Vehicle trajectory generation method, device, equipment and storage medium

Abstract

Embodiments of the present application provide a vehicle driving trajectory generation method, device and equipment, and a storage medium, relating to the field of intelligent driving. The method comprises: obtaining a historical driving trajectory of a vehicle and trajectory condition information; performing noise adding processing on the historical driving trajectory to obtain an initial noise trajectory; obtaining a first expected value corresponding to the initial noise trajectory under the trajectory condition information, and updating the initial noise trajectory based on the first expected value to obtain an intermediate noise trajectory; and performing a reverse iterative denoising process based on the intermediate noise trajectory and the trajectory condition information to obtain a target driving trajectory.

Description

Technical Field

[0001] This application relates to the field of intelligent driving, and in particular to a method, apparatus, device and storage medium for generating vehicle driving trajectory. Background Technology

[0002] Autonomous driving systems typically continuously perceive surrounding elements, understand scene semantics, and predict the behavior of traffic participants in dynamic traffic environments to generate planned trajectories that meet the requirements of the driving task. This process usually operates in a closed-loop state, meaning that the autonomous driving system regenerates the planned trajectory based on the latest environmental information at each moment and drives the vehicle to execute it through the controller.

[0003] In related technologies, learning-based planning methods, such as imitation learning or behavior replication, generate planned trajectories. These methods determine the planned trajectory by learning the state-action mapping relationship from expert driving data.

[0004] However, the above methods are insufficient to fully model the complex multimodal decision-making characteristics that are prevalent in real driving. When the actual scenario deviates from the distribution of the training data, it is difficult to guarantee the safety and quality of the planned trajectory. Summary of the Invention

[0005] This application provides a method, apparatus, device, and storage medium for generating vehicle driving trajectories. The technical solution is as follows: In one aspect, a method for generating a vehicle's driving trajectory, the method comprising: The historical driving trajectory and trajectory condition information of the vehicle are obtained, and the trajectory condition information is used to indicate the driving status of the vehicle and its neighboring vehicles respectively. The historical driving trajectory is subjected to noise processing to obtain an initial noise trajectory; Obtain the first expected value corresponding to the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory; Based on the intermediate noise trajectory and the trajectory condition information, a reverse iterative denoising process is performed to obtain the target driving trajectory.

[0006] On the other hand, a vehicle trajectory generation device includes: The acquisition module is used to acquire the historical driving trajectory and trajectory condition information of the vehicle, wherein the trajectory condition information is used to indicate the driving status of the vehicle and its neighboring vehicles respectively; The processing module is used to perform noise addition processing on the historical driving trajectory to obtain an initial noise trajectory; The acquisition module is further configured to acquire the first expected value corresponding to the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory; The denoising module is used to perform a reverse iterative denoising process based on the intermediate noise trajectory and the trajectory condition information to obtain the target driving trajectory.

[0007] On the other hand, a computer device is provided, the computer device including a processor and a memory, the memory storing at least one program, the at least one program being loaded and executed by the processor to implement the vehicle trajectory generation method as described above.

[0008] On the other hand, a computer-readable storage medium is provided, wherein at least one segment is stored in the storage medium, the at least one segment being loaded and executed by a processor to implement the vehicle trajectory generation method as described above.

[0009] On the other hand, a computer program product or computer program is provided, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the method for generating a vehicle driving trajectory as described above.

