An automatic driving risk early warning and closed-loop decision method, system, device and storage medium based on large model cooperation

Abstract

The application provides an automatic driving risk early warning and closed-loop decision method based on large model cooperation, and discloses a corresponding system, equipment and storage medium. The method fuses scene perception, vehicle state and control task information, completes scene understanding and strategy generation by a main decision large model, implements rule constraint and execution arrangement through a logic control module, and then combines vehicle feedback to perform safety checking and abnormality identification by a risk evaluation large model. When a risk is detected, an alarm large model generates early warning information and feeds back the main decision model to complete decision correction. The method constructs a "decision-control-evaluation-feedback" closed-loop architecture, and improves the risk perception, decision correction and safety control capability of the automatic driving system in a complex dynamic scene.

Description

Technical Field

[0001] This invention relates to the fields of autonomous driving and intelligent transportation technology, and in particular to a method, system, device and storage medium for autonomous driving risk warning and closed-loop decision-making based on large model collaboration. Background Technology

[0002] With the rapid development of autonomous driving technology, vehicle-to-everything (V2X) technology, and intelligent vehicle control technology, autonomous driving systems are gradually evolving from a single perception or decision-making module-driven approach to an intelligent decision-making approach characterized by multi-module collaboration, real-time closed-loop feedback, and high safety redundancy. In autonomous driving scenarios, vehicles need to continuously process multi-source heterogeneous information from cameras, LiDAR, millimeter-wave radar, positioning units, and the vehicle's bus system, and combine this information with the current driving task, traffic rules, and environmental conditions to perform scene understanding, behavioral decisions, control execution, and risk avoidance in real time. Therefore, achieving highly reliable, safe, and interpretable closed-loop decision-making for autonomous driving in complex dynamic scenarios has become a key issue in the research and engineering implementation of autonomous driving systems.

[0003] Most existing autonomous driving solutions adopt a sequential technical architecture of "perception-prediction-planning-control". While this architecture is structurally clear, it still has several shortcomings in practical applications. On the one hand, different modules typically use independent models for processing, and information transmission between modules mainly relies on fixed interfaces and predefined features, lacking higher-level semantic collaboration and global reasoning capabilities. This leads to problems such as information fragmentation, insufficient contextual understanding, and inconsistent local decisions in complex scenarios. On the other hand, traditional decision-making and control methods often rely on rule logic, supervised learning models, or single planning algorithms. Their comprehensive analysis capabilities for long-term environmental changes, sudden events, and implicit risk factors are limited, making it difficult to balance flexibility and safety.

[0004] In recent years, large language models and multimodal large models have demonstrated strong capabilities in complex task understanding, contextual reasoning, and decision support, providing new technical pathways for scene understanding, behavioral decision-making, and risk analysis in autonomous driving systems. By introducing large models, unified modeling of environmental information, vehicle state, and task objectives can be achieved at a higher semantic level, thereby enhancing the system's ability to understand and reason about complex traffic scenarios. However, current technologies applying large models to autonomous driving mainly focus on scene description, question answering assistance, offline analysis, or single-round decision output, lacking a closed-loop architecture design that deeply integrates with the vehicle control chain, and have not yet formed a complete technical solution covering "decision generation—logical constraints—risk assessment—alarm feedback—decision correction".

[0005] Furthermore, existing technologies suffer from the following problems: First, single-model-driven decision-making methods lack robustness in the face of perceptual noise, rule conflicts, control mismatch, or sudden environmental changes, making it difficult to detect potential risks and make adaptive corrections in a timely manner. Second, although some existing solutions have safety verification or rule filtering mechanisms, most of them remain at the level of traditional threshold judgment or static rule verification, lacking the ability to conduct deep semantic analysis of risk sources, risk levels, and risk evolution trends. Third, the alarm functions in existing autonomous driving control links are usually isolated from the decision-making modules, and the alarm results are difficult to effectively influence the upstream decision-making process, thus failing to form a true closed-loop optimization. Fourth, existing solutions do not adequately consider consistency control, priority management, and conservative decision-making mechanisms in abnormal situations among the collaborative outputs of multiple models, affecting the overall safety and engineering availability of the system.

[0006] Therefore, there is an urgent need to provide a technology solution for autonomous driving risk warning and closed-loop decision-making based on large-scale model collaboration, so as to achieve collaborative understanding of scene perception information, vehicle status information and driving task information. By linking the main decision-making large model, the risk assessment large model and the alarm generation module, combined with logical control constraints and feedback correction mechanisms, the accuracy, safety, robustness and interpretability of autonomous driving system decision-making in complex dynamic environments can be improved. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a method, system, device, and storage medium for autonomous driving risk warning and closed-loop decision-making based on large-model collaboration.

[0008] Firstly, this invention provides a method for risk warning and closed-loop decision-making in autonomous driving based on large-model collaboration. The technical solution of this method is as follows: Acquire scene perception information, vehicle status information, and driving task information of the target vehicle; The scene perception information, the vehicle status information, and the driving task information are input into the main decision-making model to perform scene understanding and driving strategy generation, and an initial decision result is obtained. The initial decision results are input into the logic control module for rule constraint verification, process orchestration, and control command generation to obtain vehicle control commands. The vehicle control commands, the scene perception information, and the vehicle operation feedback information are input into the risk assessment model for safety verification and anomaly identification to obtain the risk assessment results. When the risk assessment result indicates the existence of a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment result and the alarm information are fed back to the main decision-making model for decision correction to obtain the corrected decision result; When the risk assessment result indicates that there is no risk event, an autonomous driving closed-loop decision output is formed based on the vehicle control command; when the risk assessment result indicates that there is a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment result and the alarm information are fed back to the main decision model for decision correction to obtain the corrected decision result; based on the corrected decision result, the vehicle control command is regenerated and an autonomous driving closed-loop decision output is formed.

[0009] The beneficial effects of the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration of the present invention are as follows: The method of this invention collaboratively models scene perception information, vehicle status information, and driving task information, utilizes a master decision-making model to achieve high-level scene understanding and strategy generation, combines a logic control module to complete rule constraints and execution control, and constructs a feedback correction mechanism through a risk assessment model and an alarm generation module. This solves the problems of limited single-model decision-making capabilities, insufficient risk warning, difficulty in reverse-influencing the decision-making link with alarm information, and insufficient closed-loop optimization capabilities in existing autonomous driving technologies, thereby improving the decision-making accuracy, safety, robustness, and interpretability of autonomous driving systems in complex dynamic environments.

[0010] Based on the above scheme, the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration of the present invention can be further improved as follows.

