Vehicle driving task editing method and related device

Abstract

The application discloses a vehicle driving task editing method and related equipment, and relates to the technical field of vehicle control. The method comprises the following steps: obtaining an editing request for a current driving task of a target vehicle and current running state information of the target vehicle; generating a candidate trajectory of the target vehicle based on the editing request and the current running state information; performing continuous collision detection on the candidate trajectory to obtain collision distance time series data; determining a collision risk quantitative value for the candidate trajectory based on the collision distance time series data and a preset error uncertainty parameter; and determining a decision result for the editing request based on the collision risk quantitative value. Through the processing procedure from editing request acquisition, candidate trajectory generation, continuous collision detection, uncertainty risk evaluation to decision output, the safety judgment for the editing request can be realized in the dynamic driving task adjustment process, and the safety and controllability of the vehicle driving task editing are improved.

Description

Technical Field

[0001] This application relates to the field of vehicle control technology, and more specifically, to a method and related equipment for editing vehicle driving tasks. Background Technology

[0002] With the continuous development of intelligent driving technology, autonomous driving and assisted driving systems are being used more and more widely in passenger cars, commercial vehicles, and special vehicles. As a crucial function for enhancing vehicle intelligence, driving task management can plan, adjust, and execute vehicle driving tasks based on vehicle operating status, road environment, and user needs. In practical applications, drivers or remote operators may need to dynamically edit driving tasks that are being executed or awaiting execution, such as adjusting target speed, modifying driving routes, changing the task execution order, or updating target locations. Therefore, ensuring the safety and reliability of the driving task editing process while meeting user operational needs has become an important research direction in intelligent driving systems.

[0003] In related technologies, safety assessments for vehicle driving task editing typically rely on preset rules or simple constraint verification methods, such as threshold judgments for target speed, travel distance, or task parameters. These methods primarily detect single parameters or static conditions, making it difficult to fully consider the dynamic changes that occur during the actual execution of the driving task after editing. They also fail to accurately reflect the risk changes between the candidate trajectory and surrounding obstacles during continuous movement. Especially in complex traffic environments or low-speed, high-precision scenarios, vehicle trajectories are affected by various uncertainties such as positioning errors, perception errors, control execution errors, and model prediction errors. Relying solely on deterministic geometric relationships or simple rules for judgment can easily lead to inaccurate risk assessment results. In other words, related technologies suffer from low accuracy in risk detection and insufficient reliability in safety decisions when dynamically editing vehicle driving tasks. Summary of the Invention

[0004] In the summary section of this application, the relevant technical solutions are described in general terms, and a series of simplified concepts are introduced. These concepts will be further elaborated in the detailed embodiments section. This summary section should not be construed as limiting the key or essential technical features of the claimed solutions, nor is it intended to limit the scope of protection of the claimed solutions.

[0005] The vehicle driving task editing method and related equipment provided in this application can improve the safety and controllability of vehicle driving task editing by performing a processing flow from editing request acquisition, candidate trajectory generation, continuous collision detection, uncertainty risk assessment to decision output, and making safety judgments on editing requests during dynamic driving task adjustment.

[0006] In a first aspect, this application provides a method for editing a vehicle driving task, comprising: acquiring an editing request for a current driving task of a target vehicle and current operating status information of the target vehicle; generating a candidate trajectory of the target vehicle based on the editing request and the current operating status information, wherein the candidate trajectory includes trajectory points of the target vehicle at multiple future time points; performing continuous collision detection on the candidate trajectory to obtain collision distance time-series data, wherein the collision distance time-series data is used to characterize the distance between the trajectory point and an obstacle at each time point; determining a collision risk quantification value for the candidate trajectory based on the collision distance time-series data and a preset error uncertainty parameter, wherein the preset error uncertainty parameter includes at least one of positioning error, perception error, control execution error, and model prediction error; and determining a decision result for the editing request based on the collision risk quantification value.

[0007] In some implementations, determining the collision risk quantification value for the candidate trajectory based on the collision distance time series data and a preset error uncertainty parameter includes: determining the equivalent distance variance corresponding to each moment in the collision distance time series data based on the preset error uncertainty parameter; determining the instantaneous collision probability at each moment using a preset normal cumulative distribution function based on the collision distance time series data and the equivalent distance variance to obtain a collision probability sequence; and determining the collision risk quantification value based on the collision probability sequence.

[0008] In some implementations, determining the collision risk quantification value based on the collision probability sequence includes: obtaining the upper limit of the instantaneous collision probability in the collision probability sequence as a first risk metric; determining a second risk metric based on the instantaneous collision probability at each moment in the collision probability sequence; and performing a weighted summation of the first risk metric and the second risk metric to obtain the collision risk quantification value.

[0009] In some implementations, determining the second risk metric based on the instantaneous collision probability at each moment in the collision probability sequence includes: performing time-weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric; and performing a preset exponential decay function operation on the cumulative risk metric to obtain the second risk metric.

[0010] In some embodiments, the vehicle driving task editing method further includes: obtaining scene parameters of the target vehicle's current location, wherein the scene parameters include at least one of parking space type, ambient light level, and road surface adhesion coefficient; determining a target time weight sequence matching the scene parameters based on a preset correspondence between scene parameters and time weight levels, wherein each moment in the target time weight sequence corresponds to a time weight value; and performing time-weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric, including: performing weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence based on the target time weight sequence to obtain the cumulative risk metric.

[0011] In some implementations, determining the equivalent distance variance corresponding to each moment in the collision distance time series data based on the preset error uncertainty parameter includes: if the positioning error, the perception error, the control execution error, and the model prediction error are independent of each other, then the sum of squares of the preset standard deviations corresponding to each error is calculated to obtain the equivalent distance variance corresponding to each moment; if at least two of the positioning error, the perception error, the control execution error, and the model prediction error are correlated, then the preset covariance matrix corresponding to each error source is obtained, and the joint variance of each error source is calculated based on the preset covariance matrix to obtain the equivalent distance variance corresponding to each moment.

[0012] In some implementations, generating a candidate trajectory for the target vehicle based on the edit request and the current operating status information includes: performing a hard safety constraint screening on the edit request based on the current vehicle speed and obstacle information in the current operating status information to obtain a safety screening result; if the safety screening result is satisfactory, then performing a dynamic reachability analysis based on the current vehicle state in the current operating status information to generate the candidate trajectory.

[0013] In some implementations, the step of performing hard safety constraint screening on the edit request based on the current vehicle speed and obstacle information in the current operating status information to obtain a safety screening result includes: determining a safe stopping distance based on a preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request; determining the difference between the current nearest obstacle distance in the obstacle information and the safe stopping distance as a safety margin; and generating a safety screening result indicating failure if the safety margin is less than a preset safety threshold.

[0014] In some implementations, determining the safe stopping distance based on the preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request includes: determining the larger of the current vehicle speed and the target vehicle speed as the vehicle evaluation speed; determining the delayed travel distance by multiplying the vehicle evaluation speed by a preset system delay duration, wherein the preset system delay duration characterizes the response time required from triggering the braking condition to the target vehicle generating braking deceleration; determining the vehicle braking distance based on the ratio of the square of the current vehicle speed to twice the preset maximum braking deceleration; and determining the safe stopping distance by summing the delayed travel distance and the vehicle braking distance.

[0015] In some implementations, the preset system delay duration is determined by: determining the sum of the perception delay duration, decision delay duration, control command transmission delay duration, and actuator response delay duration to obtain the system delay duration; wherein, the perception delay duration represents the time interval from data acquisition by the environmental sensor to the output of structured perception results, the decision delay duration represents the time interval from receiving the perception results to generating control commands, the control command transmission delay duration represents the time interval from the transmission of control commands to the actuator via the communication bus, and the actuator response delay duration represents the time interval from the actuator receiving the command to starting to generate braking action.

[0016] In some implementations, if the safety screening result characterization is passed, then based on the current vehicle state in the current operating state information, a dynamic reachability analysis is performed to generate the candidate trajectory, including: if the safety screening result characterization is passed, then based on the edit request and the current task sequence in the current operating state information, a constraint consistency solution is performed to obtain constraint solution information; based on the constraint solution information and the current vehicle state in the current operating state information, a dynamic reachability analysis is performed to generate the candidate trajectory.

[0017] In some implementations, the constraint consistency solution based on the edit request and the current task sequence in the current running state information to obtain constraint solution information includes: parsing the constraints of the edit request to obtain incremental constraints; extracting tasks from the current task sequence to obtain at least one task constraint, wherein the task constraint includes at least one of precondition constraints, postcondition constraints, temporal dependency constraints, and safety boundary constraints; merging the incremental constraints and the task constraints to obtain a constraint set; performing consistency determination on the constraint set to obtain a determination result; if the determination result is inconsistent, identifying the conflict terms between the incremental constraints and the task constraints, and solving the minimum modification set of the constraint set based on preset relaxation rules, using the minimum modification set as the constraint solution information, wherein the minimum modification set is a set of parameter corrections that can eliminate the conflict terms; if the determination result is consistent, generating feasible assignments as the constraint solution information, wherein the feasible assignments are a set of control parameter values ​​that satisfy all constraint conditions.

[0018] In some implementations, the step of performing dynamic reachability analysis based on the constraint solution information and the current vehicle state in the current operating state information to generate the candidate trajectory includes: if the constraint solution information is the feasible assignment, then obtaining the first control parameter in the feasible assignment, and determining the first control input boundary based on the first control parameter and the current vehicle state; if the constraint solution information is the minimum modification set, then obtaining the second control parameter in the minimum modification set, and determining the second control input boundary based on the second control parameter and the current vehicle state; taking the current vehicle state as the initial state, within the first control input boundary or the second control input boundary, through a preset vehicle model... The system performs forward reachability propagation to obtain a set of reachable states, wherein the set of reachable states represents the continuous spatial state envelope region that the target vehicle can reach at each time under the constraints of the first control input boundary or the second control input boundary. If the set of reachable states meets the preset dynamic constraints, a nominal control sequence is selected from the corresponding control input boundary for forward simulation to obtain the candidate trajectory. The preset dynamic constraints include at least one of lateral acceleration constraints, yaw rate constraints, and minimum turning radius constraints. The nominal control sequence is a set of time-series control commands that satisfy the preset dynamic constraints within the first control input boundary or the second control input boundary.

[0019] In some embodiments, the step of performing continuous collision detection on the candidate trajectory to obtain collision distance time series data includes: acquiring obstacle prediction trajectories; aligning the candidate trajectory and the obstacle prediction trajectory with a preset time step using a time grid to obtain an aligned candidate trajectory time series and an aligned obstacle prediction trajectory time series; performing geometric expansion operations on the vehicle outline of the target vehicle and the obstacle outline corresponding to the obstacle prediction trajectory with a preset safe expansion distance to obtain a vehicle expansion region sequence and an obstacle expansion region sequence; for each aligned time point, determining the minimum distance between the vehicle expansion region and the obstacle expansion region corresponding to that time point to obtain a minimum distance sequence; determining the intervals in the minimum distance sequence that are continuously less than a preset distance threshold as collision intervals; and determining the minimum distance sequence and the collision intervals as the collision distance time series data.

[0020] In some implementations, determining the minimum distance between the vehicle expansion region and the obstacle expansion region at the specified time includes: performing intersection detection on the vehicle expansion region and the obstacle expansion region to obtain the minimum detection distance between them; if the minimum detection distance is positive, then determining the minimum detection distance as the minimum distance; if the minimum detection distance is zero or negative, then calculating the penetration depth between the vehicle expansion region and the obstacle expansion region using an expansion polyhedron algorithm, and determining the minimum distance based on the penetration depth.

[0021] In some implementations, determining the decision result for the edit request based on the collision risk quantification value includes: if the collision risk quantification value is greater than or equal to a first preset risk threshold, generating a first decision result of rejecting the edit request; if the collision risk quantification value is less than a second preset risk threshold, generating a second decision result of accepting the edit request, wherein the second preset risk threshold is less than the first preset risk threshold; if the collision risk quantification value is greater than or equal to the second preset risk threshold and less than the first preset risk threshold, generating a third decision result of postponing the edit request and waiting for manual confirmation.

[0022] Secondly, this application also provides a vehicle driving task editing device, comprising: a data acquisition unit, configured to acquire an editing request for a current driving task of a target vehicle and current operating status information of the target vehicle; a trajectory generation unit, configured to generate candidate trajectories of the target vehicle based on the editing request and the current operating status information, wherein the candidate trajectories include trajectory points of the target vehicle at multiple future time points; a collision timing unit, configured to perform continuous collision detection on the candidate trajectories to obtain collision distance timing data, wherein the collision distance timing data is used to characterize the distance between the trajectory point and an obstacle at each time point; a risk quantification unit, configured to determine a collision risk quantification value for the candidate trajectory based on the collision distance timing data and a preset error uncertainty parameter, wherein the preset error uncertainty parameter includes at least one of positioning error, perception error, control execution error, and model prediction error; and a decision determination unit, configured to determine a decision result for the editing request based on the collision risk quantification value.

[0023] Thirdly, this application also provides a vehicle, including: a memory and a processor, the processor being configured to execute a computer program stored in the memory to implement the steps of the vehicle driving task editing method described in the first aspect.

[0024] Fourthly, this application also provides a computer-readable storage medium storing computer-executable instructions or a computer program, wherein when the computer-executable instructions or the computer program are executed by a processor, the steps of the vehicle driving task editing method described in the first aspect are implemented.

[0025] Fifthly, this application also provides a computer program product, including a computer program or computer-executable instructions, wherein when the computer program or computer-executable instructions are executed by a processor, they implement the steps of the vehicle driving task editing method provided in the embodiments of this application.

[0026] In summary, this application obtains the edit request for the current driving task of the target vehicle and generates candidate trajectories by combining the vehicle's current operating status information. This allows the dynamic modification of the driving task to move beyond simple judgments based on preset rules. Instead, it performs a safety analysis based on the current vehicle status, task change requirements, and the actual operating environment. The candidate trajectories are generated after the edit request is correlated with the vehicle's real-time status, making the subsequent risk assessment process more closely aligned with the vehicle's actual operating conditions, thereby improving the adaptability and safety of the vehicle's driving task editing process. Furthermore, continuous collision detection is performed on the candidate trajectories to obtain collision distance time-series data characterizing the geometric proximity between the vehicle trajectory and obstacles at each moment. This is significantly more efficient than judging collisions only at a single location or moment. This method continuously analyzes the distance changes between the vehicle and obstacles along the execution process of the candidate trajectory, thus more comprehensively reflecting the potential collision risks during trajectory execution and reducing the risk omission caused by ignoring intermediate motion states. Based on collision distance time-series data and preset error uncertainty parameters including positioning error, perception error, control execution error, and model prediction error, the collision risk of the candidate trajectory is quantified. Because this risk quantification process introduces uncertainties present in vehicle operation, it avoids the limitations of relying solely on deterministic geometric distances for safety judgments, enabling the obtained collision risk quantification value to more accurately reflect the risk level under actual operating conditions and improving the reliability of risk assessment results. In summary, the vehicle driving task editing method provided in this application, through a processing flow from editing request acquisition, candidate trajectory generation, continuous collision detection, uncertainty risk assessment to decision output, can achieve safety judgment on editing requests during dynamic driving task adjustment, improving the safety and controllability of vehicle driving task editing. Attached Figure Description

[0027] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating a vehicle driving task editing method provided in this application embodiment; Figure 2 This application provides a schematic diagram of the composition of a vehicle driving task editing device according to an embodiment of the present application. Figure 3 This is a schematic diagram of the composition structure of a vehicle provided in an embodiment of this application. Detailed Implementation

[0028] The terms used in the specification, claims, and drawings of this application, such as "first," "second," "third," "fourth," etc. (if any), are used to distinguish similar objects and not to describe a specific order or sequence. Therefore, it is to be understood that these terms can be used interchangeably where appropriate, allowing the described embodiments to be used in different orders, unless specifically required by the illustrations or description. Furthermore, the terms "is" and "has," and any variations thereof, are intended to cover, non-exclusively, all possible constituent elements. For example, a process, method, system, product, or apparatus comprising several steps or units is not necessarily limited to the steps or units explicitly listed, but may also include other steps or units not explicitly listed, or steps or units inherent to the process, method, product, or apparatus.

[0029] In this application, a "module" or "unit" refers to a computer program or part of a computer program that has a specific function and works in conjunction with other related parts to achieve a predetermined goal. These modules or units can be implemented by software, hardware (e.g., processing circuitry or memory), or a combination of both. One or more processors or memories can implement one or more modules or units. Furthermore, each module or unit can also be part of a larger module or unit.

[0030] The technical solutions of this application will be described in detail below with reference to the accompanying drawings of the embodiments. It should be noted that the described embodiments are only a part of this application, and not all embodiments. In the following description, the "some embodiments" mentioned are only a subset of all possible embodiments, which may be the same or different subsets, and different embodiments can be combined with each other without conflict.

[0031] Figure 1 This is a flowchart illustrating a vehicle driving task editing method provided in an embodiment of this application. For example, see [link to example]. Figure 1 The vehicle driving task editing method provided in this application embodiment may include the following steps 101 to 105: Step 101: Obtain the edit request for the current driving task of the target vehicle and the current operating status information of the target vehicle.

[0032] In some examples, the target vehicle is an intelligent driving vehicle that is currently performing autonomous or assisted driving tasks and receiving dynamic editing instructions from the user. The vehicle is equipped with a complete intelligent driving in-vehicle system, including hardware and software modules such as environmental perception, positioning, decision planning, vehicle control, and in-vehicle human-machine interaction, which can support the implementation of the entire process of dynamic editing and safety verification of driving tasks.

[0033] The current driving task is a standardized, atomized intelligent driving task sequence that the target vehicle is executing or waiting to execute at the moment the editing request is triggered. It is generated by the upper-level task orchestration system and has a clear execution sequence, control parameters, safety boundaries, and completion conditions. It is the object of the user's dynamic editing operation. For example, the target vehicle's current driving task is a standard parallel parking reversing task. This task includes continuous sub-tasks such as constant speed reversing, lateral fine adjustment, and precise parking in the parking space. The preset initial parking speed is 0.5m / s, and the preset target parking space is the standard parallel parking space to the right of the current vehicle. It has a fixed task execution sequence and basic safety constraints.

[0034] An edit request is a dynamic task parameter modification command initiated by the user through a preset terminal based on the current driving task's running status. It is used to adjust the parameters of the intelligent driving task that is being executed or to be executed in real time. It is a dynamic incremental intervention command during driving and only applies to the currently active driving task, without affecting the global navigation or underlying system configuration. For example, if a user initiates an edit request through the in-vehicle human-machine interface to adjust the driving speed of the current parallel parking task from the initial preset 0.5m / s to 0.9m / s, this command is a compliant dynamic task parameter editing command, which can be normally captured by the in-vehicle system and trigger subsequent safety detection processes. In addition, common edit requests may also include similar dynamic modification operations such as adjusting the target parking space position, modifying the task execution order, and optimizing obstacle avoidance strategies.

[0035] The current operating status information is the full-dimensional real-time operating data of the target vehicle at the moment the edit request is triggered. It can include three types of data: vehicle body status, environmental perception status, and map scene status. It is the basic input for subsequent candidate trajectory generation and safety constraint verification, and can completely represent the current operating conditions of the vehicle and the surrounding environment. The vehicle body status includes the vehicle's real-time position, driving speed, heading angle, dynamic limit parameters, and fault status. The environmental perception status includes the ID, type, real-time position, driving speed, predicted trajectory, and perception confidence of surrounding obstacles. The map scene status includes the current lane topology, right-of-way information, road speed limit, and parking space semantic information. For example, the current operating status information of the target vehicle specifically includes: the vehicle's current driving speed is 0.5 m / s, the vehicle's heading angle matches the parallel parking reversing posture, and the vehicle's maximum braking deceleration, maximum lateral acceleration, and other dynamic parameters are normal. The perception module detects that the nearest obstacle to the vehicle is the curb, with a minimum distance of 0.9 m, and there are no dynamically moving obstacles. The map module recognizes that the current situation is a standard parallel parking scenario, the road speed limit and parking space parameters are compliant, and the vehicle has no hardware faults or system abnormalities.

[0036] For example, firstly, user operation commands can be captured in real time through multiple input interfaces of the in-vehicle human-machine interface, remote terminal, and upper-level task planner. When a valid command signal is detected, user authentication and command format parsing are immediately completed, illegal commands and abnormal interference signals are eliminated, and finally a standardized driving task editing request is encapsulated and generated. At the same time, a multi-module data synchronization acquisition mechanism is triggered to accurately align the timestamp of the editing request trigger, and simultaneously retrieve real-time sensor data of the vehicle chassis, high-precision positioning data, environmental perception obstacle detection data, and local high-precision map scene data to form complete information on the current operating status of the target vehicle. On this basis, the currently executing atomic driving task sequence is bound, and the editing request, the current driving task, and the current operating status information of the vehicle are associated and encapsulated to generate a uniquely identified editing instance to be detected. All context data of this dynamic editing operation is completely retained, providing accurate, complete, and time-aligned basic data support for subsequent candidate trajectory generation, collision detection, and risk quantification decision-making.