[0010] The beneficial effects of the technical solutions provided in this application include at least the following: An initial noisy trajectory is obtained by adding noise to historical driving trajectories. A first expected value is then determined for this initial noisy trajectory under the given trajectory conditions. This first expected value is used to update the initial noisy trajectory, resulting in an intermediate noisy trajectory. The intermediate noisy trajectory is then applied to subsequent reverse iterative denoising steps to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noisy trajectory to evolve with the noise level. This approach maintains the diversity of the initial noisy trajectory's distribution in the early stages and enhances the convergence of high-value areas in the later stages, ensuring the target driving trajectory can handle complex traffic scenarios and further improving the driving experience of the navigation path. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1This is a structural block diagram of a computer system provided in an exemplary embodiment of this application; Figure 2 This is a flowchart of a method for generating a vehicle driving trajectory provided in an exemplary embodiment of this application; Figure 3 This is a flowchart of a method for generating a vehicle driving trajectory provided in another exemplary embodiment of this application; Figure 4 This is a flowchart of a method for generating a vehicle driving trajectory provided in yet another exemplary embodiment of this application; Figure 5 This is a flowchart illustrating the training process of a differentiable value function provided in an exemplary embodiment of this application; Figure 6 This is a schematic diagram illustrating the encoding of trajectory condition information provided in an exemplary embodiment of this application; Figure 7 This is a flowchart illustrating the training process of a preset diffusion network provided in an exemplary embodiment of this application; Figure 8 This is a schematic flowchart of a directional denoising sampling process using DPM-Solver++ provided in an exemplary embodiment of this application; Figure 9 This is a schematic diagram of an optimized iLQR backend structure provided in an exemplary embodiment of this application; Figure 10 This is a schematic diagram illustrating the composition of trajectory optimization cost terms and constraint terms provided in an exemplary embodiment of this application; Figure 11 This is a structural block diagram of a vehicle trajectory generation device provided in an exemplary embodiment of this application; Figure 12 A structural block diagram of a computer device provided in an exemplary embodiment of this application is shown. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0014] In this application, the terms "first" and "second" are used to distinguish between identical or similar items that have essentially the same function. It should be understood that there is no logical or temporal dependency between "first" and "second", nor is there any limitation on the quantity or execution order.

[0015] like Figure 1 As shown, Figure 1 This is a structural block diagram of a computer system provided in an exemplary embodiment of this application. Based on this structural block diagram, the execution process of the vehicle trajectory generation method provided in this embodiment will be described. This computer system is applied to a vehicle 10, and the following process is described using any controller within the vehicle 10 as the execution entity.

[0016] Optionally, the vehicle 10 can be at least one of the following: a gasoline-powered vehicle, an electric vehicle, a hybrid vehicle, a fuel cell vehicle, or a solar-powered vehicle, wherein a hybrid vehicle is a combination of a gasoline-powered vehicle and an electric vehicle.

[0017] Optionally, the vehicle 10 may include, but is not limited to, a powertrain system, a chassis system, a body system, and electrical equipment. The powertrain system primarily provides driving force for the vehicle 10; the chassis system is primarily responsible for support, transmission, driving, steering, and braking; the body system primarily carries passengers and goods; and the electrical equipment is primarily responsible for power supply and lighting, etc.

[0018] Vehicle 10 integrates multiple controllers, each responsible for different vehicle functions. Illustratively, vehicle 10 includes, but is not limited to, a powertrain controller, chassis domain controller, autonomous driving domain controller, cockpit domain controller, and body domain controller. The powertrain controller is the central computing unit responsible for energy management, power distribution, and efficiency optimization of the vehicle's powertrain system, such as the vehicle controller, motor controller, battery management system, and engine control unit. The chassis domain controller is the central computing unit responsible for vehicle 10's dynamic control, safety redundancy, and ride quality; it is the vehicle's motion hub, such as the air suspension controller, four-wheel drive controller, and EPS. The autonomous driving domain controller is the central computing unit responsible for environmental perception fusion, path planning, and vehicle 10 decision execution; it is the intelligent brain of vehicle 10. The cockpit domain controller is primarily responsible for human-machine interaction, infotainment, and driving experience; it is the digital cockpit brain of vehicle 10. The body domain controller is primarily responsible for the logic control, state management, and functional coordination of the body's electronic systems. In this embodiment, the autonomous driving domain controller executes the vehicle trajectory generation method provided in this embodiment.

[0019] In this embodiment, the autonomous driving domain controller has intelligent navigation functionality. Intelligent navigation is a navigation system based on artificial intelligence, high-precision positioning, real-time data fusion, and dynamic path planning technologies. Illustratively, the intelligent navigation system combines high-precision maps, traffic flow data, driving environment perception data, vehicle status data, etc., to analyze user intent and scene semantics, achieving a leap from executing commands to proactive suggestions.