[0011] In one optional approach, acquiring the scene perception information, vehicle status information, and driving task information of the target vehicle includes: Data on the vehicle's surrounding environment is acquired by installing at least one of the following on the target vehicle: a camera, a lidar, or a millimeter-wave radar. The environmental data is processed by at least one of target detection, lane line recognition, traffic sign recognition, and passable area recognition to obtain the scene perception information; The vehicle's position, speed, acceleration, heading angle, and attitude information are obtained through the target vehicle's positioning unit, inertial measurement unit, and vehicle bus to obtain the vehicle's state information; The driving task information is obtained by acquiring at least one of the following: navigation path, current driving stage, driving intention, and system control objective.

[0012] The advantages of adopting the above-mentioned optional approach are as follows: by further collecting environmental data through multi-source sensors and combining the positioning unit, inertial measurement unit and vehicle bus to obtain the vehicle's operating status, and by introducing navigation path and driving task constraints, it is possible to achieve comprehensive acquisition and unified organization of autonomous driving input information, providing a reliable data foundation for subsequent large-scale model collaborative reasoning.

[0013] In one optional approach, the step of inputting the scene perception information, the vehicle state information, and the driving task information into the main decision-making model for scene understanding and driving strategy generation to obtain an initial decision result includes: The scene perception information is feature-encoded to obtain environmental representation features; The vehicle state information is state encoded to obtain motion state features; The driving task information is encoded to obtain task constraint features; The environmental representation features, motion state features, and task constraint features are fused and input into the main decision-making model; The main decision model outputs an initial decision result corresponding to the current driving scenario. The initial decision result includes at least one of the following: longitudinal control strategy, lateral control strategy, behavioral decision result, and trajectory planning result.

[0014] The beneficial effects of adopting the above-mentioned optional approach are as follows: by further encoding and fusing environmental information, vehicle status and driving tasks separately, the main decision-making model can complete scene understanding and policy reasoning from a multi-dimensional semantic context, thereby enhancing the globality, coherence and adaptability of autonomous driving decision-making.

[0015] In one alternative approach, the step of inputting the initial decision result into the logic control module for rule constraint verification, process orchestration, and control command generation to obtain vehicle control commands includes: Based on a pre-defined traffic rule library, safety constraint library, and vehicle dynamics constraints, the initial decision results are verified for compliance. When the initial decision result meets the constraints, the control flow is arranged to generate execution-level vehicle control commands. When the initial decision result does not meet the constraints, the initial decision result is logically corrected, downgraded, or its execution is blocked, and the constraint correction result is output.

[0016] The advantages of adopting the above-mentioned optional approach are as follows: by further combining the high-level decisions output by the large model with traffic rules, control logic and vehicle dynamics constraints through the logic control module, unreasonable or unexecutable decisions can be effectively avoided from being issued directly, thereby improving the compliance, security and engineering feasibility of control commands.

[0017] In one optional approach, the step of inputting the vehicle control commands, the scene perception information, and the vehicle operation feedback information into a large-scale risk assessment model for safety verification and anomaly identification to obtain risk assessment results includes: Obtain the execution feedback information of the target vehicle to the vehicle control command, wherein the execution feedback information includes at least one of control response information, trajectory deviation information and vehicle stability information; The execution feedback information, the scene perception information, and the vehicle control command are jointly input, and the risk assessment big model is used to identify at least one of the following: collision risk, boundary crossing risk, rule conflict risk, perception anomaly risk, and control mismatch risk. Output the risk assessment results, which include at least one of the following: risk type, risk level, risk location, and risk triggering cause.

[0018] The beneficial effects of adopting the above-mentioned optional methods are as follows: by further conducting joint analysis of control execution feedback, environmental status and control commands, potential anomalies and explicit risks in the autonomous driving process can be identified in a timely manner, and the sources of risks can be semantically analyzed and classified, thereby enhancing the system's risk perception capability and active protection capability.

[0019] In one optional approach, when the risk assessment result indicates the existence of a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment result and the alarm information are fed back to the main decision-making model for decision correction, resulting in a corrected decision result, including: When the risk level reaches a preset threshold, the alarm generation module is triggered to generate alarm information; The alarm information includes at least one of the following: risk description information, risk level information, suggested handling strategy, and manual takeover prompt; The risk assessment results and the alarm information are input as feedback constraint information into the master decision-making model; The main decision model is used to regenerate a revised decision result that matches the current scenario, thus forming a closed-loop decision update.

[0020] The beneficial effects of adopting the above-mentioned optional methods are as follows: by further outputting alarm information in the form of structured or natural language through the alarm generation module and feeding it back to the main decision-making model for further reasoning, the linkage between risk warning and decision correction can be realized, thereby enhancing the adaptive adjustment capability and closed-loop optimization capability of the autonomous driving system in the face of sudden risks.

[0021] In one alternative approach, it also includes: The outputs of the main decision-making model, the risk assessment model, and the alarm generation module are compared for consistency. When the consistency comparison result is less than the preset consistency threshold, at least one of the model negotiation mechanism, conservative decision-making mechanism or manual takeover mechanism is triggered. The conservative decision-making mechanism includes at least one of deceleration, stopping, lane keeping, and lane change restriction.

[0022] The beneficial effects of adopting the above-mentioned optional approach are as follows: by further verifying the consistency of multi-model outputs and the conservative decision-making processing mechanism in abnormal situations, the safety of vehicle operation can be guaranteed when there is disagreement between models or when the system uncertainty is high, and the risks caused by misjudgment, omission, and risky execution can be reduced.

[0023] Secondly, this invention provides an autonomous driving risk warning and closed-loop decision-making system based on large-scale model collaboration. The technical solution of this system is as follows: The information acquisition module is used to acquire scene perception information, vehicle status information, and driving task information of the target vehicle. The main decision module is used to input the scene perception information, the vehicle status information and the driving task information into the main decision model to perform scene understanding and driving strategy generation, and obtain the initial decision result. The logic control module is used to perform rule constraint verification, process orchestration, and control command generation on the initial decision results to obtain vehicle control commands; The risk assessment module is used to input the vehicle control commands, the scene perception information, and the vehicle operation feedback information into the risk assessment big model for safety verification and anomaly identification, and to obtain the risk assessment results. The alarm feedback module is used to output alarm information when the risk assessment result indicates the existence of a risk event, and to feed back the risk assessment result and the alarm information to the main decision module for decision correction; The closed-loop output module is used to generate closed-loop decision output for autonomous driving based on the corrected decision results.

[0024] The beneficial effects of the autonomous driving risk warning and closed-loop decision-making system based on large model collaboration of the present invention are as follows: The system of this invention achieves a complete closed-loop link from information input, strategy generation, rule verification, risk identification to feedback correction through the coordinated cooperation between the information acquisition module, main decision-making module, logic control module, risk assessment module, alarm feedback module and closed-loop output module. It solves the problems of module fragmentation, incomplete risk feedback link and insufficient safety redundancy in existing autonomous driving systems, and improves the overall coordination capability and safety assurance capability of autonomous driving systems.

[0025] Thirdly, the technical solution of an electronic device according to the present invention is as follows: It includes a memory, a processor, and a program stored in the memory and running on the processor, wherein the processor executes the program to implement the steps of the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration of the present invention.