[0037] By implementing step 101, the user's request to modify the driving task can be linked to the vehicle's current actual operating status, providing an accurate data foundation for the subsequent generation of candidate trajectories. Since the generation of candidate trajectories is based on the vehicle's real-time operating status, it can avoid analyzing the editing request without considering the vehicle's current status, thus improving the consistency between subsequent trajectory planning and safety assessment and the vehicle's actual operating conditions.

[0038] Step 102: Based on the edit request and the current running status information, generate candidate trajectories for the target vehicle.

[0039] The candidate trajectory can include the trajectory points of the target vehicle at multiple future time points.

[0040] In some examples, the candidate trajectory is a set of expected motion trajectories of the target vehicle within a preset time period after responding to the driving task editing request. It is a simulated driving path derived from the vehicle's real-time operating conditions and editing instructions, used for subsequent collision detection and risk assessment calculations. This trajectory is not the actual driving trajectory that the vehicle ultimately executes, but rather preliminary trajectory data used for safety verification. It has complete temporal characteristics and can comprehensively characterize the vehicle's motion change trend after accepting the adjustment of editing parameters. For example, in the scenario of parallel parking, after the user initiates an editing request to increase the parking speed, the vehicle's motion trajectory within the next three seconds is generated. This candidate trajectory is divided into multiple temporal nodes according to a fixed time step, and each temporal node corresponds to a trajectory point, which sequentially records the vehicle's planar position, vehicle heading attitude, and instantaneous driving speed during the reversing process, completely covering the entire parking driving process after the editing parameters take effect.

[0041] The process of generating candidate trajectories for the target vehicle based on the edit request and current operating status information involves receiving the edit request data and the vehicle's current operating status data output in step 101, converting the parameter adjustment instructions in the edit request into new task control parameters, and using the vehicle's current position, speed, attitude, and other real-time states as the initial boundaries for trajectory extrapolation. Combined with the basic driving rules under the intelligent driving scenario, iterative calculation and fitting of the time-series trajectory are completed. For example, if the vehicle is currently in a low-speed parallel parking state, the current operating status information may include real-time vehicle speed, vehicle attitude, distribution of surrounding obstacles, and parking space scenario information. If the user's edit request is to adjust the parking driving speed parameters, new speed control parameters are read. Starting from the vehicle's current initial parking state, the subsequent driving position and attitude of the vehicle are extrapolated step-by-step according to the driving logic of the parking task, ultimately generating a complete parking candidate trajectory adapted to the new speed parameters.

[0042] For example, the editing request can first be parsed to extract all control parameters adjusted by the user for the current driving task, replacing the original task preset parameters to form a new task control instruction set. Simultaneously, the initial vehicle state data and scene environment data in the current operating status information are locked, using the vehicle's real-time position, speed, and heading attitude as initial boundary conditions for trajectory extrapolation, and scene information such as obstacle distribution, parking space semantics, and road speed limits as constraints for trajectory generation. Based on this, a vehicle motion model adapted to low-speed driving scenarios is used, and forward iterative extrapolation is performed at fixed discrete time steps. The driving position, vehicle attitude, and operating state of the target vehicle under the new control parameters are calculated time-by-time, generating independent trajectory points corresponding to each time series. All time-series trajectory points are systematically integrated and fitted in chronological order to form continuous and complete candidate trajectories for the target vehicle's future driving. Finally, a time-series trajectory dataset containing trajectory points from multiple time periods is output, strictly ensuring that the generated candidate trajectories perfectly match the user's editing intent and the vehicle's real-time operating conditions.

[0043] By implementing step 102, the editing request can be converted into a continuous trajectory description of the vehicle's future motion process, expanding the subsequent analysis object from static editing parameters to the complete motion process of the vehicle. This provides a complete trajectory basis for subsequent continuous collision detection and improves the comprehensiveness of driving task editing safety analysis.

[0044] Step 103: Perform continuous collision detection on the candidate trajectory to obtain collision distance time series data, wherein the collision distance time series data is used to characterize the distance between the trajectory point and the obstacle at each time step.

[0045] In some examples, continuous collision detection is a full-time, high-density geometric safety verification mechanism for candidate trajectories of target vehicles. This detection method performs time-by-time verification of the complete driving period covered by the candidate trajectory, continuously monitoring the changes in the spatial positional relationship between the target vehicle and surrounding obstacles, and achieving full-time capture of dangerous situations such as close approach, critical contact, and spatial interference. It can use the time-synchronized vehicle trajectory and the predicted obstacle trajectory as input, and rely on standardized geometric calculations to complete uninterrupted safety distance determination, which can effectively avoid the risk of missed detection caused by traditional discrete detection and improve the integrity of safety verification in intelligent driving task editing scenarios. For example, when the target vehicle performs a parallel parking task after dynamic parameter editing, continuous collision detection covers the entire parking driving period corresponding to the candidate trajectory, performs vehicle-obstacle distance detection at each time point in the vehicle reversing process, continuously monitors the positional changes of the vehicle and obstacles such as curbs and surrounding stationary vehicles, and fully captures the dynamic process of the vehicle gradually approaching obstacles during driving, achieving uninterrupted safety monitoring.

[0046] Collision distance time-series data is a structured time-series dataset output after continuous collision detection. It is used to characterize the minimum spatial distance between the target vehicle and various obstacles in the scene at each moment of the candidate trajectory. This data uses time as the sole axis and associates the spatial distance values ​​at the corresponding moments. It can intuitively reflect the dynamic change of the safe distance between the vehicle and obstacles during future driving. It is the basic input data for subsequent risk quantification assessment. For example, the target vehicle parking candidate trajectory covers a three-second driving time. After completing the moment-by-moment detection according to a fixed time step, the corresponding collision distance time-series data is generated. This data completely records the minimum spatial distance between the vehicle trajectory point and the nearest curb obstacle at each moment. It clearly presents the dynamic change characteristics of the vehicle's initial safe distance being large and gradually decreasing as the parking driving distance decreases, and completely covers the entire driving time of the candidate trajectory.

[0047] For example, firstly, obstacle prediction trajectory data that perfectly matches the candidate trajectory time interval can be retrieved. The target vehicle candidate trajectory and obstacle prediction trajectory are unified to the same discrete time grid, realizing the temporal dimension alignment of the two types of temporal trajectory data and ensuring the temporal consistency of subsequent distance detection. On this basis, the outline of the target vehicle and the outline of obstacles in the scene are standardized geometrically. Based on a unified time step, the minimum spatial distance calculation is performed on the vehicle trajectory position and obstacle position corresponding to each time node, continuously obtaining the minimum safe distance between the target vehicle and all obstacles at each moment. During the entire detection process, each detection timestamp and the corresponding minimum distance value are continuously recorded. All detection results are arranged in an orderly and structured manner according to the chronological order to form continuous, complete, and temporally corresponding collision distance temporal data.

[0048] By implementing step 103, the geometric proximity between the target vehicle and surrounding obstacles can be continuously reflected throughout the execution of the candidate trajectory, rather than making collision judgments only at a single location or at a single moment. This allows for a more comprehensive identification of potential collision risks during vehicle operation and improves the accuracy and reliability of collision detection results.

[0049] Step 104: Based on the collision distance time series data and the preset error uncertainty parameters, determine the collision risk quantification value for the candidate trajectory.

[0050] The preset error uncertainty parameter may include at least one of the following: positioning error, perception error, control execution error, and model prediction error.

[0051] In some examples, the preset error uncertainty parameters are a standardized set of parameters used to characterize the inherent error characteristics during the entire process of vehicle data acquisition, state perception, command execution, and model inference. They are used to quantify the uncertainty deviation between real driving conditions and theoretical calculation results, and to make up for the problem that single geometric distance detection cannot cover the system error defects. For example, in low-speed automatic parking, multiple types of basic error parameter values ​​are pre-stored, covering the standard error ranges corresponding to positioning, perception, control execution, and model prediction. This can adapt to the uncertainty error quantification needs of mainstream scenarios such as conventional indoor parking lots and open-air parking areas, and provide a unified error benchmark for the risk assessment of parking trajectories.

[0052] Positioning error is the deviation between the vehicle position and attitude information output by the positioning module and the vehicle's actual physical position and attitude information. It is a source of uncertainty in the vehicle's spatial state perception. Positioning error directly affects the accuracy of the vehicle's trajectory point coordinates, leading to deviations between the theoretically calculated obstacle distance and the actual physical distance. It is a significant factor causing misjudgments and omissions in trajectory risk assessment. Error can be statistically calibrated using a High Precision Positioning System (HPPS). By combining positioning data from satellite positioning, inertial navigation, and odometer fusion, and comparing it with high-precision true data, a standard deviation parameter can be statistically obtained and used as a preset positioning error parameter. For example, in a typical parking scenario, the vehicle positioning module is affected by indoor signal obstruction and environmental interference, resulting in a fixed, small positioning deviation. A preset positioning error value of 0.05 meters standard deviation is used to quantify the uncertainty in the vehicle's position perception.

[0053] Perception error is the deviation between the obstacle position, size, and speed information detected by the environmental perception module and the actual physical state information of the obstacle. It is the source of uncertainty in the external environment perception process. Perception error will cause modeling deviations in obstacle outlines and predicted trajectories, making it impossible for theoretically calculated safe distances between vehicles and obstacles to fully match real-world environmental conditions. It can be offline statistically calibrated by the Environment Perception System (EPS). Through iterative statistical analysis of massive amounts of real-vehicle and simulation scenario data, standard parameters for perception error are fitted for different obstacle types, lighting environments, and distance ranges, and then preset and solidified by the system. For example, in a low-speed parking scenario in a parking lot, for common obstacles such as curbs and stationary vehicles, the preset perception error value is 0.08 meters standard deviation, which is used to quantify the perception uncertainty of environmental obstacle detection.

[0054] Control execution error is the deviation between the actual control quantity executed by the vehicle chassis actuator and the theoretical control command issued by the decision module. It is the source of uncertainty in the vehicle motion execution process. Control execution error manifests as the actual output deviation of control parameters such as vehicle speed, steering angle, and acceleration, which can cause a deviation between the vehicle's actual driving trajectory and the candidate trajectory deduced by the algorithm, thus changing the actual safe distance between the vehicle and obstacles. By conducting real-vehicle multi-condition driving tests, the command execution deviation under different vehicle speeds, steering angles, and road surface adhesion conditions can be statistically analyzed, and a standard parameter of control execution error can be fitted and stored as a preset parameter for the system. For example, in low-speed parking conditions, there are slight lags and deviations in vehicle steering execution and speed following. The preset control execution error is a fixed low-speed standard deviation parameter used to quantify the uncertainty of the vehicle control execution process.

[0055] Model prediction error is the deviation between the time-series state results derived from the vehicle kinematics model and obstacle trajectory prediction model and the actual motion state. It is the source of uncertainty in the algorithm model derivation process. Model prediction error originates from the approximate simplification of the vehicle's nonlinear characteristics and environmental dynamic disturbances by the theoretical model, which will cause deviations in the derivation of candidate trajectories and obstacle prediction trajectories, affecting the accuracy of collision detection results. By comparing massive scene simulations with real vehicle trajectories, the deviation distribution between the model derivation results and the actual trajectory can be statistically analyzed, and standard parameters of model prediction error can be fitted to obtain the system preset and solidified. For example, in the parallel parking trajectory derivation process, the simplified vehicle motion model cannot fully cover the impact of small road surface disturbances and tire nonlinear characteristics. Pre-set model prediction error parameters are used to quantify the model approximation deviation in the trajectory derivation process.

[0056] The collision risk quantification value is a standardized, quantifiable, and comparable risk value calculated by integrating time-series distance detection results with multi-dimensional uncertainty error parameters of the system. It is used to objectively characterize the overall collision risk level corresponding to the candidate trajectory after the target vehicle is edited. This value transforms the traditional deterministic geometric distance judgment result into a probabilistic comprehensive risk index, which can accurately distinguish the risk level of different trajectory conditions and provide a quantitative basis for safety decisions for driving task editing requests. It can be based on the time-series collision distance data obtained by continuous detection, and various preset error uncertainty parameters can be superimposed to complete probabilistic inference and risk aggregation, and finally output a single standardized global risk value. For example, after the user increases the parking speed to generate a new candidate trajectory, the corresponding collision risk quantification value is calculated by combining the time-series data of the vehicle obstacle distance throughout the process with various system error parameters. This value can be used to intuitively determine the level of parking collision risk corresponding to this speed editing operation.

[0057] For example, firstly, the system can read the error uncertainty parameters adapted to the current parking scenario from the local preset parameter configuration library. It can selectively retrieve any one or more combinations of parameters from positioning error, perception error, control execution error, and model prediction error, fully covering the uncertainty deviations of the entire link of perception, positioning, control, and model inference. Then, the collision distance time-series data recorded at each moment is matched with various preset error uncertainty parameters. The geometric detection distance at each moment is fused with the corresponding dimension of the system error to correct the working condition deviation problem of a single geometric detection result and accurately quantify the real collision uncertainty corresponding to each trajectory moment. Based on the full-time distance data and error parameters, the system completes the overall risk aggregation calculation, abandoning the traditional black-and-white collision judgment logic. It integrates the safety hazards and system error deviations of the entire trajectory through probabilistic calculation, and finally outputs a unique, quantifiable, and comparable candidate trajectory collision risk quantification value, objectively representing the overall collision risk level corresponding to this driving task editing operation.

[0058] By implementing step 104, the uncertainties that exist in the actual operation of the vehicle can be further considered on the basis of the geometric collision detection results. This allows the collision risk assessment to not only reflect the distance relationship between the trajectory and the obstacle, but also the impact of error factors on the collision risk, thereby improving the degree of consistency between the risk assessment results and the actual operating scenario.

[0059] Step 105: Based on the collision risk quantification value, determine the decision result for the edit request.

[0060] In some examples, the decision result is the final standardized processing conclusion based on the quantified collision risk output for a user-initiated dynamic editing request for a driving task. This is the final output of the entire task editing safety verification process. The result is used to guide the in-vehicle system on whether to execute the user's editing operation, thereby achieving safety control over dynamic editing tasks and adapting to the runtime task scheduling requirements of low-speed, high-safety scenarios such as intelligent driving and parking. For example, for a user's editing request to adjust the parking speed, the corresponding editing processing conclusion is output based on the calculated collision risk quantification value. When the risk value is within the safe range, the result allows the editing operation to be executed; when the risk value exceeds the safety standard, the result prohibits the editing operation, thereby achieving safety control over parking task editing behavior.

[0061] For example, the preset safety decision-making strategy and risk assessment benchmark corresponding to the current intelligent driving scenario can be retrieved. The quantitative risk value, which integrates the multi-dimensional uncertainty error of vehicle positioning, environmental perception, control execution and model prediction, is used as the judgment basis, thus abandoning the limitations of single geometric collision judgment. By accurately comparing the collision risk quantification value with the preset safety threshold standard, the overall collision risk level brought about by this driving task editing request is objectively evaluated, and the corresponding standardized decision result is generated in strict accordance with the established safety control strategy.

[0062] By implementing step 105, the execution of vehicle driving task editing can be based on the results of risk quantification analysis, enabling safe decision-making on editing requests and avoiding judgments based solely on editing content or fixed rules, thereby improving the safety and decision reliability of the dynamic editing process of vehicle driving tasks.

[0063] In summary, this application's embodiments obtain the edit request for the target vehicle's current driving task and generate candidate trajectories by combining the vehicle's current operating status information. This allows the dynamic modification of the driving task to move beyond simple judgments based on preset rules. Instead, it performs safety analysis based on the current vehicle status, task change requirements, and the actual operating environment. The candidate trajectories are generated after the edit request is associated with the vehicle's real-time status, making the subsequent risk assessment process more closely aligned with the vehicle's actual operating conditions, thereby improving the adaptability and safety of the vehicle's driving task editing process. Furthermore, continuous collision detection is performed on the candidate trajectories to obtain collision distance time-series data characterizing the geometric proximity between the vehicle trajectory and obstacles at each moment. Compared to collision judgments based on only a single location or moment, this method provides a more comprehensive assessment. The method continuously analyzes the distance changes between the vehicle and obstacles along the candidate trajectory execution process, thus more comprehensively reflecting the potential collision risks during trajectory execution and reducing the risk omission caused by ignoring intermediate motion states. Based on collision distance time-series data and preset error uncertainty parameters including positioning error, perception error, control execution error, and model prediction error, the collision risk of the candidate trajectory is quantified. Since this risk quantification process introduces uncertainties existing in vehicle operation, it avoids the limitations of relying solely on deterministic geometric distances for safety judgment, enabling the obtained collision risk quantification value to more accurately reflect the risk level under actual operating conditions and improve the reliability of risk assessment results. In summary, the vehicle driving task editing method provided in this application, through a processing flow from editing request acquisition, candidate trajectory generation, continuous collision detection, uncertainty risk assessment to decision output, can realize safety judgment for editing requests during dynamic driving task adjustment, improving the safety and controllability of vehicle driving task editing.

[0064] In some embodiments, step 104 may include: determining the equivalent distance variance corresponding to each moment in the collision distance time series data based on a preset error uncertainty parameter; determining the instantaneous collision probability at each moment using a preset normal cumulative distribution function based on the collision distance time series data and the equivalent distance variance, thereby obtaining a collision probability sequence; and determining the collision risk quantification value based on the collision probability sequence.

[0065] In some examples, the equivalent distance variance maps multi-source, multi-dimensional uncertainty errors into a unified overall fluctuation variance index of the distance dimension between the vehicle and the obstacle. It characterizes the degree of discrete deviation between the theoretical detection distance and the actual physical distance at each moment. This parameter integrates all error disturbances from perception, localization, control, and model inference, quantifying heterogeneous errors of multiple types and dimensions into a statistical variance of the distance dimension. It is a fundamental parameter for transforming deterministic distance detection into probabilistic risk detection. The equivalent distance variance can be obtained by variance fusion calculation based on preset error uncertainty parameters and the independent or correlated characteristics of each error source, generating synchronously with the collision distance time-series data moment by moment. For example, in low-speed automatic parking, integrating localization error and perception error parameters, the equivalent distance variance at each moment is obtained through sum of squares fusion calculation. This variance value accurately reflects the fluctuation range of the obstacle detection distance affected by system errors at the current moment.

[0066] The preset normal cumulative distribution function is a pre-defined standard normal distribution probability calculation function. It is used to convert real-time distance margin and equivalent distance variance into corresponding instantaneous collision probabilities based on the normal distribution characteristics of distance deviation. All kinds of sensing errors, execution errors, and model errors conform to the zero-mean normal distribution characteristics. This function can accurately adapt to the statistical distribution law of system errors and realize the standardized mapping from distance safety margin to collision risk probability. For example, the real-time minimum distance to the obstacle and the equivalent distance variance obtained in the parking scenario can be input into the preset normal cumulative distribution function. Through standardized probability mapping calculation, the probability value of the vehicle collision at the current moment is output, realizing the transformation of geometric distance data into risk probability data.

[0067] Instantaneous collision probability is the real-time probability that a target vehicle will collide with surrounding obstacles at a single discrete moment due to various uncertainties and error disturbances. This indicator is calculated independently for each time point of the trajectory, accurately characterizing the local collision risk at a single moment. It overcomes the limitation of fixed geometric distance determination in quantifying error risk, achieving refined quantification of risk at a single moment. The collision distance values ​​at each moment and the corresponding equivalent distance variance can be input into a preset normal cumulative distribution function. Through standardized probability calculation, the independent instantaneous collision probability at each moment is obtained, forming a single-moment risk quantification result. For example, at a certain sampling moment of the vehicle's parking trajectory, if the measured minimum distance to the obstacle is within the critical safety range, after superimposing system positioning and perception errors, the corresponding instantaneous collision probability is obtained through function calculation. This probability value can intuitively reflect the likelihood of the vehicle colliding due to system error disturbances at that moment.

[0068] A collision probability sequence is a time-series probability dataset formed by arranging the instantaneous collision probabilities corresponding to all discrete moments of a candidate trajectory in ascending order over time. This sequence completely covers the entire process of the vehicle's driving trajectory after editing, and can continuously characterize the dynamic change trend of collision risk throughout the vehicle's future driving process. It is the time-series data foundation for aggregating global risk and generating the final risk quantification value. For example, after sampling a parking candidate trajectory with a duration of three seconds according to a fixed time step, a corresponding number of instantaneous collision probability values ​​are generated. All probability values ​​are arranged in chronological order to form a collision probability sequence, which fully presents the dynamic change process of collision risk rising, stabilizing, or decreasing throughout the entire parking process.