[0020] Optionally, the autonomous driving domain controller acquires the vehicle's historical driving trajectory and trajectory condition information, wherein the trajectory condition information corresponds to the historical driving trajectory and is used to indicate the vehicle status of the vehicle and neighboring vehicles under the historical driving trajectory.

[0021] The autonomous driving domain controller adds noise to the historical driving trajectory to obtain an initial noise trajectory. It then determines a first expected value corresponding to the initial noise trajectory under the trajectory condition information and updates the initial noise trajectory based on this first expected value to obtain an intermediate noise trajectory.

[0022] The autonomous driving domain controller performs a reverse iterative denoising process based on intermediate noise trajectory and trajectory condition information to obtain the target driving trajectory. The target driving trajectory indicates the path planning from the origin to the destination, where the origin and destination are determined based on the driver's driving needs.

[0023] The aforementioned execution steps can be executed by other controllers within the vehicle 10, and this application does not limit them, such as any one of the following: vehicle controller, motor controller, battery management system, communication controller, information management controller, entertainment management controller, etc.

[0024] It is worth noting that the above interaction method is only an example. In other embodiments, the above method for generating the vehicle driving trajectory is implemented by the vehicle 10 and the server in coordination.

[0025] To illustrate, vehicle 10 receives an intelligent navigation request, which includes the origin and destination; vehicle 10 obtains the historical driving trajectory generated by vehicle 10 within a historical time period and the corresponding trajectory condition information. Vehicle 10 sends the historical driving trajectory and trajectory condition information to the server.

[0026] The server receives historical driving trajectories and trajectory condition information, adds noise to the historical driving trajectories to obtain an initial noisy trajectory; determines the first expected value corresponding to the initial noisy trajectory under the trajectory condition information; and updates the initial noisy trajectory based on the first expected value to obtain an intermediate noisy trajectory. Based on the intermediate noisy trajectory and trajectory condition information, a reverse iterative denoising process is performed to obtain the target driving trajectory, which is then sent to vehicle 10.

[0027] Optionally, the server can be a standalone physical server, a server cluster or distributed system consisting of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud security, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware servers, domain name services, security services, content delivery networks (CDNs), and big data and artificial intelligence platforms. In some embodiments, the server can also be implemented as a node in a blockchain system.

[0028] It should be noted that all information (including but not limited to historical driving trajectories and trajectory condition information), data (including but not limited to data used for analysis, stored data, and displayed data), and signals involved in this application have been authorized by the user or by all parties in full, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant regions. For example, the water level data and operational status data involved in this application were obtained with full authorization.

[0029] To further explain, this application can display a prompt interface, pop-up window, or output voice prompts before and during the collection of user-related data (e.g., historical driving trajectories, trajectory condition information, etc.). These prompt interfaces, pop-ups, or voice prompts are used to inform the user that their relevant data is being collected. This ensures that the application only begins the steps for collecting user-related data after receiving confirmation from the user regarding the prompt interface or pop-up window; otherwise (i.e., without receiving confirmation from the user), the steps for collecting user-related data end, meaning no user-related data is collected. In other words, all user data collected in this application is collected with the user's consent and authorization, and the collection, use, and processing of related user data must comply with the relevant laws, regulations, and standards of the relevant regions.

[0030] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0031] Based on the above, the method for generating vehicle driving trajectories provided in the embodiments of this application will be described. Figure 2 This is a flowchart illustrating a method for generating a vehicle driving trajectory according to an exemplary embodiment of this application. In this embodiment, the method is executed by a vehicle, such as... Figure 2 As shown, the method includes the following steps.

[0032] Step 200: Obtain the vehicle's historical driving trajectory and trajectory condition information.

[0033] In this embodiment, historical driving trajectory refers to the driving trajectory recorded by the vehicle within a historical time period. Historical driving trajectory includes the vehicle's historical location, navigation route, and distance traveled within the historical time period.