[0026] Fourthly, the technical solution of a computer-readable storage medium provided by the present invention is as follows: A computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as described in this invention. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart of an embodiment of the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration of the present invention; Figure 2 This is a schematic diagram of the overall framework of the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration of the present invention. Figure 3 This is a schematic diagram of an embodiment of the autonomous driving risk warning and closed-loop decision-making method and system based on large model collaboration of the present invention; Figure 4 This is a schematic diagram of an embodiment of an electronic device according to the present invention. Detailed Implementation

[0028] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the disclosure of the invention will be thorough and complete.

[0029] It should be noted that, in the description of this invention, unless otherwise expressly specified and limited, the terms "installation," "connection," "coupling," "communication," "acquisition," and "output," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integrated connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the information exchange between two modules. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0030] Furthermore, it should be noted that the "autonomous driving" involved in the embodiments of this invention can be fully autonomous driving, or assisted driving, partial autonomous driving, conditional autonomous driving, highly autonomous driving, or human-machine collaborative driving scenarios. The "target vehicle" can be a vehicle platform with autonomous perception and control capabilities, such as passenger cars, commercial vehicles, logistics delivery vehicles, unmanned shuttle vehicles, mining transport vehicles, port tractors, park patrol vehicles, low-speed unmanned vehicles, or special-purpose vehicles. The "large model collaboration" refers to two or more models with strong semantic understanding, contextual reasoning, risk analysis, or content generation capabilities jointly completing decision generation, risk identification, alarm interpretation, and closed-loop correction tasks in autonomous driving under unified input information or interactive feedback constraints.

[0031] Figure 1 This diagram illustrates a flowchart of an embodiment of an autonomous driving risk warning and closed-loop decision-making method based on large-model collaboration provided by the present invention. This method can be executed by a terminal device, an in-vehicle computing platform, an edge computing node, a cloud server, or a combination of the above devices. The terminal device can be a user equipment (UE), mobile device, user terminal, in-vehicle terminal, embedded in-vehicle device, autonomous driving domain controller, industrial controller, robot controller, or other fixed or mobile electronic device. The server can be a single server or a server cluster consisting of multiple servers. Any electronic device can implement the autonomous driving risk warning and closed-loop decision-making method based on large-model collaboration described in this invention by having its processor call computer-readable instructions stored in its memory.

[0032] like Figure 1 As shown, the method includes the following steps: S1. Acquire scene perception information, vehicle status information, and driving task information of the target vehicle.

[0033] Scene perception information refers to data acquired and processed by the sensing devices mounted on the target vehicle, representing the state of the environment surrounding the vehicle. This scene perception information may include, but is not limited to: images of the road ahead, surround-view images, point cloud information, obstacle categories, obstacle locations, obstacle speeds, lane line positions, traffic sign recognition results, traffic light status, passable area information, road boundary information, static obstacle information, dynamic target trajectory information, road curvature, slope, weather conditions, lighting conditions, and road surface adhesion. For example, the vehicle's forward-facing camera may detect a slowing truck ahead, a rapidly approaching sedan in the left lane, a traffic light ahead with 5 seconds remaining, and the current road being a two-way four-lane urban arterial road.

[0034] Vehicle status information refers to data that characterizes the kinematic, dynamic, and operational states of a target vehicle. Vehicle status information may include, but is not limited to: current vehicle position, global coordinates, local coordinates, vehicle speed, longitudinal acceleration, lateral acceleration, yaw angle, pitch angle, roll angle, steering wheel angle, brake pedal opening, accelerator pedal opening, wheel speed information, gear status, control mode status, and vehicle health status. For example, the vehicle's current speed is 42 km / h, the steering wheel has a 3-degree left turn input, the vehicle is in autonomous driving following mode, and the braking system is functioning normally.

[0035] Driving task information refers to the navigation task, behavioral objective, or control objective that the target vehicle currently needs to follow or complete. This driving task information may include, but is not limited to: navigation path, destination, current path segment, lane-level navigation intent, current driving stage, lane change requirement, overtaking requirement, turning requirement, parking requirement, obstacle avoidance requirement, speed limit constraints, priority strategies, and system operation objectives. For example, the navigation path may instruct the target vehicle to change lanes to the right 150 meters ahead and then turn right onto the auxiliary road at the intersection 500 meters later.

[0036] In some implementations, the scene perception information can be acquired by at least one of a camera, lidar, millimeter-wave radar, ultrasonic radar, inertial measurement unit, global navigation satellite system module, vehicle map module, and vehicle-to-everything (V2X) equipment. Further, a perception fusion algorithm can be used to perform time synchronization, spatial alignment, and semantic fusion of multi-source sensor data to generate a unified scene representation result. The vehicle status information can be jointly obtained by a positioning unit, inertial measurement unit, chassis controller, and vehicle bus. The driving task information can be provided by a high-precision map, navigation system, task planning module, or remote scheduling platform.

[0037] The purpose of this step is to provide a unified, complete, and real-time input foundation for the subsequent main decision-making model, logic control module, and risk assessment model, so that the system can understand not only "what the vehicle sees," but also "what state the vehicle is currently in" and "what the vehicle will do next."

[0038] S2. Input the scene perception information, the vehicle status information, and the driving task information into the main decision-making model to perform scene understanding and driving strategy generation, and obtain the initial decision result.

[0039] The main decision-making model refers to the model responsible for high-level scene semantic understanding, behavioral reasoning, and policy generation. This main decision-making model can be a specially trained or fine-tuned large language model, a multimodal large model, a visual-language-action model, a driving agent model with chained reasoning capabilities, or a combination thereof. Its input can be structured data, sequential data, multimodal features, or prompt word templates, and its output can be behavioral decisions, control suggestions, trajectory intentions, risk warnings, action sequences, etc. For example, the main decision-making model can receive inputs such as "current speed 42km / h, slow vehicle ahead on the right, fast vehicle behind on the left, right lane change 150 meters ahead, traffic light 5 seconds remaining," and output an initial decision result of "temporarily change lanes, maintain current lane, decelerate and follow, wait for the vehicle behind on the left to pass, and then find a safe lane change window."

[0040] The scenario understanding refers to the high-level semantic interpretation process of the master decision-making model for complex environmental states. For example, the model can understand "a complex mixed-traffic scenario where the vehicle is preparing to turn right at an intersection," "a potential risk of cutting in from the right-hand non-motorized vehicle lane," and "a rapidly approaching vehicle from the left rear posing a threat to an immediate lane change." The driving strategy generation refers to the master decision-making model generating behavioral decisions or control strategies that conform to the current scenario after comprehensively considering environmental constraints, vehicle status, and task objectives. Examples include maintaining lane position, changing lanes to the left or right, slowing down, stopping and waiting, driving slowly around the intersection, and proceeding cautiously.