[0069] The process of determining the collision risk quantification value based on the collision probability sequence involves reading the complete collision probability sequence, extracting the risk feature information within the sequence for integration and calculation, and then normalizing it to output a unique global collision risk quantification value, thus completing the transformation from time-series risk data to a global risk indicator. For example, after obtaining the collision probability sequence for the entire parking process, the high and low risk distribution at each moment within the sequence is integrated to obtain the overall collision risk quantification value corresponding to this parking speed editing operation, which intuitively reflects the comprehensive risk level of this task editing behavior.

[0070] For example, firstly, all preset error uncertainty parameters corresponding to the current working condition can be retrieved. Based on the distribution characteristics of various error sources, the positioning error, perception error, control execution error, and model prediction error are fused and calculated at each time step to obtain the equivalent distance variance corresponding to each time-series node, completing the transformation of multi-source errors into a unified uncertainty in the distance dimension. On this basis, the minimum distance value in the collision distance time-series data is read at each time step. The real-time distance value and the equivalent distance variance at the corresponding time step are input into a preset normal cumulative distribution function. Using the standard probability mapping characteristics of the function, the instantaneous collision probability at each discrete time step is calculated one by one. The instantaneous collision probabilities obtained at all times are arranged in chronological order and encapsulated in a structured manner to generate a collision probability sequence covering the entire candidate trajectory, fully restoring the dynamic temporal change law of collision risk during vehicle driving. Finally, the complete collision probability sequence is subjected to global risk aggregation processing, integrating temporal risk features to calculate a collision risk quantification value that can characterize the overall danger level of this editing request.

[0071] Through the implementation of the above embodiments, the equivalent distance variance corresponding to each moment of the collision distance time series data is first determined according to the preset error uncertainty parameter. Then, the instantaneous collision probability at each moment is determined by combining the collision distance time series data with the preset normal cumulative distribution function. Furthermore, a collision probability sequence is formed to determine the collision risk quantification value, so that the geometric distance can be converted into a statistically significant collision probability. This avoids risk judgment based solely on distance size and can improve the adaptability of the risk assessment results to uncertainties such as positioning error, perception error, and control error during the vehicle driving task editing process.

[0072] In some embodiments, the aforementioned determination of the collision risk quantification value based on the collision probability sequence may include: obtaining the upper limit of the instantaneous collision probability in the collision probability sequence as a first risk metric; determining a second risk metric value based on the instantaneous collision probability at each moment in the collision probability sequence; and performing a weighted summation of the first risk metric value and the second risk metric value to obtain the collision risk quantification value.

[0073] In some examples, the upper limit of instantaneous collision probability is the maximum value among all instantaneous collision probability values ​​contained in the complete collision probability sequence. It is used to characterize the most extreme single collision risk level that occurs during the entire driving process of the vehicle candidate trajectory. This value can accurately capture sudden, instantaneous high-risk collision hazards in the trajectory time sequence, effectively identify short-term high-risk conditions, and make up for the shortcomings of the average risk assessment method in covering extreme dangerous points. For example, the candidate trajectory after the vehicle performs a parking speed adjustment contains 20 time sequence sampling points, each corresponding to a different instantaneous collision probability value. The probability value is highest at a critical moment when a nearby obstacle is close. This extreme value is the upper limit of instantaneous collision probability for this trajectory risk assessment, representing the highest instantaneous collision risk that occurs in this parking editing task. The first risk metric is a quantitative indicator used to characterize the local extreme risk characteristics of the candidate trajectory. It can be directly assigned from the upper limit of instantaneous collision probability. This metric focuses on single-point extreme high-risk conditions during vehicle driving, emphasizing the fatal impact of sudden instantaneous collision risks on driving safety, prioritizing the accurate identification of extreme dangerous scenarios, and reflecting the extreme risk fallback logic of safety assessment.

[0074] The second risk metric is a cumulative risk quantification index obtained by solving the full-time-series risk distribution characteristics based on the complete collision probability sequence. It is used to characterize the overall and continuous collision risk level of the vehicle candidate trajectory. This metric covers the cumulative risk effect of all driving periods of the trajectory, and can reflect the superimposed hazards of long-term, low-probability continuous risks. It makes up for the shortcomings of a single extreme risk index that cannot assess the steady-state risk throughout the entire process, and forms a complementary assessment dimension with the first risk metric. It can integrate the instantaneous collision probabilities of all moments in the collision probability sequence, complete the full-domain risk aggregation through preset cumulative calculation logic, and obtain the second risk metric that can characterize the overall risk throughout the entire process after normalization. For example, if there are low-amplitude collision probabilities at multiple time-series nodes during the entire parking process, and the risk at a single moment does not reach the high-risk standard, but there is a continuous safety hazard throughout the entire process, the corresponding second risk metric is obtained by accumulating and integrating all time-series risks, and is used to quantify the overall continuous risk level of this parking editing task.

[0075] The process of obtaining a collision risk quantification value by weighted summation of the first and second risk metrics involves pre-setting fixed weight coefficients to perform weight adaptation and fusion calculations on the first risk metric, which represents the extreme risk at a single point, and the second risk metric, which represents the cumulative risk throughout the entire process. This unifies the two different risk indicators into a single global risk value, taking into account both extreme instantaneous risks and continuous risks throughout the entire process, avoiding the one-sidedness of a single assessment dimension. This ensures that the final collision risk quantification value can comprehensively, objectively, and evenly reflect the true danger level after the driving task is edited, providing an accurate comprehensive judgment basis for subsequent editing request decisions.

[0076] For example, firstly, the instantaneous collision probability values ​​corresponding to all time nodes within the collision probability sequence can be traversed and compared one by one to filter out the maximum probability value in the sequence. This extreme value is determined as the upper limit of the instantaneous collision probability and directly assigned as the first risk metric value, thus completing the extraction of the extreme risk of a single point in the trajectory. At the same time, a global integration calculation is performed on the instantaneous collision probability at all times in the collision probability sequence. Combined with the temporal risk distribution characteristics, the cumulative risk is solved to obtain a second risk metric value that can characterize the overall safety hazard of the entire trajectory. After obtaining the two independent risk metrics values, preset and engineering-calibrated weight coefficients are retrieved to perform weighted adaptation calculations on the first and second risk metrics values ​​respectively, balancing the evaluation weights of extreme sudden risks and continuous cumulative risks. Finally, the two weighted results are summed and aggregated to output a unique and standardized global collision risk quantification value.

[0077] By implementing the above embodiments, the upper limit of instantaneous collision probability in the collision probability sequence is obtained as the first risk metric, and the second risk metric is determined based on the entire collision probability sequence. The two types of risk metrics are then weighted and fused, so that the risk assessment can reflect both the peak risk at a certain dangerous moment during the vehicle driving task editing process and the overall risk level during the entire trajectory execution period. This avoids the risk judgment being one-sided due to a single risk indicator and improves the reliability of the editing decision results.

[0078] In some embodiments, determining the second risk metric based on the instantaneous collision probability at each moment in the collision probability sequence may include: performing time-weighted summation on the instantaneous collision probability at each moment in the collision probability sequence to obtain a cumulative risk metric; and performing a preset exponential decay function operation on the cumulative risk metric to obtain the second risk metric.

[0079] In some examples, the cumulative risk metric is the total risk value obtained by summing the instantaneous collision probabilities at all times in the collision probability sequence. It is used to characterize the total original cumulative risk of the candidate trajectory of the target vehicle over the entire driving period, combined with the difference in temporal importance. A preset time weight parameter can be called to perform a weighted product operation on the instantaneous collision probability corresponding to each time moment in the collision probability sequence. After traversing all time-series nodes, the summation is completed, and finally a unique cumulative risk metric is obtained. For example, the candidate trajectory of a vehicle parking at low speed covers multiple consecutive time-series nodes. Different time-series nodes correspond to different time weights and instantaneous collision probabilities. The weighted probabilities of all time-series nodes are summed to obtain the cumulative risk metric corresponding to this parking trajectory, which intuitively reflects the total superposition of risks throughout the process, including the comprehensive superposition effect of short-term high-risk risks and long-term low-risk risks.

[0080] The preset exponential decay function operation is a pre-calibrated, non-linear mapping operation logic embedded in the algorithm's underlying layer. It is specifically used to perform non-linear correction processing on the risk values ​​obtained by linear accumulation. This operation can avoid the problem of inflated risk caused by linear weighted summation, suppress the risk distortion caused by the continuous superposition of a large number of low-probability risks, enhance the differentiation of high effective risks, and ensure that the final risk value is highly matched with the actual driving hazard level, thus adapting to the risk quantification needs of intelligent driving dynamic editing scenarios. For example, when there are a large number of continuous low-amplitude instantaneous collision probabilities in the parking trajectory, linear accumulation will cause the risk value to be inflated and distorted. By correcting the accumulated risk metric value through the preset exponential decay function operation, the interference of invalid low-risk superposition can be weakened, the true and effective high-risk characteristics can be preserved, and the risk value can be accurately calibrated.

[0081] For example, firstly, a preset time weight parameter adapted to the current driving scenario can be retrieved, and all time-series nodes included in the collision probability sequence can be traversed. The instantaneous collision probability corresponding to each moment is calculated with time weight, and the weight correction is completed by combining the risk importance at different moments. The weighted calculation results of all time-series nodes are accumulated and summed to obtain a cumulative risk metric that fully covers the entire driving process of the candidate trajectory, accurately preserving the superposition characteristics of time-series risks and differences in the time dimension. On this basis, the uncorrected cumulative risk metric is used as an input parameter, and a built-in preset exponential decay function is called to complete the nonlinear decay calculation. This uniformly corrects the problems of inflated risk and distortion caused by low probability superposition caused by linear accumulation, weakens the interference of ineffective low-risk accumulation, and highlights the effective high-risk characteristics. After exponential decay mapping processing, a standardized and high-precision second risk metric is finally output. This metric can stably characterize the global continuous collision risk level of the candidate trajectory after vehicle editing.

[0082] By implementing the above embodiments, the instantaneous collision probabilities at each moment in the collision probability sequence are first summed using a time-weighted method, and then the cumulative risk metric is calculated using an exponential decay function to obtain a second risk metric. This allows the risk assessment to not only comprehensively consider the risk accumulation throughout the entire trajectory process, but also avoid the problem of the risk value increasing infinitely with simple accumulation over time. As a result, the obtained risk metric is more consistent with the risk evolution law in the continuous driving task editing process of the vehicle, and improves the stability of the risk quantification results.

[0083] In some embodiments, the aforementioned vehicle driving task editing method may further include: obtaining scene parameters of the target vehicle, wherein the scene parameters may include at least one of parking space type, ambient light level, and road surface adhesion coefficient; determining a target time weight sequence that matches the scene parameters based on a preset correspondence between scene parameters and time weight levels, wherein each moment in the target time weight sequence corresponds to a time weight value; performing time-weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric, which may include: performing weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence based on the target time weight sequence to obtain a cumulative risk metric.

[0084] In some examples, scene parameters are a set of structured environmental parameters used to characterize the current intelligent driving operating environment and working conditions of the target vehicle. They can objectively reflect the environmental complexity, lighting perception conditions, and road surface adhesion conditions of the parking operation, and serve as the input basis for the adaptive adjustment of the risk assessment strategy. Scene parameters can be output collaboratively by the environmental perception system, the road surface state estimation module, and the high-precision map module. They can identify and output the environmental feature parameters of the current vehicle's working conditions in real time, ensuring that the parameters are aligned with the scene state time sequence at the time of the edit request. Scene parameters can include any one or more combinations of three types of parameters: parking space type, ambient light level, and road surface adhesion coefficient.

[0085] Parking space type is a scenario classification parameter used to characterize the geometric shape, layout, and operational difficulty of a target parking space. Different parking space types correspond to different vehicle motion constraints and obstacle distribution characteristics, directly affecting the risk distribution pattern and safety margin of the parking trajectory. Parking space type is the basis for classifying parking operation conditions and can effectively distinguish the risk sensitivity of different parking scenarios. It can read the semantic information of parking spaces from high-precision maps and combine it with the parking space outline, marking shape, and surrounding layout features detected by the visual perception module to complete the accurate classification output of parking space types. For example, common parking space types include parallel parking spaces, perpendicular parking spaces, angled parking spaces, and narrow non-standard parking spaces. Different parking space types correspond to different trajectory curvature and obstacle avoidance difficulty, adapting to differentiated time weight allocation rules.

[0086] Ambient lighting level is an environmental perception quality parameter used to characterize the intensity and uniformity of light in the current driving environment. It directly determines the detection accuracy, obstacle recognition confidence, and contour extraction accuracy of the visual perception module. Low light, backlight, and shadow occlusion conditions increase perception uncertainty, requiring adjustments to the temporal risk weights and strengthening risk management during high-risk periods. Ambient brightness data can be collected by forward-looking and surround-view cameras, combined with image exposure parameters and image signal-to-noise ratio indicators, to output the corresponding lighting level according to preset grading standards. For example, it can be divided into three typical levels: strong sunlight level, normal indoor lighting level, and dim light level. Perception errors fluctuate more in low-light environments, which increases the risk weight of near-field temporal nodes and strengthens the instantaneous risk identification capability.

[0087] The road surface adhesion coefficient is a physical parameter that characterizes the friction between a vehicle's tires and the road surface. It directly determines the vehicle's maximum braking capacity, lateral stability limit, and trajectory tracking accuracy. Under low-adhesion road surface conditions, vehicle braking lag and trajectory deviation risk increase, and the risk temporal characteristics are significantly different from those of standard dry road surfaces. It is necessary to adaptively adjust the weights to match the actual safety conditions. The system can combine the road surface dryness / wetness and material characteristics identified by the perception module with the deceleration response data of the vehicle during braking to fit and output real-time road surface adhesion coefficient values ​​online and classify the corresponding adhesion levels. For example, road surface adhesion conditions include three typical conditions: dry asphalt high-adhesion road surface, wet conventional adhesion road surface, and rain / snow low-adhesion road surface. Under low-adhesion road surface conditions, the system will increase the risk weight throughout the entire process, amplify the cumulative impact of small collision probability, and avoid safety hazards caused by trajectory deviation.

[0088] The pre-defined correspondence between scenario parameters and time weight levels is a set of structured mapping rules calibrated and solidified offline in the vehicle. This establishes a one-to-one correspondence between different scenario conditions and time-series weight configuration levels. This correspondence can dynamically adjust the weight ratio of instantaneous collision probability at different times based on scenario complexity, perception reliability, and vehicle stability limits. This enables scenario-adaptive risk assessment strategies and addresses the technical shortcomings of fixed weights being unable to adapt to multiple conditions. For example, the pre-defined rule is that low-light, low-adhesion, narrow parking scenarios correspond to high-sensitivity weight levels, increasing the weight ratio of close-range time-series nodes. Standard lighting, dry road surfaces, and regular parking space scenarios correspond to standard weight levels. A balanced time-series weight allocation strategy is adopted to adapt to the risk assessment needs of different scenarios.

[0089] The target time-weighted sequence is a set of time-series weight values ​​that are precisely matched with the current vehicle scenario parameters and adapted to the current parking conditions. The sequence contains an independent time-weight value that corresponds one-to-one with each discrete sampling moment of the candidate trajectory. This sequence is used to differentiate the risk importance at different driving moments, which can fit the safety characteristics of the current scenario and achieve refined and scenario-based cumulative risk calculation. After reading the current scenario parameters, the preset scenario weight correspondence can be retrieved to match the corresponding weight level. According to the time step distribution of the candidate trajectory, standardized time-weight values ​​corresponding to each moment are generated and orderly packaged to form a complete target time-weighted sequence. For example, in a parking scenario with a wet road surface and low light, a high-sensitivity weight level is matched. The generated target time-weighted sequence shows a distribution characteristic of higher weight for close-range moments and lower weight for distant moments, which focuses on strengthening the risk assessment weight of the later stage of parking when approaching the parking space.

[0090] The process of obtaining a cumulative risk metric by weighted summation of the instantaneous collision probabilities at each moment in the collision probability sequence based on the target time weight sequence is as follows: the target time weight sequence obtained by adaptive matching of the current scenario is used to differentiate and correct the instantaneous collision probabilities at each moment in the collision probability sequence, and then the scenario-adaptive cumulative risk metric is obtained by summing the probabilities across the entire domain. This process can dynamically adjust the contribution ratio of risk at each moment according to the risk characteristics of the scenario, so that the cumulative risk result is close to the danger level of the real working conditions, which greatly improves the accuracy and robustness of risk quantification in complex scenarios.

[0091] For example, the onboard perception module and road surface estimation module can collect scene parameters of the target vehicle's current operating condition in real time, accurately obtain the quantitative parameters and classification results of the current parking space type, ambient light level, and road surface adhesion coefficient, and completely construct the scene feature dataset corresponding to the current driving task; retrieve the correspondence between the scene parameters and time weight levels stored locally, match and retrieve the real-time collected scene parameters with the offline calibrated rule levels, determine the optimal weight level suitable for the current parking condition, and generate corresponding standardized weight values ​​at each time step according to the discrete time sampling distribution of the candidate trajectory, and encapsulate them in an orderly manner to form a complete target time weight sequence; on this basis, the target time weight sequence and the collision probability sequence are aligned in the time dimension to achieve accurate matching of the instantaneous collision probability and the corresponding time weight at each trajectory sampling time, complete the probability weighted product operation at each time step, perform scene-differentiated correction on the risk contribution at different times, and finally sum all time-series weighted results across the entire domain to generate a cumulative risk metric value after scene adaptive optimization.

[0092] By implementing the above embodiments, the scene parameters of the target vehicle are obtained, and the corresponding target time weight sequence is matched according to different scene parameters. Then, the collision probability sequence is time-weighted using the target time weight sequence, so that the risk assessment of vehicle driving task editing can dynamically adjust the importance of risk at each moment for different driving scenarios such as parking space type, ambient light, and road surface adhesion conditions. This effectively solves the problem of inaccurate risk assessment under special scenarios such as complex lighting, low adhesion, and narrow parking spaces with fixed weights. It can improve the adaptability of the risk assessment model to different driving environments and make editing decisions more in line with the actual vehicle operation needs.

[0093] In some embodiments, the aforementioned determination of the equivalent distance variance corresponding to each moment in the collision distance time series data based on preset error uncertainty parameters may include: if the positioning error, perception error, control execution error, and model prediction error are independent of each other, then the sum of squares of the preset standard deviations corresponding to each error is calculated to obtain the equivalent distance variance corresponding to each moment; if at least two of the positioning error, perception error, control execution error, and model prediction error are correlated, then the preset covariance matrix corresponding to each error source is obtained, and the joint variance of each error source is calculated based on the preset covariance matrix to obtain the equivalent distance variance corresponding to each moment.

[0094] In some examples, when various errors are independent of each other, fluctuations in a single error will not cause synchronous shifts in other errors, and the variances of various errors are additive. The overall dispersion of multi-dimensional errors can be accurately aggregated through the sum of squares operation, and the final output is the equivalent distance variance at each time moment that can be used for probability risk calculation. For example, in an open-air parking scenario, the satellite positioning signal is stable and the environmental interference is singular. The positioning error, perception error, control execution error and model prediction error are not coupled with each other. The preset standard deviation values ​​of the four types of errors are read separately, and the fusion calculation is completed through the sum of squares formula to obtain the unified equivalent distance variance at the current time, which is used to solve the normal distribution of the instantaneous collision probability in the future.

[0095] The pre-defined covariance matrix is ​​a standardized second-order statistical parameter matrix used to characterize the correlation and coupling characteristics between multiple error sources. It quantifies the positive, negative, or no correlation between different errors. This matrix can comprehensively describe the interconnected fluctuation patterns of multi-source errors under real-world vehicle conditions, compensating for the inability of independent error calculation models to adapt to coupled conditions. It is a fundamental parameter for solving the joint variance in error-correlated scenarios. When at least two types of errors—positioning error, perception error, control execution error, and model prediction error—exhibit interconnected changes, the pre-defined covariance matrix can be retrieved to calculate the overall joint variance of multiple error sources, thereby obtaining the equivalent distance variance at each time step. For example, in low-light parking conditions, perception errors fluctuate due to ambient light, while model prediction trajectory deviations change synchronously with perception data offsets. These two types of errors are strongly correlated. By calling a pre-defined covariance matrix adapted to low-light scenarios, the joint variance of coupled errors can be solved, yielding an equivalent distance variance that closely matches the actual error disturbance level, ensuring the accuracy of subsequent collision probability calculations.