[0034] Optionally, the historical driving trajectory is a trajectory recorded within a historical time period based on the driver's navigation needs. In this embodiment, the historical driving trajectory can be a trajectory where the navigation needs are met, a trajectory where the navigation needs are not met, or a trajectory where route planning has been completed but navigation has not yet started.

[0035] To illustrate, at 21:00, the vehicle receives a navigation request from location a to location b. Based on this request, the vehicle generates a navigation route from location a to location b and follows this route to reach location b. The navigation route from location a to location b is considered the historical driving trajectory.

[0036] To illustrate, at 21:00, the vehicle receives a navigation request from location a to location b. Based on this request, the vehicle generates a navigation route from location a to location b. The vehicle is currently traveling along this navigation route. When acquiring historical driving trajectories, this navigation route can be considered as a historical driving trajectory.

[0037] As an illustration, a vehicle receives a navigation request from location a to location b at 21:00. The vehicle generates a navigation route from location a to location b based on the request, but the vehicle does not follow this route or has not yet started navigation. This navigation route can be considered part of the historical driving trajectory when acquiring historical driving records.

[0038] In another optional embodiment, the historical driving trajectory can also be the driving trajectory automatically recorded by the vehicle during its operation after power-on. Optionally, a dashcam device is installed in the vehicle to record the vehicle's status after power-on, including the driving route, driving speed, etc. That is, the historical driving trajectory of the vehicle is obtained through the dashcam device.

[0039] In this embodiment, trajectory condition information is used to indicate the vehicle status of the vehicle and its neighboring vehicles during the process of driving along a historical trajectory. The vehicle status includes driving speed, driving acceleration, vehicle size, vehicle category identifier, etc.

[0040] That is, the trajectory condition information includes the vehicle's historical vehicle status and the historical vehicle status of neighboring vehicles. The neighboring vehicles can be vehicles that are at a preset distance from the vehicle, or they can be communication vehicles that have established a communication connection with the vehicle. This application does not limit this.

[0041] In another optional embodiment, the trajectory condition information also includes the driving environment and navigation route conditions, wherein the driving environment is used to provide constraints on the driving road and static obstacles, and the driving environment includes, but is not limited to, a set of lane lines and a set of static objects, and the navigation route conditions are used to constrain the vehicle's global driving intention.

[0042] In this embodiment, the trajectory condition information includes the vehicle's current state constraint X0 and the historical states of neighboring vehicles. The set of lane polylines L, the set of static objects O, and the navigation route conditions G. That is, the trajectory condition information C = .

[0043] In this embodiment, the current state constraint X0 is used to provide initial boundary conditions for subsequent trajectory generation, including the vehicle's state vector at the current moment. and the state vector of at least one neighboring vehicle at the current time. That is, X0 = The vehicles include M adjacent vehicles. This refers to the state vector of the m-th neighboring vehicle at the current moment, where... , Used to indicate the real number space, with dimensions of , where M is a positive integer.

[0044] In another alternative embodiment, the state vector includes the vehicle's historical position and heading angle.

[0045] The vehicle's current state constraint X0 = (x, y, ... sin ), where (x, y) refers to the historical location coordinates of the vehicle. This refers to the vehicle's heading angle.

[0046] In this embodiment of the application, the historical state of neighboring vehicles Used to provide dynamic context for vehicle traffic interactions, including a sequence of historical states of M neighboring vehicles over the past L frames.

[0047] Optional, the historical state of the m-th neighboring vehicle. ,in, The historical state features include features such as historical location, heading angle, speed, vehicle size, and category.

[0048] Optional, a set of historical states of M neighboring vehicles. = .

[0049] In this embodiment, the set of lane breaklines L is represented using vectorization. During the vehicle's journey along its historical trajectory, there are R lane breaklines, with the r-th lane breakline consisting of P points. ,in, The set of lane polylines L includes semantic features such as the coordinates of the P-th point, lane direction, curvature, traffic light status, and speed limit information. , where P and R are both positive integers.

[0050] In this embodiment, the static object set O includes feature vectors of J static objects, where J is a positive integer. The feature vector of the j-th static object is represented as... The set of static objects O = .in, This includes information such as the coordinates, heading, size, and category of static objects.