[0041] In some optional implementations, S2 specifically includes: performing feature encoding on scene perception information to obtain environmental representation features; performing state encoding on vehicle state information to obtain motion state features; performing task encoding on driving task information to obtain task constraint features; fusing environmental representation features, motion state features, and task constraint features and inputting them into the main decision-making model; and outputting the initial decision result corresponding to the current driving scenario through the main decision-making model.

[0042] Feature encoding can be implemented using convolutional neural networks, visual Transformers, graph neural networks, point cloud networks, or dense encoders; state encoding can be implemented using fully connected networks, temporal modeling networks, or state embedding layers; and task encoding can be implemented using text embedding layers, planning prior embedding layers, or instruction understanding models. Fusion methods can include concatenation, cross-attention, gating fusion, graph structure fusion, or unified prompt word input.

[0043] The beneficial effect of this step is that by introducing a master decision-making model with high-level reasoning capabilities, we can get rid of the problems of rigid decision-making logic and insufficient contextual understanding in traditional modular systems, enabling the system to form initial decision results that are closer to human driving semantic logic in complex and dynamic traffic environments.

[0044] S3. Input the initial decision result into the logic control module for rule constraint verification, process arrangement and control instruction generation to obtain vehicle control instructions.

[0045] The logic control module refers to the functional unit used to perform rule verification, constraint correction, execution refinement, and control implementation on the output results of the main decision-making model. The logic control module can be implemented using a rule engine, finite state machine, behavior tree, conditional decision network, optimizer, or a combination thereof.

[0046] The rule constraint verification refers to checking the initial decision result based on a preset traffic rule library, safety constraint library, vehicle dynamics constraints, and system operation constraints. For example, if the main decision model outputs "change lanes to the right immediately," but there is a nearby obstacle in the current right lane or the vehicle's lateral acceleration constraint does not allow it, the logic control module can determine that the decision does not meet the execution conditions.

[0047] The process orchestration refers to generating corresponding execution processes based on different decision types. For example, for "maintaining lane speed and decelerating to follow," the process orchestration may include: maintaining current lateral control, reducing target speed, updating following distance, and monitoring the deceleration trend of the vehicle in front. For "waiting for the vehicle behind to pass before changing lanes to the right," the process orchestration may include: first entering a lane change waiting state, continuously calculating the right lane change safety clearance, and issuing a lateral displacement reference trajectory after the conditions are met.

[0048] The control command generation refers to the transformation of high-level decision results into execution-level commands. Vehicle control commands may include, but are not limited to: desired acceleration, desired deceleration, target steering angle, trajectory reference point, desired curvature, target lane number, and behavior status flags. For example, the logic control module outputs "In the next 2 seconds, reduce the target speed from 42 km / h to 28 km / h while maintaining the current lane centerline tracking."

[0049] In one alternative approach, if the logic control module determines that the initial decision violates traffic rules, safe distance constraints, or dynamic constraints, it can trigger logic correction, downgrade processing, or execution blocking. Logic correction refers to making partial adjustments to non-compliant decisions, such as changing "rapid lane change" to "delayed lane change"; downgrade processing refers to switching to a more conservative control mode, such as downgrading from active overtaking to low-speed following; execution blocking refers to preventing the current decision from directly entering the execution layer and feeding back the abnormal information upstream.

[0050] The beneficial effect of this step is that it avoids the unconstrained output of the main decision model being directly applied to the vehicle control layer, thereby improving the safety and engineering controllability of the system decision-to-execution link.

[0051] S4. Input the vehicle control command, the scene perception information, and the vehicle operation feedback information into the risk assessment big data model for safety verification and anomaly identification to obtain the risk assessment result.

[0052] Vehicle operation feedback information refers to the feedback status information after the actual execution of control commands. This feedback information may include the vehicle's actual speed, actual acceleration, actual yaw rate, actual trajectory deviation, control response delay, execution error, tire slippage, braking effect, and vehicle stability. For example, after the logic control module issues a deceleration command, the vehicle's actual deceleration is insufficient; or, the system plans to maintain the current lane, but the vehicle deviates from its trajectory due to crosswinds.

[0053] A comprehensive risk assessment model refers to a model that jointly analyzes current control strategies, scenario evolution trends, and execution feedback to identify potential risks, explicit risks, and abnormal situations. This model can be a large-scale language model trained on driving risk scenarios, a multimodal risk inference model, a scenario graph risk identification model, a time-series risk prediction model, or a combination thereof. The comprehensive risk assessment model can output information such as risk type, risk level, risk source, risk location, risk duration, risk triggers, and mitigation recommendations.

[0054] The risk types may include, but are not limited to: collision risk, rear-end collision risk, side scrape risk, boundary crossing risk, running a red light risk, entering an impassable area risk, control mismatch risk, perception anomaly risk, rule conflict risk, model output inconsistency risk, and execution lag risk. The risk level may be expressed as low risk, medium risk, high risk, or a numerical score.

[0055] For example, in one embodiment, the main decision model gives an initial decision of "can immediately merge into the right lane", and the logic control module generates a corresponding lane change command. However, the risk assessment model combines the scene perception information of an electric vehicle suddenly accelerating and approaching in the right lane, as well as the feedback of the vehicle's lateral control execution lag, to determine that the current lane change has a medium to high risk, and outputs "Risk type: Right target cutting conflict; Risk level: High; Risk triggering reason: Right dynamic obstacle accelerating and approaching and vehicle lateral execution delay".

[0056] In some implementations, the large-scale risk assessment model can further perform anomaly identification. Anomaly identification includes, but is not limited to: abrupt changes in perception results, sensor information conflicts, conflicts between model decisions and rule bases, excessive discrepancies between execution results and expectations, and inconsistencies in outputs from upstream and downstream modules. For example, if a forward-facing camera identifies the road ahead as empty, but the millimeter-wave radar continuously detects static obstacles, the large-scale risk assessment model can identify this as a "perception inconsistency anomaly."

[0057] The technical effect of this step is to extend risk identification in the autonomous driving process from traditional threshold comparison and simple collision detection to higher-level semantic reasoning and abnormal causal analysis, thereby improving the system's ability to identify complex risk scenarios.

[0058] S5. When the risk assessment result indicates the existence of a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment result and the alarm information are fed back to the main decision model for decision correction to obtain the corrected decision result.

[0059] The alarm generation module is a functional unit that generates alarm information for the system, driver, remote platform, or recording system after identifying risk events or abnormal events. The alarm generation module can be implemented using a language generation model, template generator, event trigger, visualization engine, or a combination thereof.

[0060] The alarm information may include structured alarm information and natural language alarm information. Structured alarm information may include: alarm number, alarm time, risk level, risk category, risk location, associated target, and suggested measures; natural language alarm information may include text such as: "A fast-approaching electric vehicle has been detected on the right. There is a risk of collision if you attempt to change lanes to the right. It is recommended to cancel the lane change immediately and slow down to maintain your lane."