[0096] For example, the real-time disturbance characteristics of positioning error, perception error, control execution error, and model prediction error under the current operating conditions can be read first to determine whether there are any linkage or correlation characteristics among the four types of error sources. If it is determined that the various types of errors are completely independent of each other, the preset standard deviation parameters corresponding to each error are directly retrieved, and the standard deviation of each type is squared and summed. The equivalent distance variance at the current moment is obtained by solving the independent error sum of squares algorithm. If it is determined that at least two types of errors have synchronous fluctuations and mutual influence, the preset covariance matrix matched to the current scene is retrieved. The coupling relationship between each error is quantified based on the covariance matrix, and the multi-dimensional joint variance is calculated by combining the basic statistical parameters of each error source. Finally, the equivalent distance variance that adapts to the error coupling characteristics is obtained.

[0097] For example, after solving for the equivalent distance variance at each time step, the time-by-time minimum distance output by S5 can be combined. The collision probability time-series deduction is carried out, and the complete probability calculation logic is as follows: First, we integrate all uncertainty error sources across the entire chain, including obstacle lateral / longitudinal position deviation, vehicle positioning error, perception and detection error, chassis control execution error, and motion model prediction error. The typical engineering calibration value is the positioning standard deviation. Perceived standard deviation Define the actual vehicle obstacle spacing. ,in for The total equivalent distance error term at any given time follows a zero-mean normal distribution. , for The variance of the total equivalent distance at each moment. The formula for calculating the instantaneous collision probability is: (1) In the formula, Representing the Instantaneous collision probability at discrete moments; This represents the theoretical geometric minimum distance output by the S5 continuous collision detection system. for Standard deviation of total equivalent distance at any given moment; Let be the standard normal cumulative distribution function; if the geometric distance This indicates that the trajectory has undergone geometric intrusion, directly affecting... Assign a value of 1 or a preset probability upper limit.

[0098] Regarding the total equivalent distance variance The previous text distinguished between two solution branches: independent errors and coupled errors. When the four types of errors—localization, sensing, control, and modeling—are independent, the sum of squares formula is used. (2) In the formula, for Standard deviation of location / mileage equivalent distance; for Standard deviation of obstacle perception ranging at any given time; To control the equivalent distance standard deviation of tracking bias; This represents the standard deviation of the equivalent error for the vehicle / obstacle prediction model; if multiple error classes are coupled and correlated, then the covariance matrix is ​​replaced. Calculate the joint variance to ensure that the error distribution closely matches the actual vehicle operating conditions.

[0099] Obtain the full-time collision probability sequence Then, a weighted fusion formula is used to calculate the overall global risk score. : (3) In the formula, The global quantitative risk score corresponding to the edit request; The peak risk weighting coefficient is set at 0.6 to 0.8 for engineering calibration. The maximum instantaneous collision probability across the entire time series represents the risk of the most dangerous single point on the trajectory. For the first The time weight can be a uniform fixed weight or a weight that increases over time; the exponential term... Used to suppress the artificially inflated risk caused by the continuous accumulation of a large number of low-probability risks; weight sequence The calibration needs to be completed offline through scene playback and simulation datasets with the goal of optimizing the false alarm rate, false alarm rate and overall cost. It also supports dynamic updates and adaptation based on parking space type, ambient lighting and road surface adhesion coefficient.

[0100] The calculated overall risk score With preset high-risk threshold (Typical calibration value 0.65) Comparison, if Mark this edit as a high-risk operation; after the calculation is completed, output all high-risk time intervals, corresponding obstacle numbers, and main sources of error uncertainty to provide quantitative judgment basis for the S7 arbitration module.

[0101] When sufficient computing power is available, a Monte Carlo sampling method with 500 or more iterations can be used for high-precision solution. For low-risk time slices with limited computing power, the above-mentioned normal analytical approximation can be used directly, balancing computational real-time performance and risk assessment accuracy through mixed precision.

[0102] By implementing the above embodiments, different situations are distinguished between independent and correlated error sources. The equivalent distance variance is calculated using the sum of squared standard deviations or the covariance matrix, respectively. This allows the uncertainties from different sources to be fused in a way that better reflects actual statistical characteristics. It avoids the risk estimation bias caused by simple superposition when there is correlation between errors, thereby further improving the accuracy of collision risk quantification results during vehicle driving task editing. Through the dual-branch adaptive calculation logic, all vehicle operating conditions of independent and coupled errors are fully covered, so that the equivalent distance variance output at each moment can accurately match the real system error distribution. This solves the problem of inaccurate variance estimation and large risk quantification bias of a single independent error model in complex scenarios, and provides a reliable statistical parameter basis for subsequent high-precision instantaneous collision probability sequence solutions based on the normal cumulative distribution function.

[0103] In some embodiments, step 102 may include: performing hard safety constraint screening on the edit request based on the current vehicle speed and obstacle information in the current operating status information to obtain a safety screening result; if the safety screening result is satisfactory, then performing dynamic reachability analysis based on the current vehicle state in the current operating status information to generate a candidate trajectory.

[0104] In some examples, the current vehicle speed is the real-time longitudinal speed of the target vehicle at the moment the edit request is triggered. It characterizes the instantaneous speed of the vehicle's movement and is a fundamental dynamic parameter for calculating braking distance, determining safety margin, and conducting hard safety constraint screening. Obstacle information is a set of structured environmental obstacle data output by the intelligent driving environment perception system. It is used to fully describe the distribution and spatial relationship of obstacles around the target vehicle. It is an environmental input for judging the vehicle's driving safety margin and identifying collision risks. Obstacle information can objectively reflect the density of obstacles and the distribution of close-range hazards in the current driving scenario. For example, obstacle information in parking situations includes the spatial position and closest relative distance of the roadside curb and surrounding stationary parked vehicles. Based on this type of information, the vehicle's remaining safe parking margin is calculated as a criterion for hard safety screening.

[0105] Hard safety constraint screening is a lightweight, rigid safety verification mechanism executed on the first round of user dynamic editing requests. It belongs to the underlying safety interception logic that cannot be broken. This screening mechanism relies on physical hard constraints such as vehicle braking performance, real-time vehicle speed, and obstacle distance to quickly identify editing operations that obviously violate the bottom line of driving safety. Without performing complex trajectory simulation and probability calculation, it intercepts editing requests with immediate collision risks in advance. For example, when a user requests a significant increase in parking speed, causing the theoretical braking distance to exceed the current remaining obstacle safety distance, hard safety constraint screening can directly determine that the editing operation is in violation, intercept unsafe parameter modifications in advance, and avoid subsequent invalid calculations and potential safety risks.

[0106] The safety screening result is a standardized binary judgment conclusion output after the hard safety constraint screening process is completed. It is used to identify whether the current user's editing request meets the basic physical safety baseline. The calculated safe parking distance and the real-time obstacle safety margin can be compared with the preset safety threshold, and two standardized screening conclusions, namely pass or fail, can be output. For example, if the theoretical parking distance corresponding to the editing request is less than the current remaining obstacle safety distance, the safety screening result indicates pass, and the process is allowed to proceed to the subsequent trajectory generation process. If the editing request causes insufficient safety margin, the safety screening result indicates fail, and the subsequent complex calculation process is directly terminated.

[0107] Dynamic reachability analysis is a deductive process based on the inherent dynamic limits of a vehicle to determine whether the vehicle's motion state corresponding to the edited control parameters is physically feasible. This analysis method uses the vehicle's real-time state as the initial boundary and combines inherent physical constraints such as the vehicle's maximum braking capacity and speed following limit to determine the continuous state range that the target vehicle can actually reach under the action of the edited parameters. It eliminates invalid control commands that are theoretically feasible but physically unexecutable, ensuring that the generated trajectory closely matches the vehicle's actual motion capabilities. For example, after a user edits the parking speed parameters, a dynamic reachability analysis is performed to determine whether the vehicle's acceleration / deceleration range and attitude change range under the speed parameters are within the vehicle's physical limits, verifying that the edited task parameters are actually executable.

[0108] For example, current vehicle speed and obstacle information can be extracted in real time from the current operating status information. These two types of information, along with the user's editing request parameters, are input into a fast safety filter. The safe stopping distance and real-time safety margin under the editing condition are calculated based on the vehicle's braking physical characteristics. A rigid safety judgment is made by comparing this result with a preset safety threshold, and a standardized safety screening result is output. If the safety screening result fails, the subsequent trajectory generation process is directly terminated, achieving rapid interception of unsafe editing requests and saving onboard computing resources. If the safety screening result passes, the current editing request is determined to meet the basic physical safety baseline, and subsequent dynamic calibration is then triggered. The verification and trajectory generation process involves using the vehicle's real-time position, speed, attitude, and dynamic limit parameters from the current operating status information as initial boundary conditions. A refined dynamic reachability analysis is conducted using an reachability analyzer to deduce the envelope of all reachable continuous motion states of the vehicle under the constraints of the edited parameters. Nominal control sequences that conform to the vehicle's dynamic constraints are then selected. Forward temporal simulation is performed based on the compliant control sequences to deduce the vehicle's future driving position and attitude time-by-time, ultimately generating a complete, continuous, physically feasible, and user-editable candidate trajectory for the target vehicle. This provides compliant and effective trajectory input for subsequent continuous collision detection and probabilistic risk quantification.

[0109] Through the implementation of the above embodiments, before generating candidate trajectories, the editing requests are first screened for hard safety constraints based on the current vehicle speed and obstacle information. Only after the safety screening is passed is dynamic reachability analysis performed to generate candidate trajectories. This allows driving task editing requests that obviously violate safety requirements to be filtered out in advance, reducing the amount of complex calculations in the future. At the same time, it ensures that the editing requests entering the trajectory generation stage have basic safety, thereby improving the processing efficiency and safety of vehicle driving task editing.

[0110] In some embodiments, the aforementioned hard safety constraint screening of the edit request based on the current vehicle speed and obstacle information in the current operating status information to obtain a safety screening result may include: determining a safe stopping distance based on a preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request; determining the difference between the current nearest obstacle distance in the obstacle information and the safe stopping distance as a safety margin; and generating a safety screening result indicating failure if the safety margin is less than a preset safety threshold.

[0111] In some examples, the preset maximum braking deceleration is the limit braking deceleration capability that the target vehicle can stably output under normal driving conditions. It is a fixed physical parameter of the vehicle's braking performance, used to quantify the vehicle's limit capability for the shortest braking distance. This parameter comprehensively considers the hardware performance of the vehicle's chassis braking system, normal road driving conditions, and driving safety regulations. It is a fixed parameter for calculating safe parking distance and can ensure that the hard safety constraint screening has a conservative and reliable safety fallback characteristic. For example, for low-speed automatic parking driving scenarios, the vehicle's braking system is calibrated to obtain a fixed maximum braking deceleration parameter. This parameter is adapted to the low-speed driving characteristics of parking and can accurately calculate the vehicle's limit braking capability under parking conditions, providing a stable benchmark for calculating safe parking distance.

[0112] The safe stopping distance is the total displacement distance required for a target vehicle to come to a complete stop from the moment a braking command is triggered in its current driving state or the edited target driving state. It fully covers the entire displacement process of vehicle braking response lag and physical braking slippage. It can be calculated in real time by reading the preset maximum braking deceleration, the vehicle's current speed, and the target speed corresponding to the editing request, combined with the inherent response delay characteristics of the system, using a standardized braking distance calculation formula. For example, if a user initiates an editing request to increase the vehicle speed during low-speed parking, the safe stopping distance corresponding to this editing condition can be calculated by combining the vehicle's real-time speed, the target speed after editing, and the vehicle's maximum braking performance. This distance is used to measure the minimum safe space required for emergency braking of the vehicle.

[0113] Safety margin is a metric representing the difference in remaining safety space in the current driving scenario of a target vehicle. It visually reflects the ample safety space between the vehicle's current obstacle distance and the maximum parking distance. Safety margin directly reflects the basic safety redundancy of the current driving task editing operation. A larger margin value indicates a more sufficient driving safety margin, while a smaller margin value indicates a higher risk of collision. The safety margin value can be obtained by extracting the current nearest obstacle distance from the obstacle information and subtracting the real-time calculated safe parking distance. For example, if the real-time distance of the nearest obstacle detected during the vehicle's parking process is greater than the safe parking distance calculated under the current conditions, the difference is a positive safety margin, indicating that the current editing request has basic safety conditions. If the obstacle distance cannot cover the safe parking distance, the safety margin is too low, and there is a risk of insufficient braking leading to a collision.

[0114] The preset safety threshold is a baseline value for judging the safety margin of hard safety constraints. It is a rigid critical indicator that distinguishes between safe compliance and risk violations in editing requests. This threshold is a fixed safety bottom line calibrated by the engineering and is used to avoid hidden risks caused by braking errors, environmental disturbances, and sensor deviations. It reserves sufficient safety redundancy for dynamic editing of vehicle driving tasks and belongs to the underlying safety constraint rules that cannot be broken. For example, in the low-speed automatic parking scenario, the preset safety threshold is a fixed calibrated value. When the real-time calculated safety margin is less than this threshold, it is determined that the safety margin of the current editing request is insufficient and there is an immediate risk of collision, resulting in a failed screening result.

[0115] For example, when executing the hard safety constraint screening process for a driving task editing request, the system first retrieves the preset maximum braking deceleration parameter fixed in the vehicle system, using it as the calculation benchmark for the vehicle's ultimate braking capability. Simultaneously, it acquires the target vehicle's current speed and the target speed included in the editing request. Combining these two speed parameters with the preset maximum braking deceleration, it calculates the safe stopping distance corresponding to the current editing condition using standardized braking distance calculation logic, accurately reconstructing the displacement required for the vehicle from braking to complete stillness. Subsequently, it analyzes the real-time obstacle information, extracts the distance data of the nearest obstacle to the target vehicle within the scene, and uses the nearest obstacle distance... The difference between the safety margin and the safe parking distance is calculated in real time to obtain the safety margin corresponding to the current scenario, quantifying the basic safety redundancy level of this editing operation. The real-time calculated safety margin is then compared with the preset safety threshold for judgment, and the screening judgment is completed in strict accordance with rigid safety constraint rules. When the safety margin is less than the preset safety threshold, it means that the remaining safety space in the current scenario cannot cover the vehicle's extreme braking needs, and there is an immediate collision risk in this editing request. The safety screening result indicating failure is then generated to intercept unsafe task editing operations in advance, terminate the subsequent complex trajectory generation and risk calculation process, and realize the underlying safety fallback protection in the dynamic editing scenario of driving tasks.

[0116] Through the implementation of the above embodiments, the safe stopping distance is calculated based on the maximum braking deceleration, the current vehicle speed and the target vehicle speed, and the safety margin is calculated in combination with the current nearest obstacle distance. When the safety margin is lower than the safety threshold, the editing request is directly determined to be rejected. This can quickly assess whether the editing request has a basic safety margin from the perspectives of vehicle braking capability and current environmental space conditions, which is conducive to timely interception of driving task editing requests with obvious collision risks.

[0117] In some embodiments, determining the safe stopping distance based on the preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request may include: determining the larger of the current vehicle speed and the target vehicle speed as the vehicle evaluation speed; determining the delayed travel distance by multiplying the vehicle evaluation speed by a preset system delay duration, wherein the preset system delay duration characterizes the response time required from triggering the braking condition to the target vehicle generating braking deceleration; determining the vehicle braking distance based on the ratio of the square of the current vehicle speed to twice the preset maximum braking deceleration; and determining the safe stopping distance by summing the delayed travel distance and the vehicle braking distance.

[0118] In some examples, the vehicle assessment speed is a conservative baseline speed used to calculate the safe parking distance. It can be obtained by comparing the target vehicle's current speed with the target speed corresponding to the edit request and selecting the larger value. This speed selection logic follows the principle of safety conservatism, prioritizing the use of higher speeds as the basis for braking distance calculation. It can cover the worst driving conditions after the speed increase in the edit operation, avoiding the problem of underestimating the braking distance and omitting safety risks due to selecting low speed parameters. For example, if the target vehicle's current parking speed is 0.4 meters per second, and the user initiates an edit request to adjust the target speed to 0.6 meters per second, after comparing the two speed values, 0.6 meters per second is selected as the vehicle assessment speed. The braking safety distance is calculated based on the extreme conditions after the speed increase, ensuring the conservatism and safety of the screening results.

[0119] The preset system delay duration is a standardized fixed duration parameter characterizing the inherent lag characteristics of the entire braking response process of an intelligent driving system. It refers to the complete time interval from the moment the onboard system triggers the braking control condition to the moment the target vehicle's chassis actuator stably outputs braking deceleration. This parameter covers the lag characteristics of the entire intelligent driving response chain and can quantify the inherent delays in the system's perception, decision-making, transmission, and execution responses, thus compensating for the technical deficiency of purely physical braking distance calculations that ignore system software and hardware lags. For example, in low-speed automatic parking, a fixed preset system delay duration is obtained through full-link calibration. This duration is adapted to the low-speed parking control logic and the response characteristics of the onboard hardware, and can accurately calculate the vehicle's sliding distance during the delay phase, thus improving the comprehensive calculation of safe parking distance.

[0120] Based on this, the formula for vehicle braking distance is: (4) In the formula, Represents the vehicle's purely physical braking and coasting distance; The vehicle's current speed in real time; The maximum usable braking deceleration of a vehicle under current road conditions can be determined by the road surface adhesion coefficient. You can obtain it by looking up a table or by using a formula. Online real-time estimation, This is the braking efficiency coefficient. It represents gravitational acceleration.

[0121] The formula for total safe parking distance is: (5) In the formula, The total distance required for a complete stop, including system response delay and physical braking; That is, the preset system delay duration defined above; To conservatively estimate the vehicle speed, the current speed is used. Target speed after editing The higher of the two values ​​is used to cover the extreme and dangerous operating conditions after the speed increase.

[0122] The formula for the remaining safety margin is: (6) In the formula, This represents the remaining safe distance between the vehicle and the nearest obstacle. The minimum real-time distance from the vehicle to surrounding obstacles is output by the perception module; after calculating the safety margin, it is compared with the hard safety threshold calibrated in the engineering. (Comparison of typical calibration value of 0.3m in low-speed parking scenarios) If The screening for this editing request was deemed unsuccessful.

[0123] In addition to safety margin determination, the screening also simultaneously verifies lateral motion constraints and task time constraints: if the expected lateral acceleration corresponding to the edited parameters exceeds the scene calibration limit... If the remaining execution time of the task cannot match the edited action sequence, the screening will also be deemed a failure. This includes the lateral acceleration limit. Instead of directly using the road surface theoretical adhesion limit Instead, it is dynamically assigned values ​​through a joint calibration table, taking into account road surface adhesion level and parking stability requirements; the adhesion coefficient range of dry asphalt pavement [0.75, 0.90] corresponds to a maximum lateral acceleration of 1.2 m / s². 2 The adhesion coefficient range of [0.45, 0.75] on a wet road surface corresponds to 1.0 m / s2 The adhesion coefficient range of low-adhesion pavement [0.20, 0.45] corresponds to 0.8 m / s. 2 This calibration method preserves the influence of road surface adhesion on longitudinal braking capability while avoiding safety loopholes caused by setting excessively wide thresholds in low-speed parking scenarios.

[0124] For example, a conservative screening of the baseline speed can be performed first. The current speed of the target vehicle and the target speed corresponding to the editing request are collected simultaneously. The larger value between the two is selected as the vehicle evaluation speed by numerical comparison. The most stringent vehicle speed condition is used as the distance calculation benchmark to avoid the problem of insufficient safety prediction caused by low-speed evaluation. The preset system delay time parameter fixed on the vehicle is retrieved and multiplied with the screened vehicle evaluation speed to accurately solve the delayed travel distance of the vehicle during the continuous sliding stage of the system braking response lag, which fully covers the displacement increment caused by the response delay of the entire intelligent driving link. At the same time, based on the real-time current vehicle speed, the pure physical braking distance after the vehicle stabilizes is calculated by calculating the ratio of the square of the vehicle speed to twice the preset maximum braking deceleration, which accurately reflects the sliding displacement corresponding to the vehicle's hardware limit braking capability. Finally, the delayed travel distance and the vehicle braking distance are accumulated and fused to obtain a complete safe stopping distance that takes into account both the system software response characteristics and the vehicle hardware braking characteristics. This achieves a full-condition, high-precision, and conservative solution for the safe stopping distance, providing accurate and reliable judgment parameters for subsequent vehicle safety margin calculation and hard safety constraint screening of driving task editing requests.