[0051] In this embodiment of the application, the navigation route condition G includes a set of routes and lanes selected along historical driving trajectories. Where K represents the number of lanes on the route, and W represents the number of discrete points for each lane on the route. This indicates the dimension of the route lane features.

[0052] In this embodiment of the application, the current state of the aforementioned vehicle is constrained by X0, and the historical state of neighboring vehicles is constrained by X0. The lane polyline set L, the static object set O, and the navigation route conditions G are conditionally encoded to obtain driving condition information. In another optional embodiment, the vehicle's current state constraint X0 and the historical states of neighboring vehicles are used to... The driving condition information is obtained by fusing the lane polyline set L, the static object set O, and the navigation route conditions G. Condition encoding and fusing representation are common techniques in this field, and the specific encoding and fusing processes are not detailed in this embodiment.

[0053] Step 210: Perform noise processing on the historical driving trajectory to obtain the initial noise trajectory.

[0054] In this embodiment, noise schedule parameters are set, including noise steps, cumulative noise coefficient, single-step noise coefficient, noise variance coefficient, and noise scale. The cumulative noise coefficient is the square of the proportion of historical driving trajectory retained up to the current time step, the single-step noise coefficient is the square of the proportion of historical driving trajectory retained in the current time step, the noise variance coefficient is the intensity of the noise injected in the current time step, and the noise scale is the instantaneous noise intensity from the perspective of stochastic differential equations.

[0055] Optionally, the historical driving trajectory can be noise-added according to the noise schedule parameters to obtain the initial noise trajectory.

[0056] In an optional embodiment, an initialization process is performed on the initial noise trajectory to obtain a processed initial noise trajectory, wherein the initial noise trajectory can be obtained by sampling a standard Gaussian distribution.

[0057] Step 220: Obtain the first expected value of the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory.

[0058] Optionally, the first expected value of the initial noise trajectory under the trajectory condition information can be determined by a simulation evaluator, a rule scorer, or a preset preference model. That is, the initial noise trajectory and trajectory condition information are input into the simulation evaluator, the rule scorer, or the preset preference model to obtain the first expected value.

[0059] In this embodiment, a first expected value corresponding to the initial noise trajectory under trajectory condition information is determined by a trained differentiable value function. That is, the initial noise trajectory and trajectory condition information are input into the differentiable value function to obtain the first expected value.

[0060] The training process for the differentiable value function is as follows.

[0061] The sample noise trajectory, sample trajectory condition information, and the entire trajectory corresponding to the sample trajectory condition information are obtained, as well as the reward or reward value corresponding to the sample noise trajectory under the sample trajectory condition information. The reward or reward value is evaluated by the simulation evaluator, rule scorer, or preset preference model.

[0062] Establish an initial differentiable value function. The output of the initial differentiable value function is used to characterize the expected report of the sample noise trajectory under the sample trajectory condition information. It must satisfy the following expression, which can be found in Formula 1.

[0063] Formula 1: ; in, R refers to the initial differentiable value function, R refers to the reward or return value, and C refers to the sample trajectory condition information. It indicates the control sequence corresponding to the sample noise trajectory under the sample trajectory condition information. It refers to the network parameters of the value function, which is initially differentiable.

[0064] Optionally, the above-mentioned initial differentiable value function can be trained through supervised learning so that the loss function of the initial differentiable value function satisfies the following expression, as detailed in Formula 2.

[0065] Formula 2: ; In Formula 2, R refers to the output reward or output return value of the initial differentiable value function, while R refers to the reward or return value corresponding to the sample noise trajectory under the sample trajectory condition information.

[0066] Optionally, if the loss function satisfies Formula 2 above, then training of the initial differentiable value function should be stopped.

[0067] In another alternative embodiment, the initial differentiable value function can also output an advantage function. Or return forecast Among them, the dominance function This refers to the advantage value of a sample noise trajectory relative to other noise trajectories under the given sample trajectory conditions, where the sample noise trajectory and other noise trajectories are different planned trajectories corresponding to the same starting point and the same destination; return prediction. This refers to the performance of a sample noise trajectory under sample trajectory condition information.