[0061] In one implementation, the alarm generation module not only outputs interpretable alarms for human-computer interaction, but also provides feedback constraints to the main decision-making model. For example, when the risk assessment result is "the current lane change risk level is high, it is recommended to cancel the lane change and prioritize maintaining lane speed and deceleration", this information is fed back to the main decision-making model as a new constraint context, prompting it to regenerate a more conservative corrective decision result.

[0062] The decision correction refers to the process by which the main decision-making model, after receiving risk assessment results and warning information, reassesses the current scenario and task objectives, and outputs an updated driving strategy. For example, if the initial decision is "prepare to change lanes and turn right immediately," after receiving high-risk feedback, the corrected decision could be "slow down and give way, maintain the current lane, and assess lane-changing conditions after the target on the right has passed."

[0063] The key value of this step lies in enabling alarm results to go beyond one-way notifications and instead substantially participate in the autonomous driving decision-making process, forming a closed-loop optimization mechanism of "risk detection - alarm generation - decision feedback - re-decision".

[0064] S6. The corrected decision result is re-input into the logic control module to generate updated vehicle control commands, and an autonomous driving closed-loop decision output is formed based on the updated vehicle control commands.

[0065] The closed-loop decision output for autonomous driving refers to the final, executable control result determined after collaborative processing by the main decision-making model, logic control module, risk assessment model, and alarm generation module. The closed-loop decision output can be either a specific control command or a behavioral strategy with state constraints. For example, "Maintain lane speed and decelerate to 20 km / h for the next 3 seconds, continuously monitor the safe clearance of the right lane, and execute a right lane change if the condition is met for 1 consecutive second."

[0066] In some implementations, the closed-loop decision output can further enter the trajectory tracking controller, motion controller, or actuator control layer to drive the vehicle's steering, braking, and powertrain systems to perform corresponding actions. Simultaneously, new perception information and execution feedback can re-enter the aforementioned process, thus forming a continuously iterative closed-loop control chain.

[0067] The technical solution of this embodiment collaboratively models scene perception information, vehicle status information, and driving task information. It utilizes a master decision-making model to achieve high-level scene understanding and driving strategy generation. The logic control module performs rule verification and execution refinement on the model output. Furthermore, it establishes a risk identification, feedback correction, and closed-loop optimization mechanism through a risk assessment model and an alarm generation module. This solution addresses the problems of limited single-model decision-making capabilities, incomplete risk warning links, difficulty in using alarm information to participate in decision-making, and insufficient system robustness in existing autonomous driving solutions. It improves the decision-making accuracy, safety, robustness, and interpretability of the autonomous driving system.

[0068] In one alternative approach, S1 specifically includes: S11. Obtain environmental data around the vehicle by using at least one of a camera, lidar, or millimeter-wave radar installed on the target vehicle.

[0069] The cameras can be front-view cameras, surround-view cameras, rear-view cameras, fisheye cameras, or binocular cameras; the lidar can be mechanical lidar, solid-state lidar, or hybrid lidar; and the millimeter-wave radar can provide relative distance, relative speed, and angle information. Through the coordination of these sensors, comprehensive perception of both the static environment and dynamic traffic participants can be achieved.

[0070] S12. Perform at least one of the following processing on the environmental data: target detection, lane line recognition, traffic sign recognition, and passable area recognition, to obtain the scene perception information.

[0071] For example, a visual detection network can be used to identify vehicles, pedestrians, non-motorized vehicles, and traffic cones; a lane line segmentation network can be used to identify solid lines, dashed lines, stop lines, and road boundaries; a traffic light recognition network can be used to obtain the light color status; and a point cloud segmentation network can be used to identify three-dimensional obstacles and free space.

[0072] S13. Obtain the vehicle's position, speed, acceleration, heading angle, and attitude information through the target vehicle's positioning unit, inertial measurement unit, and vehicle bus to obtain the vehicle's state information.

[0073] The positioning unit can use GNSS, RTK, visual positioning, laser positioning, or a combination of navigation methods; the inertial measurement unit can acquire angular velocity and linear acceleration; and the vehicle bus can acquire the chassis execution status.

[0074] S14. Obtain at least one of the following information: navigation path, current driving stage, driving intention, and system control target, to obtain the driving task information.

[0075] For example, the current driving intention could be "go straight along the main road", "prepare to pull over", "turn right at the intersection ahead", or "pass through the construction area at low speed".

[0076] The advantages of adopting the above-mentioned optional methods are as follows: through the joint work of multi-source sensors, positioning system, vehicle bus and task planning module, comprehensive collection and unified representation of input information can be achieved, providing a high-quality data foundation for subsequent large-scale model collaborative reasoning.

[0077] In one alternative approach, S2 specifically includes: S21. The scene perception information is feature-encoded to obtain environmental representation features.

[0078] Environmental representation features refer to the semantic feature representations of surrounding targets, road structures, traffic rule elements, and environmental conditions after encoding. For example, the encoding results can include environmental semantics such as "slow vehicle ahead", "non-motorized vehicle on the right", "straight ahead but not right turn allowed in the current lane", and "construction barriers ahead".

[0079] S22. The vehicle state information is encoded to obtain motion state features.

[0080] Motion state characteristics may include the vehicle's current speed, acceleration and deceleration trends, yaw stability, directional control trends, and actuator response status.

[0081] S23. Perform task encoding on the driving task information to obtain task constraint features.

[0082] Task constraint features can reflect destination constraints, lane-level navigation requirements, timeliness priorities, comfort requirements, and safety priority strategies.

[0083] S24. The environmental representation features, motion state features, and task constraint features are fused and input into the main decision-making model.

[0084] The fusion methods can include vector concatenation, cross-modal cross-attention, sequential text prompts, and unified tokenized input. For example, structured states can be converted into text prompts such as "Current speed 42km / h, slow car ahead on the right, fast car behind on the left, navigation requires changing lanes to the right in 150 meters," and then input together with visual features into the main decision-making model.

[0085] S25. Output the initial decision result corresponding to the current driving scenario through the main decision model.

[0086] Initial decision results may include, but are not limited to: behavior category, priority ranking, target lane selection, target speed range, expected turning trend, and suggested trajectory scheme.

[0087] The advantages of adopting the above-mentioned optional approach are that it enables comprehensive reasoning of three types of information—environment, state, and task—in a unified semantic space, thereby enhancing the global consistency of decision-making results and the adaptability to different scenarios.

[0088] In one alternative approach, S3 specifically includes: S31. Based on the preset traffic rule library, safety constraint library and vehicle dynamics constraints, perform compliance verification on the initial decision results.

[0089] The traffic rule library may include road traffic regulations, lane rules, traffic light rules, yielding rules, etc.; the safety constraint library may include minimum safe distance, minimum lane change clearance, collision time threshold, etc.; vehicle dynamics constraints may include maximum lateral acceleration, minimum turning radius, maximum braking force, etc.

