structure to generate the vehicle reachable set estimation value at t+1 time point, that is, the maximum allowable motion reachable range of the vehicle, which is transmitted to the large language model judgment module; The scene text encoder converts the driving environment scene of the vehicle S to be tested into a structured text description and transmits it to the large language model judgment module; The large language model judgment module integrates the maximum allowable motion reachable range and the structured text description to judge the consistency of the current expected driving intention and the reference trajectory in terms of feasibility, safety and rule compliance, and outputs the auxiliary decision value to the path re-planning and motion control module; The path re-planning and motion control module selects to execute the reference trajectory or local re-planning or replaces the trajectory intention according to the auxiliary decision value, and generates the corresponding chassis control instructions and sends them to the optional actuator interface; The optional actuator interface is used to issue control instructions to steering, braking and driving actuators and collect feedback to realize control loop.

3. The method of claim 1, wherein, The expected driving intention in step two includes keeping the current lane, changing lanes to the left, changing lanes to the right, accelerating, decelerating and parking waiting.

4. The method of claim 1, wherein, The step seven is specifically: through introducing a secondary motion planning module, dynamically correcting the original planned reference trajectory sequence, marking the new lane-changing trajectory as . .

Images