[0068] In another alternative embodiment, the trained differentiable value function has the property of being differentiable with respect to the initial noise trajectory and outputs the value gradient of the initial noise trajectory for guiding the subsequent diffusion sampling process.

[0069] Step 230: Based on the intermediate noise trajectory and trajectory condition information, perform a reverse iterative denoising process to obtain the target driving trajectory.

[0070] Optionally, the intermediate noise trajectory is input into a preset diffusion network to obtain noise prediction. The noise prediction is used to indicate the noise information contained in the intermediate noise trajectory.

[0071] Based on intermediate noise trajectory, noise prediction, and trajectory condition information, a reverse iterative denoising process is performed to obtain the target driving trajectory.

[0072] Determine the drift function corresponding to the intermediate noise trajectory. The drift function is used to characterize the denoising direction of the evolution from the current time step to the next time step.

[0073] Optionally, the drift function is determined based on the noise schedule parameters in step 210 above.

[0074] Optionally, the drift function is determined by the relevant personnel based on the intermediate noise trajectory.

[0075] In the reverse iterative denoising process, the target driving trajectory is determined by the drift function, which can be found in the following three methods.

[0076] The first method involves performing explicit integration on the drift function to obtain the integration result, which is then used as the target driving trajectory. The determination of the target driving trajectory can be found in Formula 3 below.

[0077] Formula 3: ; In Formula 3, This represents the noisy trajectory within the current denoising time step. This refers to inferring the noisy trajectory from the previous denoising time step based on the current denoising time step. It refers to the interval corresponding to the current time step. This refers to the time stamp corresponding to the current denoising time step, and 'c' refers to the trajectory condition information. This refers to the noise prediction corresponding to the current denoising time step. When the reverse iteration process reaches k-1 equal to 0, the output result of Formula 3 above will be determined as the target driving trajectory.

[0078] The second method involves determining the first drift value corresponding to the drift function at the first time point within the current denoising time step, and determining the second drift value corresponding to the drift function at the second time point within the current denoising time step. Here, the first time point refers to the middle time point of the current denoising time step, and the second time point refers to the maximum time point within the current denoising time step.

[0079] A drift prediction is performed on the first drift value and the second drift value using a preset correction method to obtain the prediction result, and the prediction result is determined as the target driving trajectory. The determination of the target driving trajectory can be found in Formula 4 below.

[0080] Formula 4: ; In Formula 4, This represents the noisy trajectory within the current denoising time step. This refers to inferring the noisy trajectory from the previous denoising time step based on the current denoising time step. It refers to the interval corresponding to the current time step. This refers to the time stamp corresponding to the current denoising time step, and 'c' refers to the trajectory condition information. This refers to the noise prediction corresponding to the current denoising time step. It refers to the midpoint of the time interval. This refers to noise prediction at the intermediate time point. This is an intermediate prediction state. ( ) represents a second-order combination function. When the reverse iteration process reaches k-1 equal to 0, the output result of the above formula four is determined as the target driving trajectory.

[0081] The third method involves introducing multiple sampling points within the current denoising time step and combining the drift values ​​corresponding to the multiple sampling time points. The determination of the target driving trajectory can be found in Formula 5 below.

[0082] Formula 5: ; In Formula 5, This represents the noisy trajectory within the current denoising time step. This refers to inferring the noisy trajectory from the previous denoising time step based on the current denoising time step. It refers to the interval corresponding to the current time step. This refers to the time stamp corresponding to the current denoising time step, and 'c' refers to the trajectory condition information. This refers to the drift values ​​calculated at different points in time. ( ) represents a third-order combination function. When the reverse iteration process reaches k-1 equal to 0, the output result of the above formula 3 is determined as the target driving trajectory.

[0083] Optionally, in conjunction with the above formula, the denoising update of the target driving trajectory is repeated K times, where K is the denoising time step preset by relevant personnel, K is a positive integer, and the range of k is (K, ..., 1).