[0090] S32. When the initial decision result meets the constraints, the control flow is arranged on the initial decision result to generate execution-level vehicle control instructions.

[0091] For example, the decision to "change lanes at low speed" can be programmed as follows: "First, reduce speed to 25 km / h — confirm that there are no vehicles to the right rear — perform a smooth lateral displacement — resume the target speed after completion."

[0092] S33. When the initial decision result does not meet the constraints, the initial decision result is logically corrected, downgraded, or its execution is blocked, and the constraint correction result is output.

[0093] For example, change "accelerate to overtake immediately" to "keep following", or block automatic lane changing when an anomaly is detected.

[0094] The beneficial effects of adopting the above optional methods are: by re-filtering the output of large models through explicit rules and control logic, the direct execution of unsafe decisions can be effectively suppressed.

[0095] In one alternative approach, S4 specifically includes: S41. Obtain the target vehicle's execution feedback information on the vehicle control command.

[0096] S42. The execution feedback information, the scene perception information, and the vehicle control command are jointly input, and the risk assessment big model is used to identify at least one of the following: collision risk, boundary crossing risk, rule conflict risk, perception anomaly risk, and control mismatch risk.

[0097] S43. Output the risk assessment results, which include at least one of risk type, risk level, risk location, and risk triggering cause.

[0098] For example, in a specific scenario, the system plans to turn right, but the risk assessment model identifies a pedestrian suddenly appearing in the right-turn area, and the vehicle's deceleration is insufficient. In this case, the system outputs "High risk, right-turn conflict, it is recommended to brake immediately and suspend the turning action".

[0099] The advantages of adopting the above-mentioned optional methods are: the system status can be continuously tracked after and during the execution of control, and complex risks that are difficult to describe by traditional rule-based methods can be identified in a timely manner.

[0100] In one alternative approach, S5 specifically includes: S51. When the risk level reaches the preset threshold, the alarm generation module is triggered to generate alarm information.

[0101] S52, wherein the alarm information includes at least one of risk description information, risk level information, suggested handling strategy, and manual takeover prompt.

[0102] S53. Input the risk assessment results and the alarm information as feedback constraint information into the main decision-making model.

[0103] S54. The main decision model is used to regenerate the corrected decision result that matches the current scenario, so as to form a closed-loop decision update.

[0104] In one example, when the system detects that "the large vehicle on the left is obstructing the view of the intersection ahead, and the risk of passing through the intersection is high", the alarm generation module can output "the view is obstructed, it is recommended to slow down and observe or stop and wait", and feed this constraint back to the main decision model to generate a new conservative decision.

[0105] The beneficial effects of adopting the above-mentioned optional methods are: it enables strategy reconstruction driven by risk identification, giving the system dynamic self-correction capabilities.

[0106] In one alternative approach, it also includes: S61. Perform a consistency comparison on the output results of the main decision-making model, the risk assessment model, and the alarm generation module.

[0107] S62. When the consistency comparison result is less than the preset consistency threshold, at least one of the following mechanisms is triggered: model negotiation mechanism, conservative decision-making mechanism, or manual takeover mechanism.

[0108] S63, wherein the conservative decision-making mechanism includes at least one of deceleration, stopping, lane keeping, and lane change restriction.

[0109] For example, if the main decision-making model believes that "it is safe to overtake", while the risk assessment model believes that "there is a medium to high risk", the system can trigger a conservative mechanism and prioritize slowing down and following.

[0110] The advantages of adopting the above-mentioned optional methods are: when there are conflicts of opinion or uncertain outputs among multiple models, consistency comparison and conservative strategy selection can help improve the overall level of safety redundancy.

[0111] like Figure 2 As shown, the overall framework of this embodiment specifically includes the following processing steps: 1) The target vehicle collects scene perception information, vehicle status information, and driving task information through cameras, lidar, millimeter-wave radar, positioning units, inertial measurement units, and on-board bus. Scene perception information may include dynamic targets, static obstacles, lane structure, traffic signals, passable areas, etc.; vehicle status information may include speed, acceleration, attitude, and execution status, etc.; driving task information may include navigation path, lane change requirements, and steering requirements, etc.

[0112] 2) The collected information is uniformly fed into the main decision-making model. The main decision-making model can perform high-level semantic modeling of the current driving scenario through multimodal feature fusion, contextual semantic understanding, and chain reasoning, and output initial decision results. The initial decision results can be behavioral strategies, target speeds, target lanes, or trajectory references, etc.

[0113] 3) Initial decision results are input into the logic control module. The logic control module verifies the results based on the rule base, safety constraint base, and vehicle dynamics constraints, and converts the decision results that meet the conditions into execution-level control instructions; if the decision is found to be unreasonable, it is corrected, downgraded, or blocked.

[0114] 4) After the vehicle executes the control command, the system acquires vehicle operation feedback information, including actual speed, trajectory deviation, execution delay, and stability indicators. Simultaneously, the latest scene perception information and control commands are input into the risk assessment model. The risk assessment model, combining current environmental evolution and control feedback, analyzes risks such as potential collisions, boundary crossings, perception anomalies, control mismatches, and rule conflicts, yielding the risk assessment results.

[0115] 5) When the risk assessment results indicate the existence of a risk event, the alarm generation module generates alarm information. The alarm information can be provided to the driver, remote monitoring platform, or recording system, or it can be fed back to the main decision-making model as a new decision constraint, causing the main decision-making model to re-output the revised decision result.

[0116] 6) The revised decision result enters the logic control module again and is issued for execution, thus forming a closed-loop decision architecture of "multi-source perception - main decision reasoning - logic control - risk assessment - alarm feedback - re-decision".

[0117] Through the above overall framework, this invention not only realizes risk identification and safety protection in the autonomous driving control link, but also realizes collaborative linkage and feedback optimization between large models, enhancing the continuous decision-making ability of the autonomous driving system in complex dynamic environments.

[0118] Figure 3 The diagram shows a structural schematic of an embodiment of an autonomous driving risk warning and closed-loop decision-making system 200 based on large model collaboration provided by the present invention.

[0119] like Figure 3 As shown, the system 200 includes: 1) Information acquisition module 201, used to acquire scene perception information, vehicle status information and driving task information of the target vehicle; 2) The main decision module 202 is used to input the scene perception information, the vehicle status information and the driving task information into the main decision model to perform scene understanding and driving strategy generation, and obtain the initial decision result; 3) Logic control module 203 is used to perform rule constraint verification, process arrangement and control instruction generation on the initial decision results to obtain vehicle control instructions; 4) Risk assessment module 204 is used to input the vehicle control command, the scene perception information and the vehicle operation feedback information into the risk assessment big model for safety verification and anomaly identification, and obtain the risk assessment result; 5) Alarm feedback module 205, used to output alarm information when the risk assessment result indicates the existence of a risk event, and to feed back the risk assessment result and the alarm information to the main decision module for decision correction; 6) Closed-loop output module 206, used to generate closed-loop decision output for autonomous driving based on the corrected decision results.