[0125] By implementing the above embodiments, the larger of the current vehicle speed and the target vehicle speed is used as the vehicle evaluation speed, and the safe stopping distance is determined by comprehensively considering the delayed travel distance and the vehicle braking distance. This ensures that the safe stopping distance not only considers the vehicle braking process, but also the distance the vehicle continues to travel during the system response period, thereby improving the consistency between the calculated safe stopping distance and the actual vehicle braking process, and thus improving the accuracy of hard safety screening.

[0126] In some embodiments, the preset system delay duration is determined as follows: the system delay duration is obtained by summing the perception delay duration, decision delay duration, control command transmission delay duration, and actuator response delay duration; wherein, the perception delay duration represents the time interval from data acquisition by the environmental sensor to the output of structured perception results, the decision delay duration represents the time interval from receiving the perception results to generating control commands, the control command transmission delay duration represents the time interval from the transmission of control commands to the actuator via the communication bus, and the actuator response delay duration represents the time interval from the actuator receiving the command to starting to generate braking action.

[0127] In some examples, the perception delay duration is the time interval of the complete data processing chain of the intelligent driving perception system. Specifically, it represents the entire time interval from the moment the environmental sensors start collecting raw data to the moment the perception system outputs a structured environmental perception result that can be used for subsequent decision-making calculations. This duration covers the total time consumed by raw data sampling, data preprocessing, target detection and clustering, multi-sensor fusion, and result filtering output. For example, in low-speed automatic parking, the vehicle's surround-view camera and near-field radar simultaneously collect surrounding environmental data. After image decoding, target recognition, obstacle clustering, and fusion filtering, the environmental perception result is output. The fixed time corresponding to this complete processing flow is the perception delay duration.

[0128] Decision delay duration is the computational time interval of the intelligent driving decision-making and planning link. Specifically, it represents the time interval from the moment the decision module fully receives the valid structured perception results to the moment the decision module completes trajectory planning, safety verification, instruction parsing, and generates standardized chassis control instructions. This duration covers the total algorithmic time consumption of data parsing, condition determination, trajectory extrapolation, safety logic operations, and control instruction encapsulation. For example, in a parking scenario, after the system receives perception data on obstacle distribution and vehicle position and attitude, it completes computational processes such as trajectory fine-tuning, safety margin calculation, and vehicle speed instruction generation. The fixed time consumption of this entire algorithmic processing flow is the decision delay duration.

[0129] The control command transmission delay is the signal transmission time interval of the vehicle bus communication link. Specifically, it represents the time interval from when the decision module generates a standardized control command to when the control command is completely transmitted to the vehicle chassis actuator control unit. This time covers the entire process of command encoding, bus protocol parsing, channel transmission, data verification and decoding. For example, after the decision module generates a braking control command, it transmits it to the electronic stability control system execution unit through the vehicle controller local area network bus. The fixed time generated by the command transmission and verification process is the control command transmission delay.

[0130] Actuator response delay is the time interval for the mechanical response of a vehicle chassis actuator. Specifically, it represents the time interval from when the chassis actuator control unit fully receives a valid control command until the actuator completes pressure build-up, torque output, and begins to generate effective braking action. This time covers the entire process of command parsing, mechanism initiation, pressure build-up, and mechanical action response. For example, after receiving a braking command, the electronic stability control system needs to complete actions such as hydraulic pressure build-up and mechanism drive. The fixed time from receiving the command to the vehicle generating detectable braking deceleration is the actuator response delay.

[0131] For example, the inherent latency parameters of the four links in the entire intelligent driving process can be independently calibrated first. The perception latency of the perception link, the decision latency of the decision link, the control command transmission latency of the communication link, and the actuator response latency of the chassis execution link are collected respectively. The perception latency quantifies the entire process time from environmental data acquisition to structured perception result output. The decision latency quantifies the algorithm calculation time from receiving the perception result to generating the control command. The control command transmission latency quantifies the communication time of the control command bus transmission. The actuator response latency quantifies the mechanical response time from receiving the command to the braking action taking effect. After completing the calibration of the four independent latency parameters, the four latency times are linearly summed, and the total time obtained is determined as the preset system latency time adapted to the current working condition.

[0132] By implementing the above embodiments, the system delay duration is determined by the perception delay, decision delay, control command transmission delay, and actuator response delay. This allows the system delay to cover the main time-consuming links in the vehicle control link. Compared with using fixed empirical delay values, this can more realistically reflect the actual response capability of the vehicle, improve the accuracy of the safe stopping distance calculation, and thus improve the reliability of the safety judgment in vehicle driving task editing.

[0133] In some embodiments, if the aforementioned safety screening result characterization passes, then based on the current vehicle state in the current operating state information, a dynamic reachability analysis is performed to generate a candidate trajectory. This may include: if the safety screening result characterization passes, then based on the edit request and the current task sequence in the current operating state information, a constraint consistency solution is performed to obtain constraint solution information; based on the constraint solution information and the current vehicle state in the current operating state information, a dynamic reachability analysis is performed to generate a candidate trajectory.

[0134] In some examples, the current task sequence is a structured intelligent driving task sequence queue that the target vehicle is executing at the time the edit request is triggered. It is formed by the orderly combination of multiple sub-tasks with sequential execution logic and constraint relationships. The current task sequence defines the vehicle's established driving execution logic, scenario adaptation rules, and task switching constraints, which can limit the parameter adjustment boundaries of this edit request and ensure that the editing operation will not destroy the overall execution logic and timing rationality of the original driving task. For example, if the target vehicle is performing an automatic parallel parking driving task, the corresponding current task sequence includes multiple consecutive sub-tasks such as vehicle alignment, lateral position fine-tuning, low-speed reversing into the parking space, attitude correction, and parking in the parking space. Each sub-task has a fixed execution sequence and switching constraints, which constitute a complete current task sequence to constrain the effective range of this vehicle speed parameter editing.

[0135] Constraint consistency solving is a compliance matching process performed on user editing requests and the rules of the vehicle's current execution task. It verifies whether the editing parameters are compatible with the timing constraints, scenario constraints, and execution constraints of the current task sequence. This process can eliminate invalid editing parameters that conflict with the established task logic, correct control instructions that exceed the task constraint boundaries, and ensure that the editing request does not disrupt the execution timing and scenario adaptation logic of the original driving task after it takes effect, thus achieving consistent adaptation between editing instructions and the native task system. For example, the parking sub-task has a preset low-speed execution constraint. When the user initiates an editing request to increase the vehicle speed, constraint consistency solving will verify whether the adjusted vehicle speed fits the speed constraint range of the parking task, and solve for the valid vehicle speed boundary parameters that comply with the task rules, avoiding task execution disorder caused by parameter exceeding limits.

[0136] Constraint solution information is a standardized, structured set of information output after constraint consistency solution is completed. It is used to characterize the effective constraint boundaries and compliance control parameters of the current editing request to fit the current task sequence. Constraint solution information may include the task temporal constraint range, parameter adjustment upper limit, legal control interval, and task compatibility judgment result. It serves as the prerequisite constraint basis for subsequent dynamic reachability analysis and candidate trajectory generation, ensuring that the trajectory generation process conforms to the original driving task logic. For example, after constraint consistency solution, the generated constraint solution information includes parameters such as the maximum driving speed allowed by the current parking sub-task, attitude adjustment constraints, and trajectory offset threshold, which limit the effective execution boundaries of the current editing operation.

[0137] The process of generating candidate trajectories by performing dynamic reachability analysis based on the current vehicle state from constraint solution information and current operating state information involves using constraint solution information to define the task compliance boundary, using the current real vehicle state as the initial physical condition, and combining the inherent limit constraints of vehicle dynamics to deduce the feasible motion state range of the vehicle and generate compliant candidate trajectories. This process simultaneously considers the task rule logic constraints and vehicle physical motion constraints, ensuring that the edited trajectory does not violate the execution sequence and scenario rules of the original driving task, and ensuring that the trajectory is within the executable range of vehicle dynamics throughout, achieving a dual guarantee of rule compliance and physical feasibility. For example, relying on constraint solution information to define the speed and trajectory deviation constraints of the parking task, and combining the current vehicle speed, body posture, and chassis dynamic limit state to conduct reachability analysis, within the dual boundaries of rule constraints and physical constraints, a complete candidate trajectory that adapts to the user's editing intent, fits the parking task logic, and conforms to the vehicle's motion characteristics is deduced and generated.

[0138] For example, firstly, the complete task queue currently being executed by the target vehicle can be read from the cache of the onboard task scheduling system. This queue is then parsed to obtain the current task sequence with temporal logic and constraint rules, clarifying the execution boundaries, parameter limitations, and timing switching conditions of the subtasks in the current stage. Next, the user-initiated edit request parameters are correlated and matched with the inherent constraint rules of the current task sequence. The onboard constraint parser performs constraint consistency calculations, verifying the compatibility between the edit parameters and task rules. This yields the effective parameter constraint boundaries and compliance control conditions adapted to the current driving task, which are then structured and encapsulated to form complete constraint solution information, achieving adaptation and unification between the edit command and the native driving task logic. Based on this, the vehicle information included in the current operating status information is retrieved synchronously. Initial state parameters such as real-time location, driving speed, vehicle attitude, and dynamic limits are used as upper-level task rule constraints, and the current vehicle state is used as the lower-level physical initial boundary. A full-dimensional dynamic reachability analysis is conducted using an onboard reachability analyzer to deduce a feasible state interval that simultaneously satisfies both task rule constraints and vehicle physical motion constraints. Within this feasible state interval, time-series trajectory simulation is performed, generating future vehicle trajectory points hour by hour. Finally, these are integrated to form a continuous, compliant, and executable candidate trajectory for the target vehicle. This ensures that dynamic editing operations do not disrupt the original driving task execution logic and that the generated trajectory fully adapts to the vehicle's real-time dynamic characteristics, providing accurate and effective trajectory input support for subsequent continuous collision detection and probabilistic risk quantification.

[0139] Through the implementation of the above embodiments, before the dynamic reachability analysis, constraint consistency is solved based on the editing request and the current task sequence, and candidate trajectories are generated in combination with the constraint solution information. This ensures that the generated candidate trajectories not only meet the vehicle dynamic constraints, but also meet the logical and execution constraints of the current driving task, reducing the generation of invalid trajectories due to task conflicts and improving the overall executability of vehicle driving task editing.

[0140] In some embodiments, the aforementioned constraint consistency solution based on the current task sequence in the edit request and current running status information to obtain constraint solution information may include: parsing the constraints of the edit request to obtain incremental constraints; extracting tasks from the current task sequence to obtain at least one task constraint, wherein the task constraint may include at least one of precondition constraints, postcondition constraints, temporal dependency constraints, and safety boundary constraints; merging the incremental constraints and task constraints to obtain a constraint set; performing consistency determination on the constraint set to obtain a determination result; if the determination result is inconsistent, identifying the conflict terms between the incremental constraints and task constraints, and solving the minimum modification set of the constraint set based on preset relaxation rules, using the minimum modification set as constraint solution information, wherein the minimum modification set is a set of parameter corrections that can eliminate conflict terms; if the determination result is consistent, generating feasible assignments as constraint solution information, wherein the feasible assignments are a set of control parameter values ​​that satisfy all constraint conditions.

[0141] In some examples, constraint parsing is a standardized process for structurally decomposing and extracting constraints from user-issued editing requests. It extracts quantifiable, consistency-compliant parameter constraints from unstructured or semi-structured editing instructions. This process can identify the parameter adjustment dimensions, adjustment ranges, target control quantities, and change constraint relationships corresponding to the editing request, transforming the user's subjective editing intent into an objective machine constraint model. Upon receiving the original editing request from the upper layer, it can use a preset instruction parsing template to identify parameter dimensions, extract numerical boundaries, and model constraint relationships, automatically outputting standardized constraint parameters. For example, if a user initiates an editing request to increase the vehicle speed while parking at low speed, the constraint parsing process accurately extracts the speed adjustment dimension, target speed value, and speed change rate constraint corresponding to this editing, forming a standardized parameter constraint expression that departs from the semantic form of the original interaction statement and adapts to the backend constraint operation logic.

[0142] Incremental constraints are the newly added set of constraints obtained after constraint parsing. They are generated entirely by the driving task editing request issued by the user in real time. They are used to characterize the parameter change requirements and new control boundaries brought about by this dynamic editing operation. Incremental constraints are different from the vehicle's original fixed task constraints. They are dynamic constraints that are superimposed in real time and are the source of variables that trigger the consistency verification of the current constraints. For example, after the user modifies the parking constant speed driving speed parameter, the generated incremental constraints include the new target speed limit, speed smooth change constraints, and trajectory control constraints for the corresponding stage. They are used to express all the parameter restrictions added in this editing operation.

[0143] Taking a typical parking scenario editing requirement as an example, the constraint coding logic can be intuitively demonstrated: If a user initiates an editing request to "increase the parking speed from 0.5m / s to 0.9m / s", the constraint parsing process will extract four types of core variables to complete formal modeling. The definitions of each variable are as follows: This represents the target parking speed set by the user after editing. This represents the maximum safe speed allowed under the current parking conditions; This represents the instantaneous lateral acceleration generated when the vehicle performs a turn at that speed; This represents the lateral acceleration safety limit calibrated under the current road surface adhesion conditions; This represents the actual time required to perform this parking maneuver; This represents the maximum remaining allowed execution time for the current task.

[0144] Write assertion constraint statements for the above variables according to the SMT-LIB specification. The complete code and the meaning of each constraint are as follows: (assert(=v_target0.9)): Assignment constraint, limiting the target speed variable to the user's edited value of 0.9m / s, used to express the user's original intention to speed up the editing; (assert(<=v_targetv_max_safe)): Speed ​​safety boundary constraint, requiring that the speed of the target vehicle after editing must not exceed the upper limit of the scene's allowed safe speed; (assert(<=a_lata_lat_max)): Dynamic boundary constraint, requiring that the lateral acceleration generated during driving does not exceed the vehicle stability limit; (assert(<=t_exect_window)): Timing window constraint, requiring that the time taken for a complete parking action does not exceed the remaining available time of the task.

[0145] After all constraints are merged, the solution is fed into the SMT solver for consistency checking. If the solution result is UNSAT (unsatisfactory), the minimum conflict kernel is automatically located. In this example, under the vehicle motion conditions corresponding to the current parking space curvature and tire adhesion coefficient, and If the two sets of constraints cannot be met simultaneously, the solver will output the conflict reason as "the target speed is too high, causing the lateral acceleration to exceed the limit". Simultaneously, it will generate a minimum parameter modification set based on the preset relaxation rules (such as reducing the target vehicle speed to 0.7m) as constraint solution information, which will be used by the subsequent trajectory generation module to correct the control input boundary.

[0146] The current task sequence is a complete structured driving sub-task sequence that the target vehicle is executing at the moment of trajectory detection. It can be arranged in an orderly manner according to preset driving logic and execution sequence. The current task sequence defines the vehicle's motion constraints, safety avoidance logic, trajectory execution rules and state switching conditions at the current stage. It is the upper-level task basis for constraining the detection range of the expansion area and limiting the safety judgment logic, which can ensure that the distance detection process conforms to the task attributes of the current driving conditions. For example, when the vehicle is performing an automatic parking driving task, the current task sequence includes continuous sub-tasks such as trajectory following, attitude fine-tuning and close-range obstacle avoidance, which are adapted to the low-speed close-range safety detection logic and provide a working condition constraint benchmark for the distance detection of the expansion area.

[0147] Task extraction is a process of structured decomposition and constraint filtering of the currently executing task sequence. It is used to extract effective subtasks and corresponding constraint rules that are strongly related to the current editing period from the complete temporal task queue. Task extraction can filter out unrelated task constraints, retain the effective constraint dimensions that are currently in effect, and avoid invalid constraints from interfering with the accuracy of the current consistency solution. It can locate the currently active subtask based on the task execution node at the time the editing request is triggered, extract all the constraint rules of the corresponding subtask, and form the current effective task constraint set. For example, when a vehicle is in the parking execution phase, the task extraction operation can mask the task constraints of the previous positioning and driving, and only retain the effective constraint conditions corresponding to the current parking subtask, which are used for consistency verification with the incremental constraints of the current editing request.

[0148] Task constraints are the set of inherent standardized constraints inherent to the currently activated driving subtask. They are the underlying rules that ensure the safe, orderly, and stable execution of intelligent driving tasks according to preset logic. They are used to limit the parameter value range, execution order, switching conditions, and safety boundaries of tasks at each stage. They can include four basic constraint types: precondition constraints, postcondition constraints, temporal dependency constraints, and safety boundary constraints.

[0149] Precondition constraints are the prerequisite state conditions that limit the normal start and execution of the current subtask. They are used to specify the vehicle state, environmental state, and basic task progress conditions that must be met to enter the current task stage. If any precondition constraint is not met, the current subtask cannot start normally or run stably, which can effectively prevent the task from starting abnormally in a non-compliant state. For example, the precondition constraints of the parking subtask include entry conditions such as the vehicle being aligned, the lateral deviation of the vehicle body being less than a set threshold, and the vehicle being in a low-speed standby state, which are used to ensure that the parking action starts in a compliant initial state.

[0150] Post-condition constraints are closing state constraints that ensure the successful completion of the current subtask and allow transition to the next subtask. They specify the closing indicators such as position, attitude, and speed that the vehicle needs to achieve after the current task is completed. Post-condition constraints ensure that each subtask has clear completion criteria, enabling a smooth transition between task sequences. For example, the post-condition constraints for parking include closing indicators such as vehicle centering deviation, vehicle angle deviation, and final stationary speed. Only when the corresponding indicators are met can the parking be considered complete and the vehicle can enter the parking subtask.

[0151] Temporal dependency constraints are temporal rules that limit the execution order, stage duration, and state progression relationships among multiple subtasks. They are used to ensure that the overall driving task progresses according to a fixed logical temporal sequence, preventing abnormal situations such as out-of-order execution of tasks, overlapping stage execution, or skipping levels of execution. For example, the parking task strictly follows the temporal dependency constraint of first aligning, then fine-tuning, then parking, and finally stopping. Any editing operation must not disrupt the established task progression logic, ensuring the standardization of the parking process.

[0152] Safety boundary constraints are safety parameter boundary constraints that must be followed at all times during the operation of the current subtask. They are used to limit the range of safety-related parameters such as vehicle speed limits, turning angle limits, trajectory deviation limits, and lower limits of safe distances. They are the dynamic safety bottom line during the operation of the driving task. For example, the safety boundary constraints of the low-speed parking scenario include hard limits such as the maximum driving speed, the maximum lateral offset, and the minimum safe distance from obstacles. No parameter editing results may exceed this safety boundary.

[0153] A constraint set is a complete constraint system formed by merging incremental constraints generated by the user in real time with various task constraints extracted from the current task. It can include constraints added in this editing and inherent constraints of the original task in the system. For example, merging the incremental constraints of the user adjusting the vehicle speed with the timing constraints, safety boundary constraints, and state switching constraints of the parking task yields a complete constraint set under the current editing condition, covering all rule restrictions for this parameter adjustment.

[0154] The judgment result is the standardized conclusion information output after performing a global consistency check on the complete constraint set. It is used to characterize whether there are rule conflicts or parameter boundary contradictions between the incremental constraints and the original task constraints. The judgment result is divided into two categories: consistent and inconsistent, which directly determines the branch operation logic of subsequent constraint solving. For example, if the target vehicle speed increased by the user does not exceed the task safety boundary and does not conflict with the timing rules, the judgment result is consistent. If the user edits the parameters to exceed the task safety limit or disrupts the task timing logic, the judgment result is inconsistent.

[0155] Conflict items are constraint pairs in the constraint set that have logical contradictions, overlapping parameter ranges, or mutually exclusive execution conditions. They refer to specific rule dimensions and parameter values ​​that are incompatible between incremental constraints and existing task constraints. Conflict items can accurately locate the point of rule contradiction caused by the edit request, providing a clear object for subsequent constraint relaxation and parameter correction. For example, if the task safety boundary constraint limits the parking speed to a preset threshold, and the incremental constraint in the user's edit request sets the target speed to be higher than this threshold, the two form a conflict item in the parameter boundary dimension.

[0156] Preset relaxation rules are constraint conflict correction criteria that are calibrated and solidified offline. They are used to moderately modify incremental constraint parameters when constraint conflicts exist, based on the principles of minimal modification, highest safety priority, and maximum preservation of user editing intent. Preset relaxation rules follow the priority logic of safety constraints not being broken, timing constraints not being violated, and fine-tuning incremental parameters to achieve the rational resolution of conflicts. For example, a preset relaxation rule stipulates that speed over-limit conflicts only compress the incremental speed to the upper limit of the safety boundary, without changing the original task timing and attitude constraints, eliminating conflicts with minimal parameter changes, while preserving user editing intent to the maximum extent.

[0157] The minimum modification set is a set of optimal parameter corrections obtained by solving based on preset relaxation rules. It is a standardized parameter adjustment result that can completely eliminate constraint conflicts and minimize the degree of change to the original editing intent. The minimum modification set ensures that the conflict resolution process does not introduce additional rule contradictions, while maintaining the integrity of the original execution logic of the driving task. It can be based on the conflict dimension, with the minimum amount of change as the optimization objective, to iteratively solve for the appropriate parameter correction offset and form a structured minimum modification set. For example, if the user sets the vehicle speed to exceed the safety threshold, the minimum modification set obtained is to correct the target vehicle speed to the maximum allowable value of the safety boundary. This correction method only adjusts the parameters that exceed the limit, completely eliminates the conflict and retains the user's speed increase requirement to the maximum extent.

[0158] Feasible assignments are a set of standardized control parameter values ​​generated under conditions where the constraint set is completely consistent. This set of values ​​can simultaneously meet the rule requirements of both incremental constraints and all task constraints, without any logical conflicts or parameter overruns in any dimension. Feasible assignments fully inherit the user's legitimate editing intent, require no parameter correction, and can be directly used for subsequent dynamic reachability analysis and candidate trajectory generation. For example, if the user slightly increases the parking speed without exceeding the task safety boundary and timing constraints, the speed control parameters that fit the user's set target are directly generated, and these parameters are the feasible assignments that satisfy all constraints.

[0159] For example, the user's original editing request can first be parsed to extract quantifiable parameter adjustment rules and numerical boundaries from the unstructured editing instructions, and incremental constraints corresponding to this editing can be generated in a structured manner. At the same time, task extraction operations are performed on the currently executing task sequence to accurately locate the currently active subtask and extract the corresponding task constraints, fully covering four types of rule dimensions: precondition constraints, postcondition constraints, temporal dependency constraints, and safety boundary constraints. The incremental constraints generated in real time and the extracted multi-dimensional task constraints are merged and deduplicated to construct a complete and effective constraint set under the current editing conditions. Then, a global traversal and verification is performed on the constraint set to determine whether there are logical conflicts or parameter over-limit issues between the incremental constraints and the original task constraints, and the corresponding judgment results are output. When the judgment results are inconsistent, the rules within the constraint set are compared dimension by dimension to accurately locate the conflicting and mutually exclusive terms. The offline fixed preset relaxation rules are retrieved, and the minimum modification set that can completely eliminate the conflict is obtained through iterative solution with the optimization goals of minimum modification, safety priority, and intention preservation. This parameter correction set is then encapsulated as the final constraint solution information. When the judgment results are consistent, it is confirmed that the current incremental constraint is fully compatible with the existing task constraints and all rules in the constraint set can be satisfied simultaneously. The feasible assignment that fits the user's editing intention is obtained directly within the feasible domain of the constraint, and this set of compliance control parameters is used as the constraint solution information.

[0160] Through the implementation of the above embodiments, the edit request is parsed to form incremental constraints, the current task sequence is extracted to form task constraints, and the consistency between the two is determined. When there is a constraint conflict, the minimum modification set is further solved. When the constraints are consistent, a feasible assignment is output. This enables vehicle driving task editing not only to determine whether there is a task conflict, but also to provide a correctable solution for conflicting edit requests, thereby improving the flexibility and recoverability of vehicle driving task editing.

[0161] In some embodiments, the aforementioned dynamic reachability analysis based on the current vehicle state from constraint solution information and current operating state information to generate candidate trajectories may include: if the constraint solution information is a feasible assignment, then obtaining the first control parameter from the feasible assignment, and determining the first control input boundary based on the first control parameter and the current vehicle state; if the constraint solution information is a minimum modification set, then obtaining the second control parameter from the minimum modification set, and determining the second control input boundary based on the second control parameter and the current vehicle state; taking the current vehicle state as the initial state, and within the first control input boundary or the second control input boundary, performing forward propagation using a preset vehicle model. The reachable state propagation yields a set of reachable states, which represents the continuous spatial state envelope region that the target vehicle can reach at each time step under the constraints of the first or second control input boundary. If the set of reachable states meets the preset dynamic constraints, a nominal control sequence is selected from the corresponding control input boundary for forward simulation to obtain a candidate trajectory. The preset dynamic constraints may include at least one of lateral acceleration constraints, yaw rate constraints, and minimum turning radius constraints. The nominal control sequence is a set of time-series control commands that satisfy the preset dynamic constraints within the first or second control input boundary.

[0162] In some examples, the first control parameter is a control variable extracted from the set of compliant feasible control parameters when the constraint solution information is feasible. It is used to characterize the compliant control instruction value corresponding to the user's original editing request. This parameter fully satisfies all the constraints of the current task sequence, requires no parameter correction, and can completely preserve the user's editing intent. It is the input parameter for vehicle trajectory extrapolation and control boundary calculation. For example, if the user slightly increases the target driving speed within the parking compliance parameter range, and there is no conflict in the constraint set, the corresponding target speed and trajectory following parameters are extracted from the feasible assignment to form the first control parameter required for this trajectory generation.

[0163] The first control input boundary is the compliant control input range determined jointly by the first control parameter and the current real-time vehicle state. It is used to limit the upper and lower limits of all output control commands during the subsequent trajectory simulation of the vehicle. This boundary takes into account the user's legitimate editing intention, the current initial dynamic state of the vehicle, and the inherent constraints of the task. The first control parameter can be used as the control target, combined with the initial vehicle state such as the current vehicle speed, body posture, and chassis state, to solve for the continuous upper and lower limits of control input adapted to the current working conditions, forming a standardized first control input boundary. For example, based on the compliantly increased parking target speed and the current stationary initial state of the vehicle, the legitimate input range of speed control and steering angle control can be solved. This range is the first control input boundary of this trajectory simulation, and all subsequent control commands must not exceed this boundary range.

[0164] The second control parameter is a compliant control variable obtained after correction by preset relaxation rules under the condition that the constraint solution information is the minimum modification set. It is used to characterize the optimal correction control command value after conflict resolution. This parameter eliminates all conflicts between incremental constraints and task constraints while minimizing changes to the user's original editing intention, and adapts to the rule system and safety boundary requirements of the current driving task. For example, if the parking speed set by the user exceeds the task safety boundary, the over-limit speed is corrected to the compliant maximum value through preset relaxation rules. The corrected compliant speed parameter is the second control parameter.

[0165] The second control input boundary is a compliant control input range jointly determined based on the corrected second control parameters and the current real-time vehicle state. It serves as the trajectory control constraint benchmark under conflicting conditions. This boundary uses the corrected compliant parameters as a benchmark and combines the real-time vehicle dynamics state to limit the range of control command output, ensuring that subsequent trajectory extrapolation fully conforms to mission safety rules and vehicle physical limits. The second control parameters can be used as the corrected control target, matched with the current initial motion state of the vehicle and the mission constraint boundary, to iteratively solve for a conflict-free and executable control input range, thus generating the second control input boundary. For example, based on the corrected compliant parking speed parameters and the current lateral attitude state of the vehicle, the corresponding speed adjustment range and steering angle control range can be solved to form a second control input boundary adapted to conflict correction conditions.

[0166] The preset vehicle model is a standardized vehicle dynamics and kinematics simulation model that is calibrated and solidified offline. It is used to accurately depict the state evolution of the target vehicle under different control inputs. This model is fully adapted to the vehicle motion characteristics under low-speed parking conditions and can characterize the correspondence between vehicle speed response, steering response, attitude change and displacement evolution. It is the underlying model for forward state propagation and trajectory simulation. Through massive low-speed real vehicle tests, multi-dimensional motion response data of the vehicle can be collected and fitted to obtain a standardized vehicle model adapted to low-speed scenarios of intelligent driving. For example, the preset vehicle model can accurately map the correspondence between vehicle steering input and lateral displacement and yaw attitude under parking conditions, providing accurate model support for subsequent reachable state deduction and candidate trajectory generation.

[0167] Forward reachability propagation starts with the vehicle's current real-time state and, within defined control input boundaries, it iteratively calculates all possible future reachable motion states of the vehicle using a pre-defined vehicle model. This propagation process does not output a single fixed trajectory but rather traverses all feasible motion states corresponding to legal control inputs, achieving full-domain coverage of the vehicle's future motion space. For example, starting with the vehicle's current initial parking posture, within the compliant speed and turning angle input boundaries, it iteratively calculates all feasible positions and postures of the vehicle, completing the forward reachability propagation of the entire parking process.

[0168] The reachable state set is the global state envelope set output after the forward reachable state propagation is completed. It is used to characterize all continuous spatial state regions that the target vehicle can physically reach at all future times under given control boundary constraints. This set completely defines the physically feasible motion range of the vehicle under the current editing condition, eliminates all invalid states that exceed the vehicle's dynamic limits, and provides a feasible domain benchmark for subsequent nominal control sequence selection. For example, the reachable state set under the parking condition includes all valid states such as the vehicle's feasible driving position, body attitude, and driving speed at all future times, completely covering all feasible parking motion ranges corresponding to the current editing parameters.

[0169] Preset dynamic constraints are a set of vehicle physical limit constraints pre-defined for the trajectory generation stage. They are used to filter valid states within the reachable state set that conform to the vehicle's actual motion characteristics and eliminate unreasonable states that exceed dynamic response limits. These constraints serve as a fallback standard for the final compliance of the trajectory and include at least one of lateral acceleration constraints, yaw rate constraints, and minimum turning radius constraints, ensuring that the generated trajectory is smooth, controllable, and conforms to the vehicle chassis dynamic limits. Based on the vehicle chassis hardware limits, driving stability specifications, and low-speed safety control standards, the limit thresholds of various dynamic parameters can be calibrated offline and solidified into a unified set of dynamic constraints. For example, the preset dynamic constraints under low-speed parking conditions limit the vehicle's maximum lateral acceleration, maximum yaw rate, and minimum turning radius, preventing unsafe conditions such as sharp turns and excessive lateral deviations from the trajectory.

[0170] Lateral acceleration constraints are dynamic constraints that limit the maximum value of a vehicle's lateral acceleration. They are used to control the intensity of the vehicle's lateral dynamic response, avoid vehicle instability and attitude deviation caused by excessive steering or sudden changes in lateral acceleration and deceleration, and ensure the smoothness and stability of low-speed driving. For example, during the parking trajectory generation process, the instantaneous lateral acceleration of all trajectory segments must not exceed a preset threshold to avoid attitude loss caused by excessive lateral impact on the vehicle.

[0171] Yaw rate constraint is a dynamic constraint that limits the maximum angular velocity of a vehicle body around its vertical centerline. It is used to control the vehicle's steering rate and attitude change rate, prevent body swaying and trajectory instability caused by rapid steering, and ensure a smooth and continuous steering process. For example, when the vehicle is finely adjusting its parking posture, the body rotation angular velocity strictly follows the yaw rate constraint to ensure a smooth and continuous steering process without sudden swaying.

[0172] The minimum turning radius constraint is a physical constraint that limits the minimum turning radius that a vehicle can achieve. It corresponds to the hardware limit of the vehicle's chassis steering mechanism and is used to prevent the generation of theoretical trajectories that exceed the vehicle's steering capabilities, ensuring that the trajectory is realistically executable. For example, during the trajectory generation process, the turning radius of all arc turning segments is greater than or equal to the preset minimum turning radius to avoid generating over-limit steering trajectories that the vehicle cannot execute.

[0173] The nominal control sequence is a set of time-standardized control commands that fully satisfy preset dynamic constraints within the effective control input boundary. It consists of vehicle speed commands and turning angle commands at consecutive moments. This sequence is the optimal control timing command that is smooth, safe, and conforms to the user's editing intentions, selected from all feasible control commands in the entire domain. It is the direct input basis for generating candidate trajectories in forward simulation. All feasible control command combinations can be traversed within the control input boundary to select the timing command combination that satisfies the dynamic constraints and has the best trajectory smoothness throughout the entire process, thus generating the nominal control sequence. For example, within the compliant vehicle speed and turning angle boundary, timing control commands that fully satisfy the acceleration, yaw rate, and turning radius constraints throughout the entire process are selected to form the nominal control sequence required for parking trajectory derivation.

[0174] Forward simulation is a simulation process that uses the nominal control sequence as the temporal control input and the initial state of the vehicle as the starting point for deduction. It iteratively calculates the future position and attitude of the vehicle time-by-time using a pre-set vehicle model, continuously generating a complete driving trajectory. This simulation process strictly follows the laws of vehicle kinematics and dynamics, accurately reproducing the vehicle's motion response corresponding to control commands and outputting realistic and executable candidate trajectories. For example, by loading a parking nominal control sequence and simulating the vehicle's displacement and attitude changes time-by-time, a continuous, smooth, and dynamically compliant complete parking candidate trajectory is ultimately generated.

[0175] For example, in low-speed parking situations, a bicycle model is preferentially selected as the preset vehicle model, and this model defines a state vector. Optionally increase yaw rate Expanding dimensions; where For the vehicle's lateral coordinates, For vehicle longitudinal coordinates, For vehicle body heading angle, For the longitudinal speed of the vehicle, The vehicle body yaw rate; model control input. ,in For the front wheel steering angle, For longitudinal acceleration and deceleration of the vehicle, the discrete simulation time step is... The value range is 0.05s to 0.1s. Taking the vehicle's real-time current state as the initial state for reachability deduction, and combining it with the first or second control input boundary obtained in the previous solution, the global range of values ​​for the steering angle and acceleration is determined.

[0176] The reachable state set is calculated iteratively using two methods: interval propagation and linear overapproximation. For critical and dangerous time slices, a refined envelope using Zonotope (fully symmetric multicellular object) can be employed to improve safety conservatism. The complete set of recursive calculation formulas and explanations of each parameter are as follows: The recurrence relation for the locally linearized state is: (7) In the formula, the subscript Representing the Discrete time steps for Vehicle state vector at any given time; This was obtained through single-step deduction. Real-time vehicle status; for The linearized matrix of the system state at each time step; for The input linearization matrix is ​​controlled at any time. for The control input quantity applied at any given time; This is the affine compensation constant term for the linearization process.

[0177] The basic formula for Zonotope is: (8) In the formula, represent Vehicles can reach the Zonotope envelope at any time; The center state vector of the Zonotope; The Zonotope generator matrix; To limit the value to The generating vector of the interval; For the generator dimension.

[0178] The formula for recursively updating the center is: (9) In the formula, For the post-deduction The center of Zonotope at all times; for Time reference center control input; This is the residual compensation term for single-step linearization.

[0179] The formula for the recursive update generator is: (10) In the formula, For the post-deduction The Zonotope generator matrix at time step; The generation matrix is ​​used to control input uncertainty; the matrix concatenation symbol represents merging the state generation term with the control disturbance generation term.

[0180] To ensure the real-time computing performance of the onboard system, the generator matrix is ​​completed each time. After the update, order reduction processing can be performed to limit the continuous expansion of the matrix dimension. After completing the calculation of the reachable set for all time steps, all state points within the reachable set are traversed one by one to check whether they violate the lateral acceleration constraint, yaw rate constraint, minimum turning radius constraint, and obstacle safety boundary constraint. If there are out-of-bounds states, the corresponding out-of-bounds constraint type and discrete time are recorded. If all reachable states in the entire domain satisfy all dynamic and safety constraints, a smooth and compliant nominal control sequence is selected from the control input boundary to conduct forward simulation, generating a deterministic candidate trajectory with complete temporal information. ,in Sampling time, To correspond to the vehicle's position, attitude, and speed at any given moment, a complete candidate trajectory is output for subsequent continuous collision detection processes.

[0181] For example, firstly, branch parameter extraction logic can be executed based on the type of constraint solution information. When the constraint solution information is a feasible assignment, the first compliant and valid control parameters are parsed and extracted from the feasible assignment set. Combined with the current real-time vehicle speed, vehicle posture, chassis status, and other current vehicle states, the first control input boundary adapted to the conflict-free editing condition is obtained, limiting the range of all legal control inputs for this trajectory extrapolation. When the constraint solution information is a minimum modification set, the optimized second control parameters are extracted from the conflict correction results. Based on the corrected compliant parameters and the current initial vehicle state, the second control input boundary that eliminates constraint conflicts is iteratively solved, ensuring that the control boundary fully adapts to the corrected editing requirements and task rules. Taking the current vehicle state corresponding to the time the editing request is triggered as the unified initial extrapolation starting point, the corresponding control input boundaries are matched according to different working conditions. The preset vehicle model is called to carry out global forward reachable state propagation, time-by-time. The system iteratively simulates all feasible motion states of the vehicle within the legal control input range, integrates and fits them to form a continuous and complete set of reachable states, comprehensively covering all physically feasible driving spaces and attitude ranges of the vehicle under the current editing conditions. Further, it performs dynamic compliance verification on the reachable state set, matching preset dynamic constraints consisting of lateral acceleration constraints, yaw rate constraints, and minimum turning radius constraints. It then filters out effective control input intervals that fully meet the vehicle's physical limits and stability requirements. Within these effective intervals, it selects control command combinations that are time-continuous, have optimal smoothness, and align with the user's editing intent, generating a standardized nominal control sequence. Finally, using the nominal control sequence as the time-series input, it conducts a full-process forward simulation through a preset vehicle model, updating the vehicle's position and attitude state time-by-time, continuously stitching together all simulation trajectory points, and ultimately generating a high-quality candidate trajectory that is dynamically compliant throughout the process, adapts to task constraints, and can be directly used for subsequent risk quantification assessment.

[0182] By implementing the above embodiments, different control input boundaries are determined based on the constraint solution results, and vehicle dynamics forward reachability propagation is performed within the control input boundaries. Candidate trajectories are then generated when the dynamic constraints are met, ensuring that the candidate trajectories are always within the controllable range of vehicle dynamics. This avoids generating trajectories that the vehicle cannot actually execute, and improves the consistency between the vehicle driving task editing results and the actual vehicle control capabilities.

[0183] In some embodiments, step 103 may include: acquiring the obstacle prediction trajectory; performing time grid alignment processing on the candidate trajectory and the obstacle prediction trajectory according to a preset time step to obtain the aligned candidate trajectory time series and the aligned obstacle prediction trajectory time series; performing geometric expansion operations on the vehicle outline of the target vehicle and the obstacle outline corresponding to the obstacle prediction trajectory with a preset safe expansion distance to obtain the vehicle expansion region sequence and the obstacle expansion region sequence; determining the minimum distance between the vehicle expansion region and the obstacle expansion region corresponding to each aligned time for each time, to obtain the minimum distance sequence; determining the intervals in the minimum distance sequence that are continuously less than a preset distance threshold as collision intervals; and determining the minimum distance sequence and the collision intervals as collision distance time series data.

[0184] In some examples, the obstacle prediction trajectory is a sequence of future temporal motion states of obstacles output by the intelligent driving environment prediction system. It is used to characterize the continuous changes in the position, attitude, and speed of various obstacles in the driving scene within a preset time period in the future. This trajectory covers the motion characteristics of obstacles at multiple sampling moments in the future, and can accurately predict the dynamic driving trend of obstacles. It serves as the environmental input basis for vehicle dynamic collision detection and distance temporal analysis. Based on the historical motion data, contour size, motion attitude, and scene constraints of obstacles collected by the perception module, a continuous motion trajectory sequence of obstacles for a fixed time period in the future can be deduced through a temporal motion prediction model. For example, in the case of automatic parking, dynamic or static obstacles such as adjacent parked vehicles and slow-moving pedestrians are detected. The position change sequence of each obstacle in the next few seconds can be obtained through motion deduction. This structured motion sequence is the obstacle prediction trajectory.

[0185] The preset time step is a pre-defined unified sampling interval parameter for time-series data. It is used to standardize the temporal sampling accuracy of vehicle trajectory and obstacle trajectory, and to achieve standardized and equalized discretization of multi-source time-series data. This parameter is a fixed and calibrated temporal accuracy parameter, which can unify the temporal resolution of vehicle and obstacle trajectories, eliminate the temporal misalignment problem caused by inconsistent sampling frequencies of multi-source trajectories, and ensure the accuracy of subsequent spatiotemporal matching and distance calculation. For example, in a low-speed parking scenario, a fixed preset time step is used as the temporal sampling benchmark to uniformly discretize the candidate vehicle trajectory and the predicted obstacle trajectory into equally spaced temporal nodes, thereby achieving standardized alignment of the temporal dimensions of the two trajectories.