[0120] In one optional embodiment, the information acquisition module 201 is specifically used to: acquire environmental data through cameras, lidar, millimeter-wave radar, etc.; acquire vehicle status information through positioning units, inertial measurement units, and vehicle bus; and acquire driving task information through navigation modules and task planning modules.

[0121] In one alternative approach, the main decision module 202 is specifically used to: encode environmental features of scene perception information, encode state features of vehicle state information, encode task features of driving task information, and input the fused encoding results into the main decision model to output the initial decision result.

[0122] In an alternative approach, the logic control module 203 is specifically used to: perform traffic rule constraint verification, safety constraint determination, and dynamic executability analysis on the initial decision results, and convert the high-level decisions that meet the requirements into execution-level control instructions.

[0123] In one alternative approach, the risk assessment module 204 is specifically used to: acquire control execution feedback information and perform joint analysis with scene perception information and vehicle control commands to output risk type, risk level and risk cause.

[0124] In one alternative approach, the alarm feedback module 205 is specifically used to: generate structured or natural language alarm information when a high risk or abnormal risk is detected, and feed back the risk constraints to the main decision module 202 so that the main decision module 202 can output the corrected decision result.

[0125] In an alternative embodiment, the system 200 further includes a consistency management module for performing consistency analysis on the outputs of the main decision-making module 202, the risk assessment module 204, and the alarm feedback module 205, and triggering a conservative decision-making mechanism or a manual takeover mechanism when output conflicts occur.

[0126] It should be noted that the beneficial effects of the system 200 provided in the above embodiments are basically the same as those of the method embodiments described above, and will not be repeated here. Furthermore, the functional modules in system 200 can be implemented in software, hardware, or a combination of both.

[0127] In some embodiments, system 200 may be a computer program running on a vehicle domain controller or deployed in a cloud-edge-device collaborative architecture. For example, the main decision-making module 202 may run on a high-performance computing platform, the logic control module 203 may run on a real-time controller, and the risk assessment module 204 and alarm feedback module 205 may run on a vehicle-side safety computing unit or an edge server.

[0128] An electronic device according to an embodiment of the present invention includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements any of the above-mentioned autonomous driving risk warning and closed-loop decision-making methods based on large model collaboration.

[0129] In one alternative embodiment, such as Figure 4 As shown, the electronic device 4000 includes a processor 4001, a bus 4002, a memory 4003, and a transceiver 4004. The processor 4001 and the memory 4003 can be connected via the bus 4002. The transceiver 4004 can be used for data exchange between the electronic device 4000 and other electronic devices, sensor nodes, remote servers, or roadside devices.

[0130] The processor 4001 can be a central processing unit (CPU), a graphics processing unit (GPU), a digital signal processor (DSP), a neural network processor (NPU), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or any combination thereof. The processor 4001 can be used to perform operations such as scene perception information processing, vehicle status information parsing, driving task information management, master decision reasoning, logical constraint verification, risk assessment, and alarm generation.

[0131] Bus 4002 may include an address bus, a data bus, and a control bus, used for transmitting information between various components of electronic device 4000. For ease of explanation, Figure 4 The image shows only one bus, but this does not mean that the device has only one bus.

[0132] The memory 4003 can be random access memory (RAM), read-only memory (ROM), flash memory, solid-state drive (SSD), embedded multimedia card, disk storage medium, or other media suitable for storing programs and data. The memory 4003 can store application code, model parameters, rule bases, security constraint libraries, log data, and runtime configurations used to execute the methods of this invention.

[0133] The transceiver 4004 can be an Ethernet interface, CAN communication interface, FlexRay interface, 5G communication module, Wi-Fi module, Bluetooth module, vehicle-road cooperative communication module or other communication unit, used to receive sensing data, navigation information, remote control information or send alarm data, vehicle status data and execution result data.

[0134] Among them, the electronic device 4000 can be an in-vehicle autonomous driving controller, a driving assistance domain controller, an edge computing device, a cloud control server, a remote dispatch terminal, a vehicle-road cooperative edge node, etc.

[0135] It should be noted that, Figure 4 The electronic devices shown are merely examples and should not be construed as limiting the scope of the invention.

[0136] This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the above-mentioned autonomous driving risk warning and closed-loop decision-making methods based on large model collaboration.

[0137] Optionally, the computer-readable storage medium may be a read-only memory (ROM), random access memory (RAM), flash memory, optical disc, magnetic disk, magnetic tape, portable memory card, or other tangible medium capable of storing program instructions. The program may run on a single device or distributed across multiple devices.

[0138] In an exemplary embodiment, a computer program product or computer program is also provided. The computer program product or computer program includes computer instructions that can be stored in a computer-readable storage medium. When a processor of an electronic device reads and executes the computer instructions, the electronic device performs the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as shown in any of the above embodiments.

[0139] The program code can be written in one or more programming languages, including but not limited to object-oriented programming languages ​​such as Java, C++, and Python, as well as procedural programming languages ​​such as C or similar languages. The program code can be executed entirely on the user's device, partially on the user's device and partially on a remote device, or entirely on a remote server.

[0140] It should be understood that the flowcharts and block diagrams in the accompanying drawings illustrate possible implementation architectures, functions, and operations of systems, devices, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function.

[0141] It should also be noted that in some alternative implementations, the functions marked in the boxes may occur in a different order than those marked in the accompanying drawings. For example, two consecutively given boxes may actually be executed in substantially parallel order, or they may be executed in reverse order, depending on the functions involved. Each box and its combination in the block diagram and flowchart can be implemented either by a dedicated hardware system that performs the specified function or action, or by a combination of dedicated hardware and computer instructions.

[0142] The computer-readable storage medium provided in the embodiments of the present invention may be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any combination thereof. More specific examples of computer-readable storage media include, but are not limited to: portable hard disks, hard disks, RAM, ROM, EPROM, flash memory, optical fiber, CD-ROM, optical storage devices, magnetic storage devices, etc.

[0143] The computer-readable storage medium described above carries one or more programs, which, when executed by an electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0144] The above description is merely a preferred embodiment of the present invention and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of protection involved in this invention is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of the present invention. For example, technical solutions formed by substituting the above-described features with technical features having similar functions disclosed in this invention should all fall within the scope of protection of this invention.

[0145] It should be noted that the terms "first," "second," etc., used in this application specification and claims are only used to distinguish the objects being described and do not represent a specific order, priority, or hierarchical relationship. Where appropriate, the order of use for similar objects can be interchanged so that embodiments of this application can be implemented in an order other than the order illustrated or described.