[0186] Temporal grid alignment is a time-series standardization matching process for candidate vehicle trajectories and predicted obstacle trajectories. It addresses the temporal misalignment caused by inconsistencies in the original sampling times and frequencies of the two types of trajectories. This process uses a preset time step as a unified time grid reference. Through temporal interpolation, resampling, and time matching, it converts two heterogeneous temporal trajectories into synchronous temporal sequences with one-to-one correspondence, providing a temporal alignment basis for time-by-time spatial distance calculation. A globally uniform time grid can be constructed with a preset time step, and interpolation and resampling can be performed on the original candidate trajectory and the predicted obstacle trajectory, respectively, to output dual-path temporal trajectories with strictly synchronized times. For example, if there are slight differences in the original sampling intervals of the candidate vehicle trajectory and the predicted obstacle trajectory, temporal grid alignment can map the two trajectories to equal time grid nodes, achieving accurate correspondence of trajectory data at each calculation time.

[0187] The aligned candidate trajectory time series is the standardized time-series trajectory sequence of the target vehicle obtained after time grid alignment processing. It is composed of trajectory parameters such as vehicle position, attitude, and driving state at uniform time intervals. This sequence is fully adapted to the grid time series reference with a preset time step and is strictly synchronized with the obstacle time series trajectory. For example, the sampling nodes of the original parking candidate trajectory are unevenly distributed. After time grid alignment processing, a standardized vehicle trajectory time series with a fixed step size, uniform time series, and time that can be matched one-to-one with the obstacle trajectory is generated.

[0188] The aligned obstacle prediction trajectory time series is the standardized temporal motion sequence of obstacles obtained after time grid alignment processing. It is composed of orderly motion parameters such as obstacle position, attitude, and contour boundary at uniform time intervals. This sequence is completely synchronized with the vehicle's aligned trajectory time series, eliminating multi-source trajectory time series deviations and accurately matching the vehicle's driving state at each moment. For example, the sampling frequency of the original predicted trajectory of a dynamic pedestrian is inconsistent with the vehicle's trajectory. After time alignment processing, an obstacle time series that perfectly matches the vehicle's trajectory time is generated.

[0189] The vehicle profile is a standardized geometric profile model that accurately represents the external physical boundary of a target vehicle. It is constructed based on the actual length, width, and body shape of the vehicle and is used to define the physical space occupied by the vehicle during driving. This profile is a closed and continuous geometric boundary and serves as the basic spatial reference for vehicle safety distance detection, geometric expansion processing, and collision determination. A standardized two-dimensional planar closed geometric profile model can be constructed based on the overall physical size parameters of the target vehicle. For example, in parking situations, the standard rectangular outer profile of the vehicle is used as the vehicle profile, which completely covers the entire physical occupied area of ​​the vehicle body and serves as the reference boundary for spatial distance detection.

[0190] An obstacle profile is a structured geometric profile model that represents the external physical boundary of an obstacle in a driving scene. It is constructed in real time based on the obstacle's real-time perceived size, shape features, and attitude angle to define the real-time spatial occupancy range of the obstacle. This profile is dynamically updated as the obstacle moves and serves as the benchmark for determining the spatial area of ​​the obstacle and calculating safe distances. It can be fitted and generated based on obstacle point cloud and image data collected by sensors to generate closed and continuous obstacle geometric profiles and output the spatial boundary range of the obstacle at the current moment. For example, for obstacles such as parked vehicles, curbs, and pedestrians in the scene, closed geometric profiles with corresponding shapes can be fitted and generated to accurately represent the actual spatial occupancy area of ​​each type of obstacle.

[0191] The preset safety expansion distance is a static safety redundancy parameter used to spatially expand and compensate for the contours of the vehicle and obstacles, quantifying the dynamic safety margin during vehicle operation. This parameter can cover uncertainties such as vehicle perception errors, control disturbances, and trajectory deviations, avoiding missed detection of risks due to contour boundary fit, and improving the safety and robustness of collision detection. For example, in low-speed automatic parking, a fixed preset safety expansion distance is configured to uniformly expand and compensate for the contours of the vehicle and obstacles, reserving sufficient dynamic driving safety margin and avoiding collision risks caused by minor disturbances.

[0192] The vehicle expansion region sequence is a time-series set of the target vehicle's safe occupancy area obtained after performing geometric expansion processing at each time step. It is composed of the vehicle expansion regions corresponding to each aligned time step. This sequence superimposes a safety redundancy space on the original vehicle body physical contour, fully representing the dynamic safe occupancy range of the vehicle at each time step. For example, based on the vehicle body posture at each time step of the parking trajectory, the vehicle contour is safely expanded frame by frame to generate a vehicle expansion region sequence covering the entire parking process.

[0193] The obstacle expansion region is the safe occupancy area obtained by expanding the obstacle outline by a preset safe expansion distance at a single moment. It is used to characterize the overall dangerous space range of the obstacle, including its physical size and safety redundancy. This region can cover the high-risk warning area around the obstacle. For each alignment moment, the obstacle outline can be expanded outward according to the preset safe expansion distance to generate a closed danger region containing safety redundancy. For example, for adjacent vehicle obstacles in a parking scenario, the original outline can be expanded outward with safety to obtain an obstacle expansion region containing safety avoidance margin.

[0194] The minimum distance sequence is a structured numerical sequence of minimum safe distances throughout the entire time series. It is composed of the shortest spatial distances between the vehicle's expansion region and the obstacle's expansion region at each alignment time. This sequence can fully reflect the dynamic safe distance change pattern between the vehicle and surrounding obstacles throughout the entire driving process. It is a numerical basis for identifying close-range risk zones and quantifying temporal collision risks. It can traverse all temporal alignment nodes, calculate the shortest Euclidean distance between the vehicle's expansion region and the obstacle's expansion region frame by frame, and integrate them in temporal order to generate a continuous minimum distance sequence.

[0195] The collision interval is a continuous high-risk time sequence interval selected from the minimum distance sequence. It refers to a time sequence segment in which the minimum distance between regions at multiple consecutive moments is less than a preset distance threshold. This interval can accurately locate the time period in which a vehicle is continuously in a close-range high-risk state during its driving process, distinguish between instantaneous small distance fluctuations and continuous high-risk risks, and effectively avoid misjudgment problems caused by single-point noise interference. For example, when a vehicle is continuously close to an adjacent vehicle during parking, the safe distance between multiple consecutive time sequence nodes is less than the preset distance threshold. This continuous high-risk time sequence segment is the collision interval.

[0196] For example, firstly, the predicted trajectories of effective obstacles within the scene can be obtained through an in-vehicle environment prediction system, fully understanding the future temporal movement and change patterns of obstacles. A preset time step is used as a unified temporal benchmark, and temporal grid alignment processing is performed synchronously on the candidate trajectory of the target vehicle and the predicted obstacle trajectory. Temporal standardization and adaptation of heterogeneous trajectories are completed through temporal interpolation and resampling, generating precisely synchronized aligned candidate trajectory time series and aligned obstacle prediction trajectory time series, completely eliminating the problem of temporal misalignment of multi-source trajectories. Simultaneously, the vehicle's fixed standard self-vehicle contour is retrieved, and obstacle contours at each moment are obtained by fitting real-time perception. A unified preset safety expansion distance is used to perform geometric expansion on both the self-vehicle contour and the obstacle contour. The process involves generating vehicle and obstacle expansion regions with safety redundancy at each time step, and integrating them sequentially to form a complete vehicle expansion region sequence. For each time alignment node, the minimum spatial distance between the vehicle and obstacle expansion regions is calculated frame by frame, and the minimum distance sequence covering the entire driving process is obtained by arranging and integrating them sequentially. The minimum distance sequence is further traversed, and continuous risk screening is performed based on a preset distance threshold. Time segments with consecutive distances less than the threshold are clustered and identified as collision intervals, accurately extracting continuous high-risk periods throughout the entire process. Finally, the minimum distance sequence containing global distance change features and the collision intervals containing local high-risk period features are structurally fused and encapsulated to generate collision distance time series data with complete dimensions and comprehensive features.

[0197] Through the implementation of the above embodiments, the candidate trajectory and the predicted obstacle trajectory are aligned by time grid, and the vehicle outline and obstacle outline are geometrically expanded respectively. Then, the minimum distance at each time moment is calculated and the continuous collision interval is identified, so that continuous collision detection can take into account both time synchronization and spatial safety margin, improve the integrity and accuracy of collision detection in the process of vehicle driving task editing, and provide more accurate distance time series data for subsequent risk quantification.

[0198] In some embodiments, the minimum distance between the vehicle expansion region and the obstacle expansion region corresponding to the aforementioned determined time may include: performing intersection detection on the vehicle expansion region and the obstacle expansion region to obtain the minimum detection distance between the vehicle expansion region and the obstacle expansion region; if the minimum detection distance is a positive value, then the minimum detection distance is determined as the minimum distance; if the minimum detection distance is zero or negative, then the penetration depth between the vehicle expansion region and the obstacle expansion region is calculated using the dilatation polyhedron algorithm, and the minimum distance is determined based on the penetration depth.

[0199] In some examples, intersection detection is a two-dimensional spatial topological relationship detection operation performed on the vehicle expansion region and the obstacle expansion region. It is used to determine whether the two closed geometric regions have spatial overlap, boundary contact, or separation. Based on standardized polygon boundary parameters, the boundary intersection detection of the vehicle expansion region and the obstacle expansion region can be completed edge by edge, and the spatial position topological relationship and initial distance parameters of the two regions can be output. For example, when the vehicle is close to the obstacle, intersection detection can accurately identify the different spatial states of the vehicle expansion region and the obstacle expansion region, such as complete separation, boundary contact, or partial overlap.

[0200] The minimum detection distance is the shortest spatial distance parameter between two regions output after intersection detection. It is used to quantify the relative interval between the vehicle's expansion region and the obstacle's expansion region. When the two regions are separated, the minimum detection distance represents the shortest Euclidean distance between the boundaries of the two regions. When the two regions overlap or intrude, the minimum detection distance is set to zero and the intrusion state is marked, which is used to trigger the subsequent penetration depth calculation branch. For example, when the vehicle maintains a sufficient safe distance from surrounding obstacles during parking, the two expansion regions are completely separated, and the minimum detection distance is a positive value. This value is directly used as the minimum distance at the current moment to construct the minimum distance sequence. When the vehicle's attitude shifts and the vehicle's expansion region and the obstacle's expansion region partially overlap, the detection distance is set to zero, the expansion polyhedron algorithm is started to calculate the penetration depth, and the intrusion depth is used as the minimum distance at the current moment to accurately represent the high-risk collision state.

[0201] The dilatational polyhedron algorithm is a refined depth-solving algorithm for intrusion scenarios in two-dimensional closed geometric regions. It is used to quantify the degree of intrusion between two overlapping polygonal regions. Based on the standardized dilatational polygon boundary topology, the algorithm can accurately calculate the spatial penetration degree of two regions through region reconstruction, overlapping surface fitting, and depth iteration. For example, when the dilatational region of a vehicle and the dilatational region of an obstacle partially overlap, the traditional boundary distance cannot quantify the degree of risk. The dilatational polyhedron algorithm accurately solves the region penetration depth by fitting the boundary of the overlapping region, thereby realizing the risk quantification of the intrusion situation.

[0202] Penetration depth is a quantitative indicator of the degree of intrusion obtained by solving the dilatation polyhedron algorithm when two expansion regions spatially overlap. It is used to characterize the maximum depth distance of the vehicle's safe zone intruding into the obstacle's safety warning zone. The larger the penetration depth value, the larger the overlap range between the two regions, and the higher the risk of close-range collision between the vehicles. It can accurately distinguish the risk levels of different intrusion degrees. Based on the topological relationship of the overlapping polygons, the geometric boundary of the overlapping region can be reconstructed, the maximum intrusion distance between the two regions can be calculated iteratively, and a standardized penetration depth value can be output. For example, if the vehicle's close-range deviation causes the safe expansion region to slightly intrude into the obstacle expansion region, the corresponding penetration depth can be calculated to quantify the severity of the current region overlap and provide a basis for assigning distance values ​​in high-risk working conditions.

[0203] For example, firstly, by combining the driving constraint logic corresponding to the current task sequence, high-precision intersection detection can be performed on the vehicle expansion region and the obstacle expansion region at the current moment. The spatial topological relationship of the two sets of polygon boundaries is traversed to obtain the initial minimum detection distance. Based on the numerical state of the minimum detection distance, branch decision logic is executed to complete the differential minimum distance solution. When the minimum detection distance is positive, it is determined that the two expansion regions are in a completely separated safe state with no region overlap or boundary intrusion risk. The minimum detection distance is directly assigned as the minimum distance at the current moment, which truly reflects the safety margin of the current working condition. When the minimum detection distance is zero or negative, it is determined that the two expansion regions have boundary contact or spatial overlap. The traditional boundary minimum distance cannot quantify the degree of intrusion risk. Then, the preset expansion polyhedron algorithm is called to perform geometric reconstruction and depth iteration solution on the overlapping region, accurately calculate the penetration depth between the two regions, and use the penetration depth as the minimum distance at the current moment to realize the risk quantification characterization of high-risk intrusion conditions.

[0204] Through the implementation of the above embodiments, the intersection detection of the vehicle expansion area and the obstacle expansion area is first performed. When no intersection occurs, the minimum detection distance is directly used. When an intersection occurs, the penetration depth is further calculated using the expansion polyhedron algorithm to determine the minimum distance. This enables the collision detection to not only accurately reflect the safe distance between the vehicle and the obstacle before a collision occurs, but also accurately describe the degree of spatial overlap after a collision, thereby improving the accuracy of the collision distance time series data.

[0205] In some embodiments, step 105 may include: if the collision risk quantification value is greater than or equal to a first preset risk threshold, generating a first decision result of rejecting the editing request; if the collision risk quantification value is less than a second preset risk threshold, generating a second decision result of accepting the editing request, wherein the second preset risk threshold is less than the first preset risk threshold; if the collision risk quantification value is greater than or equal to the second preset risk threshold and less than the first preset risk threshold, generating a third decision result of delaying the editing request and waiting for manual confirmation.

[0206] In some examples, the first preset risk threshold is a preset high-risk critical judgment benchmark value, which belongs to the highest level of risk constraint threshold of the driving task editing request. It is used to determine whether there is a high-risk collision hazard in the driving condition corresponding to the current editing request. This threshold is a rigid safety interception threshold, used to distinguish between high-risk unexecutable conditions and medium-to-low-risk assessable conditions. For example, in the automatic parking dynamic parameter editing condition, the first preset risk threshold corresponds to the risk critical value of a high probability of close-range collision with the vehicle. When the collision risk quantification value reaches this threshold, it is determined that the current editing parameters will cause an extremely high collision hazard, and the editing request must be rejected directly.

[0207] The second preset risk threshold is a preset low-risk safety judgment benchmark value. Its value is always less than the first preset risk threshold. It is used to determine whether the driving condition corresponding to the current editing request is in an absolutely safe state without collision risks. This threshold is a safety release threshold value, used to distinguish between completely safe conditions and critical conditions with potential risks. It is the basis for judgment directly through editing requests and release parameter modification operations. For example, during the low-speed parking parameter editing process, when the collision risk quantification value is lower than the second preset risk threshold, it indicates that the vehicle trajectory after the current editing has no close-range collision risks throughout the entire process, and various error disturbances will not cause safety risks. The editing request can be directly accepted and executed.

[0208] The first decision result is a standardized decision instruction output for high-risk editing scenarios. Specifically, it rejects the driving task editing request initiated by the current user. This decision result corresponds to a high-risk scenario where the collision risk quantification value is greater than or equal to the first preset risk threshold. This means that the vehicle's driving trajectory corresponding to the currently edited parameters has an extremely high probability of collision and does not meet the execution conditions. By rejecting the editing, the immediate safety risk is avoided, and the underlying safety fallback is achieved. For example, after the user edits the parking speed and trajectory parameters, the calculated collision risk quantification value exceeds the high-risk threshold. The first decision result is generated, directly rejecting the parameter editing operation and maintaining the original safe driving parameters to continue executing the task.

[0209] The second decision result is a standardized decision instruction output for safe editing conditions. Specifically, it accepts the driving task editing request initiated by the current user. This decision result corresponds to an absolutely safe scenario where the collision risk quantification value is less than the second preset risk threshold. This means that the vehicle's driving trajectory corresponding to the current editing parameters is controllable throughout the entire process, with no collision hazards. Various system errors and environmental disturbances will not cause driving safety issues, and the editing request has the conditions to be fully executed. For example, if the user fine-tunes the parking trajectory smoothing parameters within the compliant and safe range, and the calculated collision risk quantification value is lower than the low-risk threshold, the second decision result is generated, confirming the validity of this editing request and updating the vehicle driving control parameters.

[0210] The third decision result is a standardized decision instruction output for critical risk conditions. Specifically, it postpones the execution of the current editing request and waits for manual confirmation. This decision result corresponds to an intermediate critical condition where the collision risk quantification value is between the second and first preset risk thresholds. It means that the driving risk corresponding to the current editing request is in a fuzzy controllable range, with neither extremely high immediate collision risk nor absolute safety standards, and there is potential uncertainty risk. It requires manual judgment to assist in the decision-making process. For example, if the user slightly increases the parking speed, the calculated collision risk quantification value is between the safe threshold and the high-risk threshold. The risk status is uncertain, and the third decision result is generated. The editing operation is temporarily not executed, and a manual interaction prompt is sent. The decision is made after the driver confirms the action.

[0211] For example, the system first retrieves the vehicle's fixed first and second preset risk thresholds. These two thresholds form a three-tiered standardized risk judgment interval: a high-risk interception interval, a critical risk pending confirmation interval, and a safe passage interval. The second preset risk threshold is always lower than the first preset risk threshold, ensuring no overlap or blind spots in the interval division. The real-time calculated collision risk quantification value is compared and judged level by level with the two preset thresholds, executing differentiated decision logic. When the collision risk quantification value is greater than or equal to the first preset risk threshold, it is determined that the driving trajectory corresponding to the current editing request has a high-risk collision risk, and the modified driving condition cannot meet the safety execution requirements. A first decision result of rejecting the editing request is then generated, intercepting the dynamic editing operation and maintaining the original vehicle driving parameters and trajectory execution logic, preventing the high-risk condition from taking effect. When the collision risk quantification value is less than the second preset threshold, it is determined that the driving trajectory corresponding to the current editing request is safe and controllable throughout, with no potential collision hazards, and that various error disturbances and environmental changes will not cause driving safety problems. A result of accepting the editing request is then generated. The second decision allows the current editing request to take effect immediately, updating the vehicle's driving task control parameters and trajectory planning logic. When the collision risk quantification value is greater than or equal to the second preset risk threshold and less than the first preset risk threshold, it is determined that the current operating condition is in a critical fuzzy range between safe and high-risk, and the risk is uncertain. The machine cannot make a safety judgment through autonomous logic. The third decision then generates a delayed editing request and waits for manual confirmation, temporarily freezing the effective process of the current editing parameters, delaying the update of driving task parameters, and triggering a human-machine interaction prompt. The subsequent editing release or rejection operation is then executed after the manual confirmation instruction is received.

[0212] By implementing the above embodiments, different risk thresholds are set for acceptance, rejection, and manual confirmation. Different decision results are output based on the collision risk quantification value. This allows vehicle driving task editing to move beyond a simple binary judgment method and adopt differentiated processing strategies for different risk levels. High-risk editing requests are directly rejected, low-risk editing requests are directly accepted, and medium-risk editing requests await manual confirmation. This balances vehicle operation safety with the flexibility of driving task editing, improving the rationality and reliability of vehicle driving task editing decisions.

[0213] Furthermore, as an implementation of the aforementioned method embodiments, this application also provides a vehicle driving task editing device for implementing the aforementioned method embodiments. This device embodiment corresponds to the aforementioned method embodiments. For ease of reading, this vehicle driving task editing device embodiment will not repeat the details of the aforementioned method embodiments one by one, but it should be understood that the device in this application embodiment can correspondingly implement all the contents of the aforementioned method embodiments. For example... Figure 2As shown, the vehicle driving task editing device 20 includes: a data acquisition unit 201, a trajectory generation unit 202, a collision timing unit 203, a risk quantification unit 204, and a decision determination unit 205. The data acquisition unit 201 acquires an editing request for the current driving task of the target vehicle and the current operating status information of the target vehicle. The trajectory generation unit 202 generates candidate trajectories for the target vehicle based on the editing request and the current operating status information. The candidate trajectories may include trajectory points of the target vehicle at multiple future times. The collision timing unit 203 performs continuous collision detection on the candidate trajectories to obtain collision distance timing data, which characterizes the distance between the trajectory point and the obstacle at each time point. The risk quantification unit 204 determines a collision risk quantification value for the candidate trajectory based on the collision distance timing data and a preset error uncertainty parameter. The preset error uncertainty parameter may include at least one of positioning error, perception error, control execution error, and model prediction error. The decision determination unit 205 determines a decision result for the editing request based on the collision risk quantification value.

[0214] In some embodiments, the risk quantification unit 204 is further configured to determine the equivalent distance variance corresponding to each moment in the collision distance time series data based on a preset error uncertainty parameter; determine the instantaneous collision probability at each moment based on the collision distance time series data and the equivalent distance variance using a preset normal cumulative distribution function to obtain a collision probability sequence; and determine the collision risk quantification value based on the collision probability sequence.

[0215] In some embodiments, the risk quantification unit 204 is further configured to obtain the upper limit of the instantaneous collision probability in the collision probability sequence as the first risk metric; determine the second risk metric based on the instantaneous collision probability at each moment in the collision probability sequence; and perform a weighted summation of the first risk metric and the second risk metric to obtain the collision risk quantification value.

[0216] In some embodiments, the risk quantification unit 204 is further configured to perform time-weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric; and to perform a preset exponential decay function operation on the cumulative risk metric to obtain a second risk metric.

[0217] In some embodiments, the risk quantification unit 204 is further configured to obtain scene parameters of the target vehicle, wherein the scene parameters include at least one of parking space type, ambient light level and road surface adhesion coefficient; based on the preset correspondence between scene parameters and time weight levels, determine a target time weight sequence that matches the scene parameters, wherein each moment in the target time weight sequence corresponds to a time weight value; based on the target time weight sequence, perform a weighted summation of the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric value.

[0218] In some embodiments, the risk quantification unit 204 is further configured to: if the positioning error, perception error, control execution error, and model prediction error are independent of each other, then solve for the sum of squares of the preset standard deviations corresponding to each error to obtain the equivalent distance variance at each time step; if at least two of the positioning error, perception error, control execution error, and model prediction error are correlated, then obtain the preset covariance matrix corresponding to each error source, and calculate the joint variance of each error source based on the preset covariance matrix to obtain the equivalent distance variance at each time step.

[0219] In some embodiments, the trajectory generation unit 202 is further configured to perform hard safety constraint screening on the editing request based on the current vehicle speed and obstacle information in the current operating status information, and obtain a safety screening result; if the safety screening result characterization is passed, then based on the current vehicle state in the current operating status information, a dynamic reachability analysis is performed to generate a candidate trajectory.

[0220] In some embodiments, the trajectory generation unit 202 is further configured to determine a safe stopping distance based on a preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the editing request; determine the difference between the current nearest obstacle distance in the obstacle information and the safe stopping distance as a safety margin; and generate a safety screening result indicating failure if the safety margin is less than a preset safety threshold.

[0221] In some embodiments, the trajectory generation unit 202 is further configured to determine the larger of the current vehicle speed and the target vehicle speed as the vehicle evaluation speed; determine the delayed travel distance by multiplying the vehicle evaluation speed by a preset system delay duration, wherein the preset system delay duration characterizes the response time required from triggering the braking condition to the target vehicle generating braking deceleration; determine the vehicle braking distance based on the ratio of the square of the current vehicle speed to twice the preset maximum braking deceleration; and determine the safe stopping distance by summing the delayed travel distance and the vehicle braking distance.

[0222] In some embodiments, the preset system delay duration is determined as follows: the system delay duration is obtained by summing the perception delay duration, decision delay duration, control command transmission delay duration, and actuator response delay duration; wherein, the perception delay duration represents the time interval from data acquisition by the environmental sensor to the output of structured perception results, the decision delay duration represents the time interval from receiving the perception results to generating control commands, the control command transmission delay duration represents the time interval from the transmission of control commands to the actuator via the communication bus, and the actuator response delay duration represents the time interval from the actuator receiving the command to starting to generate braking action.

[0223] In some embodiments, the trajectory generation unit 202 is further configured to, if the safety screening result characterization is passed, perform constraint consistency solution based on the current task sequence in the editing request and the current running status information to obtain constraint solution information; and perform dynamic reachability analysis based on the constraint solution information and the current vehicle state in the current running status information to generate candidate trajectories.

[0224] In some embodiments, the trajectory generation unit 202 is further configured to perform constraint parsing on the editing request to obtain incremental constraints; extract tasks from the current task sequence to obtain at least one task constraint, wherein the task constraint includes at least one of precondition constraints, postcondition constraints, temporal dependency constraints, and safety boundary constraints; merge the incremental constraints and task constraints to obtain a constraint set; perform consistency determination on the constraint set to obtain a determination result; if the determination result is inconsistent, determine the conflict terms between the incremental constraints and task constraints, and solve the minimum modification set of the constraint set based on preset relaxation rules, using the minimum modification set as constraint solution information, wherein the minimum modification set is a set of parameter corrections that can eliminate conflict terms; if the determination result is consistent, generate feasible assignments as constraint solution information, wherein the feasible assignments are a set of control parameter values ​​that satisfy all constraint conditions.

[0225] In some embodiments, the trajectory generation unit 202 is further configured to: if the constraint solution information is a feasible assignment, obtain the first control parameter in the feasible assignment, and determine the first control input boundary based on the first control parameter and the current vehicle state; if the constraint solution information is a minimum modification set, obtain the second control parameter in the minimum modification set, and determine the second control input boundary based on the second control parameter and the current vehicle state; take the current vehicle state as the initial state, and within the first control input boundary or the second control input boundary, perform forward reachable state propagation through a preset vehicle model to obtain a reachable state set, wherein the reachable state set represents the continuous spatial state envelope region that the target vehicle can reach at each time under the constraints of the first control input boundary or the second control input boundary; if the reachable state set meets the preset dynamic constraint conditions, select a nominal control sequence from the corresponding control input boundary for forward simulation to obtain a candidate trajectory, wherein the preset dynamic constraint conditions include at least one of lateral acceleration constraint, yaw rate constraint and minimum turning radius constraint, and the nominal control sequence is a set of time-series control commands that satisfy the preset dynamic constraint conditions within the first control input boundary or the second control input boundary.

[0226] In some embodiments, the collision timing unit 203 is further configured to acquire the obstacle prediction trajectory; perform time grid alignment processing on the candidate trajectory and the obstacle prediction trajectory according to a preset time step to obtain the aligned candidate trajectory time sequence and the aligned obstacle prediction trajectory time sequence; perform geometric expansion operation on the vehicle outline of the target vehicle and the obstacle outline corresponding to the obstacle prediction trajectory by a preset safe expansion distance to obtain the vehicle expansion region sequence and the obstacle expansion region sequence; for each aligned time, determine the minimum distance between the vehicle expansion region and the obstacle expansion region corresponding to the time to obtain the minimum distance sequence; determine the intervals in the minimum distance sequence that are continuously less than a preset distance threshold as collision intervals; and determine the minimum distance sequence and the collision intervals as collision distance time series data.

[0227] In some embodiments, the collision timing unit 203 is further configured to perform intersection detection on the vehicle expansion region and the obstacle expansion region to obtain the minimum detection distance between the vehicle expansion region and the obstacle expansion region; if the minimum detection distance is a positive value, the minimum detection distance is determined as the minimum distance; if the minimum detection distance is zero or negative, the penetration depth between the vehicle expansion region and the obstacle expansion region is calculated by the expansion polyhedron algorithm, and the minimum distance is determined based on the penetration depth.

[0228] In some embodiments, the decision determination unit 205 is further configured to generate a first decision result of rejecting the editing request if the collision risk quantification value is greater than or equal to a first preset risk threshold; generate a second decision result of accepting the editing request if the collision risk quantification value is less than a second preset risk threshold, wherein the second preset risk threshold is less than the first preset risk threshold; and generate a third decision result of delaying the editing request and waiting for manual confirmation if the collision risk quantification value is greater than or equal to the second preset risk threshold and less than the first preset risk threshold.

[0229] This application also provides a computer-readable storage medium storing computer-executable instructions or computer programs, which, when executed by a processor, will cause the processor to perform any step of the vehicle driving task editing method provided in this application.

[0230] In some embodiments, the computer-readable storage medium may be a random access memory (RAM), a read-only memory (ROM), flash memory, a magnetic surface memory, an optical disc, or a compact disc read-only memory (CD-ROM); or it may be a variety of devices that include one or any combination of the above-mentioned memories.

[0231] In some embodiments, computer-executable instructions may take the form of programs, software, software modules, scripts, or code, written in any form of programming language (including compiled or interpreted languages, or declarative or procedural languages), and may be deployed in any form, including as stand-alone programs or as modules, components, subroutines, or other units suitable for use in a computing environment.

[0232] In some embodiments, computer-executable instructions may, but do not necessarily, correspond to files in a file system, and may be stored as part of a file that holds other programs or data, for example, in one or more scripts in a HyperText Markup Language (HTML) document, in a single file dedicated to the program in question, or in multiple co-located files (e.g., files that store one or more modules, subroutines, or code sections).

[0233] like Figure 3 As shown, this application also provides a vehicle 30, including a memory 310, a processor 320, and a computer program 311 stored in the memory 310 and executable on the processor. When the processor 320 executes the computer program 311, it implements any step of the above-described vehicle driving task editing method.

[0234] This application also provides a computer program product comprising a computer program or computer-executable instructions stored in a computer-readable storage medium. A vehicle's processor reads the computer program or computer-executable instructions from the computer-readable storage medium and executes the computer program or computer-executable instructions, causing the vehicle to perform any step of the vehicle driving task editing method described above.

[0235] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such 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 this application.

Claims

1. A method for editing vehicle driving tasks, characterized in that, include: Obtain the edit request for the current driving task of the target vehicle and the current operating status information of the target vehicle; Based on the edit request and the current running status information, a candidate trajectory of the target vehicle is generated, wherein the candidate trajectory includes trajectory points of the target vehicle at multiple future time points; Continuous collision detection is performed on the candidate trajectory to obtain collision distance time series data, wherein the collision distance time series data is used to characterize the distance between the trajectory point and the obstacle at each moment; Based on the collision distance time series data and the preset error uncertainty parameter, a collision risk quantification value for the candidate trajectory is determined, wherein the preset error uncertainty parameter includes at least one of positioning error, perception error, control execution error and model prediction error; Based on the collision risk quantification value, a decision result is determined for the edit request.

2. The vehicle driving task editing method according to claim 1, characterized in that, The step of determining the collision risk quantification value for the candidate trajectory based on the collision distance time series data and preset error uncertainty parameters includes: Based on the preset error uncertainty parameter, the equivalent distance variance corresponding to each moment in the collision distance time series data is determined; Based on the collision distance time series data and the equivalent distance variance, the instantaneous collision probability at each moment is determined using a preset normal cumulative distribution function, thus obtaining a collision probability sequence. The collision risk quantification value is determined based on the collision probability sequence.

3. The vehicle driving task editing method according to claim 2, characterized in that, Determining the collision risk quantification value based on the collision probability sequence includes: Obtain the upper limit of the instantaneous collision probability in the collision probability sequence as the first risk metric; A second risk metric is determined based on the instantaneous collision probability at each moment in the collision probability sequence. The collision risk quantification value is obtained by weighted summation of the first risk metric value and the second risk metric value.

4. The vehicle driving task editing method according to claim 3, characterized in that, The step of determining the second risk metric based on the instantaneous collision probability at each moment in the collision probability sequence includes: The instantaneous collision probabilities at each moment in the collision probability sequence are time-weighted and summed to obtain the cumulative risk metric. The second risk metric is obtained by performing a preset exponential decay function operation on the cumulative risk metric.

5. The vehicle driving task editing method according to claim 4, characterized in that, The vehicle driving task editing method also includes: Obtain the scene parameters of the target vehicle, wherein the scene parameters include at least one of parking space type, ambient light level and road surface adhesion coefficient; Based on the preset correspondence between scene parameters and time weight levels, a target time weight sequence matching the scene parameters is determined, wherein each moment in the target time weight sequence corresponds to a time weight value. The step of performing time-weighted summation on the instantaneous collision probabilities at each moment in the collision probability sequence to obtain a cumulative risk metric includes: Based on the target time weight sequence, the instantaneous collision probabilities at each moment in the collision probability sequence are weighted and summed to obtain the cumulative risk metric.

6. The vehicle driving task editing method according to claim 2, characterized in that, The step of determining the equivalent distance variance corresponding to each moment in the collision distance time series data based on the preset error uncertainty parameter includes: If the positioning error, the perception error, the control execution error, and the model prediction error are independent of each other, the sum of squares of the preset standard deviations corresponding to each error is calculated to obtain the equivalent distance variance corresponding to each time step. If at least two of the positioning error, the perception error, the control execution error, and the model prediction error are correlated, then a preset covariance matrix corresponding to each error source is obtained, and the joint variance of each error source is calculated based on the preset covariance matrix to obtain the equivalent distance variance corresponding to each time step.

7. The vehicle driving task editing method according to claim 1, characterized in that, The step of generating candidate trajectories for the target vehicle based on the edit request and the current running status information includes: Based on the current vehicle speed and obstacle information in the current operating status information, a hard safety constraint screening is performed on the edit request to obtain the safety screening result; If the safety screening result is satisfactory, then based on the current vehicle status in the current operating status information, a dynamic reachability analysis is performed to generate the candidate trajectory.

8. The vehicle driving task editing method according to claim 7, characterized in that, Based on the current vehicle speed and obstacle information in the current operating status information, the edit request is subjected to hard safety constraint screening to obtain safety screening results, including: The safe stopping distance is determined based on the preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request. The difference between the current nearest obstacle distance in the obstacle information and the safe parking distance is determined as the safety margin; If the security margin is less than the preset security threshold, a security screening result indicating failure is generated.

9. The vehicle driving task editing method according to claim 8, characterized in that, The process of determining a safe stopping distance based on a preset maximum braking deceleration, the current vehicle speed in the current operating status information, and the target vehicle speed in the edit request includes: The larger of the current vehicle speed and the target vehicle speed is determined as the vehicle evaluation speed; The product of the vehicle's assessed speed and the preset system delay duration is determined as the delayed travel distance, wherein the preset system delay duration represents the response time required from the triggering of the braking condition to the target vehicle generating braking deceleration; The vehicle braking distance is determined based on the ratio of the square of the current vehicle speed to twice the preset maximum braking deceleration. The sum of the delayed driving distance and the vehicle braking distance is determined as the safe stopping distance.

10. The vehicle driving task editing method according to claim 9, characterized in that, The preset system delay duration is determined in the following way: The system delay time is obtained by summing the sensing delay time, decision delay time, control command transmission delay time, and actuator response delay time; wherein, the sensing delay time represents the time interval from data acquisition by the environmental sensor to the output of structured sensing results, the decision delay time represents the time interval from receiving the sensing results to generating control commands, the control command transmission delay time represents the time interval from the transmission of control commands to the actuator via the communication bus, and the actuator response delay time represents the time interval from the actuator receiving the command to starting to generate braking action.

11. The vehicle driving task editing method according to claim 7, characterized in that, If the safety screening result is satisfactory, then based on the current vehicle status in the current operating status information, a dynamic reachability analysis is performed to generate the candidate trajectory, including: If the security screening result is satisfactory, then constraint consistency is solved based on the editing request and the current task sequence in the current running status information to obtain constraint solution information. Based on the constraint solution information and the current vehicle state in the current operating state information, a dynamic reachability analysis is performed to generate the candidate trajectory.

12. The vehicle driving task editing method according to claim 11, characterized in that, The constraint consistency solution is performed based on the edit request and the current task sequence in the current running status information to obtain constraint solution information, including: The edit request is parsed to obtain incremental constraints; Task extraction is performed on the current task sequence to obtain at least one task constraint, wherein the task constraint includes at least one of the following: precondition constraint, postcondition constraint, temporal dependency constraint, and safety boundary constraint. The incremental constraints are merged with the task constraints to obtain a constraint set; A consistency determination is performed on the constraint set to obtain the determination result; If the determination result is inconsistent, then the conflict term between the incremental constraint and the task constraint is determined, and the minimum modification set of the constraint set is solved based on the preset relaxation rule. The minimum modification set is used as the constraint solution information, wherein the minimum modification set is a set of parameter corrections that can eliminate the conflict term. If the determination result is consistent, then a feasible assignment is generated as the constraint solution information, wherein the feasible assignment is a set of control parameter values ​​that satisfy all constraint conditions.

13. The vehicle driving task editing method according to claim 12, characterized in that, The step of performing dynamic reachability analysis based on the constraint solution information and the current vehicle state in the current operating state information to generate the candidate trajectory includes: If the constraint solution information is the feasible assignment, then the first control parameter in the feasible assignment is obtained, and the first control input boundary is determined based on the first control parameter and the current vehicle state; If the constraint solution information is the minimum modification set, then the second control parameters in the minimum modification set are obtained, and the second control input boundary is determined based on the second control parameters and the current vehicle state; Taking the current vehicle state as the initial state, within the first control input boundary or the second control input boundary, forward reachable state propagation is performed through a preset vehicle model to obtain a set of reachable states. The set of reachable states represents the continuous spatial state envelope region that the target vehicle can reach at each time under the constraints of the first control input boundary or the second control input boundary. If the set of reachable states meets the preset dynamic constraints, a nominal control sequence is selected from the corresponding control input boundary for forward simulation to obtain the candidate trajectory. The preset dynamic constraints include at least one of lateral acceleration constraints, yaw rate constraints, and minimum turning radius constraints. The nominal control sequence is a set of time-series control commands that satisfy the preset dynamic constraints within the first control input boundary or the second control input boundary.

14. The vehicle driving task editing method according to claim 1, characterized in that, The step of performing continuous collision detection on the candidate trajectory to obtain collision distance time-series data includes: Obtain the predicted trajectory of obstacles; The candidate trajectory and the obstacle prediction trajectory are aligned by time grid according to a preset time step to obtain the aligned candidate trajectory time series and the aligned obstacle prediction trajectory time series. For the vehicle profile of the target vehicle and the obstacle profile corresponding to the predicted obstacle trajectory, perform geometric expansion operation with a preset safe expansion distance to obtain the vehicle expansion region sequence and the obstacle expansion region sequence. For each aligned time point, determine the minimum distance between the vehicle's expansion region and the obstacle's expansion region corresponding to that time point to obtain a minimum distance sequence; The intervals in the minimum distance sequence that are consecutively less than a preset distance threshold are defined as collision intervals; The minimum distance sequence and the collision interval are determined as the collision distance time series data.

15. The vehicle driving task editing method according to claim 14, characterized in that, The minimum distance between the vehicle's expansion region and the obstacle's expansion region at the determined time includes: Intersection detection is performed on the vehicle expansion region and the obstacle expansion region to obtain the minimum detection distance between the vehicle expansion region and the obstacle expansion region; If the minimum detection distance is a positive value, then the minimum detection distance is determined as the minimum distance; If the minimum detection distance is zero or negative, the penetration depth between the vehicle's expansion region and the obstacle's expansion region is calculated using the dilatation polyhedron algorithm, and the minimum distance is determined based on the penetration depth.

16. The vehicle driving task editing method according to claim 1, characterized in that, The step of determining the decision result for the edit request based on the collision risk quantification value includes: If the collision risk quantification value is greater than or equal to the first preset risk threshold, a first decision result of rejecting the editing request is generated; If the collision risk quantification value is less than the second preset risk threshold, a second decision result accepting the editing request is generated, wherein the second preset risk threshold is less than the first preset risk threshold; If the collision risk quantification value is greater than or equal to the second preset risk threshold and less than the first preset risk threshold, a third decision result is generated to postpone the editing request and wait for manual confirmation.

17. A vehicle driving task editing device, characterized in that, include: The data acquisition unit is used to acquire the editing request for the current driving task of the target vehicle and the current operating status information of the target vehicle; The trajectory generation unit is used to generate a candidate trajectory of the target vehicle based on the edit request and the current running status information, wherein the candidate trajectory includes trajectory points of the target vehicle at multiple future time points; The collision timing unit is used to perform continuous collision detection on the candidate trajectory to obtain collision distance timing data, wherein the collision distance timing data is used to characterize the distance between the trajectory point and the obstacle at each moment; The risk quantification unit is used to determine the collision risk quantification value for the candidate trajectory based on the collision distance time series data and the preset error uncertainty parameter, wherein the preset error uncertainty parameter includes at least one of positioning error, perception error, control execution error and model prediction error; The decision-making unit is used to determine the decision result for the edit request based on the collision risk quantification value.

18. A vehicle comprising: The memory and processor are characterized in that the processor, when executing a computer program stored in the memory, implements the steps of the vehicle driving task editing method as described in any one of claims 1 to 16.

19. A computer-readable storage medium having stored thereon computer-executable instructions or a computer program, characterized in that, When the computer-executable instructions or the computer program are executed by a processor, the steps of the vehicle driving task editing method as described in any one of claims 1 to 16 are implemented.

20. A computer program product comprising a computer program or computer-executable instructions, characterized in that, When the computer program or computer-executable instructions are executed by the processor, the steps of the vehicle driving task editing method as described in any one of claims 1 to 16 are implemented.

Images