[0146] Those skilled in the art should understand that this invention can be implemented as a system, method, or computer program product. Therefore, this invention can be specifically implemented in a completely hardware form, a completely software form, or a combination of hardware and software. It is generally referred to herein as a "module," "unit," or "system."

[0147] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the spirit and principles of the present invention, and such changes, modifications, substitutions and variations should also fall within the protection scope of the present invention.

Claims

1. A method for risk warning and closed-loop decision-making in autonomous driving based on large-scale model collaboration, characterized in that, include: Acquire scene perception information, vehicle status information, and driving task information of the target vehicle; The scene perception information, the vehicle status information, and the driving task information are input into the main decision-making model to perform scene understanding and driving strategy generation, and an initial decision result is obtained. The initial decision results are input into the logic control module for rule constraint verification, process orchestration, and control command generation to obtain vehicle control commands. The vehicle control commands, the scene perception information, and the vehicle operation feedback information are input into the risk assessment model for safety verification and anomaly identification to obtain the risk assessment results. When the risk assessment results indicate that there are no risk events, an autonomous driving closed-loop decision output is formed based on the vehicle control commands. When the risk assessment results indicate the existence of a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment results and the alarm information are fed back to the main decision-making model for decision correction to obtain the corrected decision result; Based on the revised decision results, vehicle control commands are regenerated, and a closed-loop decision output for autonomous driving is formed.

2. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as described in claim 1, characterized in that, The acquisition of scene perception information, vehicle status information, and driving task information of the target vehicle includes: acquiring environmental data around the vehicle by using at least one of a camera, lidar, and millimeter-wave radar installed on the target vehicle; and processing the environmental data by at least one of target detection, lane line recognition, traffic sign recognition, and passable area recognition to obtain the scene perception information. The vehicle's position, speed, acceleration, heading angle, and attitude information are obtained through the target vehicle's positioning unit, inertial measurement unit, and vehicle bus to obtain the vehicle's state information; The driving task information is obtained by acquiring at least one of the following: navigation path, current driving stage, driving intention, and system control objective.

3. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration according to claim 2, characterized in that, The process involves inputting the scene perception information, vehicle state information, and driving task information into the main decision-making model for scene understanding and driving strategy generation to obtain initial decision results, including: The scene perception information is feature-encoded to obtain environmental representation features; The vehicle state information is state encoded to obtain motion state features; The driving task information is encoded to obtain task constraint features; The environmental representation features, motion state features, and task constraint features are fused and input into the main decision-making model; The main decision model outputs an initial decision result corresponding to the current driving scenario. The initial decision result includes at least one of the following: longitudinal control strategy, lateral control strategy, behavioral decision result, and trajectory planning result.

4. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration according to claim 3, characterized in that, The initial decision result is input into the logic control module for rule constraint verification, process orchestration, and control command generation to obtain vehicle control commands, including: The initial decision results are verified for compliance based on a pre-defined traffic rule library, safety constraint library, and vehicle dynamics constraints. When the initial decision result meets the constraints, the control flow is arranged to generate execution-level vehicle control commands. When the initial decision result does not meet the constraints, the initial decision result is logically corrected, downgraded, or its execution is blocked. When logic correction or downgrade processing is performed, updated vehicle control commands are generated based on the correction results. When the block is executed, output the minimum risk maneuver command, the current lane maintenance command, or the manual takeover request, instead of continuing the original control chain.

5. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration according to claim 4, characterized in that, The process involves inputting the vehicle control commands, scene perception information, and vehicle operation feedback information into a large-scale risk assessment model for safety verification and anomaly identification, resulting in a risk assessment outcome, including: Obtain the execution feedback information of the target vehicle to the vehicle control command, wherein the execution feedback information includes at least one of control response information, trajectory deviation information and vehicle stability information; The execution feedback information, the scene perception information, and the vehicle control command are jointly input, and the risk assessment big model is used to identify at least one of the following: collision risk, boundary crossing risk, rule conflict risk, perception anomaly risk, and control mismatch risk. Output the risk assessment results, which include at least one of the following: risk type, risk level, risk location, and risk triggering cause.

6. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as described in claim 5, characterized in that, When the risk assessment result indicates the existence of a risk event, the alarm generation module is triggered to output alarm information, and the risk assessment result and the alarm information are fed back to the main decision-making model for decision correction, resulting in a corrected decision result, including: When the risk level reaches a preset threshold, the alarm generation module is triggered to generate alarm information; wherein, the alarm information includes at least one of risk description information, risk level information, suggested handling strategy, and manual takeover prompt; The risk assessment results and the alarm information are input as feedback constraint information into the master decision-making model; The main decision model is used to regenerate a revised decision result that matches the current scenario, thus forming a closed-loop decision update.

7. The autonomous driving risk warning and closed-loop decision-making method based on large model collaboration according to any one of claims 1 to 6, characterized in that, Also includes: The outputs of the main decision-making model, the risk assessment model, and the alarm generation module are compared for consistency. The consistency comparison includes at least determining whether there is a conflict between the execution direction, risk level, and suggested handling strategy of the initial decision results. When the consistency comparison result is less than a preset consistency threshold, at least one of the following mechanisms is triggered: model negotiation mechanism, conservative decision-making mechanism, or manual takeover mechanism. The conservative decision-making mechanism includes at least one of the following: deceleration, stopping, lane keeping, and lane changing restriction. The conservative decision-making mechanism includes at least one of deceleration, stopping, lane keeping, and lane change restriction.

8. A risk warning and closed-loop decision-making system for autonomous driving based on large-scale model collaboration, characterized in that, include: The information acquisition module is used to acquire scene perception information, vehicle status information, and driving task information of the target vehicle. The main decision module is used to input the scene perception information, the vehicle status information and the driving task information into the main decision model to perform scene understanding and driving strategy generation, and obtain the initial decision result. The logic control module is used to perform rule constraint verification, process orchestration, and control command generation on the initial decision results to obtain vehicle control commands; The risk assessment module is used to input the vehicle control commands, the scene perception information, and the vehicle operation feedback information into the risk assessment big model for safety verification and anomaly identification, and to obtain the risk assessment results. The alarm feedback module is used to output alarm information when the risk assessment result indicates the existence of a risk event, and to feed back the risk assessment result and the alarm information to the main decision module for decision correction; The closed-loop output module is used to generate an autonomous driving closed-loop decision output based on vehicle control commands when the risk assessment result indicates that there is no risk event. When the risk assessment results indicate the presence of a risk event, the vehicle control commands are regenerated based on the revised decision results, and an autonomous driving closed-loop decision output is formed.

9. An electronic device, characterized in that, The electronic device includes a processor and a memory, wherein the memory stores at least one computer program, and when the at least one computer program is loaded and executed by the processor, the electronic device enables the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one computer program, which, when executed by a processor, implements the autonomous driving risk warning and closed-loop decision-making method based on large model collaboration as described in any one of claims 1 to 7.

Images