[0084] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0085] Based on the above, the method for generating vehicle driving trajectories provided in the embodiments of this application will be described. Figure 3 This is a flowchart of a method for generating a vehicle driving trajectory provided in another exemplary embodiment of this application. In this embodiment, the method is executed by a vehicle, such as... Figure 3 As shown, the method includes the following steps.

[0086] Step 300: Perform backend optimization on the target driving trajectory.

[0087] In this embodiment of the application, in order to further improve the performance of the target driving trajectory in terms of dynamic feasibility, constraint satisfaction, smoothness, comfort and control feasibility, the target driving trajectory output by the above embodiment is locally refined according to the preset dynamic constraints, and the optimized target driving trajectory is output. That is, the target driving trajectory is adjusted according to the preset dynamic constraints to obtain the adjusted target driving trajectory.

[0088] Optionally, preset dynamic constraints include reference trajectory proximity conditions.

[0089] Based on the vehicle's current status and trajectory conditions, a reference driving trajectory is determined. The current status includes the vehicle's speed, acceleration, heading, lane information, traffic light information corresponding to the lane, and information on obstacles encountered along the way. The reference driving trajectory is selected by relevant personnel from multiple candidate driving trajectories based on their experience.

[0090] In another optional embodiment, the starting point and destination of the target driving trajectory are determined, and multiple candidate driving trajectories that correspond one-to-one with the above starting point and destination are selected based on historical driving trajectories. That is, the multiple candidate driving trajectories are consistent with the starting point and destination of the target driving trajectory.

[0091] Using the differentiable value function described above, multiple reward values ​​corresponding to multiple candidate driving trajectories are determined, and the candidate driving trajectory with the largest reward value among the multiple reward values ​​is determined as the reference driving trajectory.

[0092] Optionally, based on the number of traffic lights waiting, the driving trajectory with the fewest traffic lights waiting can be selected from multiple candidate driving trajectories as the reference driving trajectory.

[0093] Optionally, based on the driving distance, the driving trajectory with the shortest driving distance can be selected from multiple candidate driving trajectories as the reference driving trajectory.

[0094] Optionally, the driving trajectory with the fewest obstacles among multiple candidate driving trajectories can be used as a reference driving trajectory, based on the number of obstacles.

[0095] Optionally, based on driving speed, the driving trajectory with the fastest driving speed is selected from multiple candidate driving trajectories as the reference driving trajectory. It should be noted that the driving speed of the vehicles in multiple candidate driving trajectories all meet the road speed limit requirements.

[0096] Determine the deviation between the reference driving estimate and the target driving trajectory.

[0097] A control deviation weight matrix is ​​generated using the deviation value. The target driving trajectory is then adjusted based on the control deviation weight matrix to obtain the adjusted target driving trajectory.

[0098] Optionally, preset dynamic constraints include control constraints.

[0099] Obtain the vehicle's longitudinal acceleration and front wheel steering angle.

[0100] Based on the pre-set first non-negative weight parameter, the pre-set second non-negative weight parameter, the longitudinal acceleration, and the front wheel angle, control adjustment parameters are generated.

[0101] The target driving trajectory is adjusted by using control and adjustment parameters to obtain the adjusted target driving trajectory.

[0102] In an optional embodiment, a first squared value of the longitudinal acceleration is determined, and a second squared value of the front wheel steering angle is determined. A first product of the first squared value and a first non-negative weighting parameter is determined, and a third product of the second squared value and a second non-negative weighting parameter is determined. The sum of the first and second products is determined, and this sum is used as a control adjustment parameter.

[0103] Optionally, preset dynamic constraints include ride comfort constraints.

[0104] Comfort adjustment parameters are generated based on the vehicle's longitudinal acceleration, lateral acceleration, longitudinal jerk, and steering change rate.

[0105] Optionally, the longitudinal acceleration cost is determined based on the longitudinal acceleration, the lateral comfort cost is determined based on the lateral acceleration, the longitudinal jerk cost is determined based on the longitudinal jerk, and the steering rate of change cost is determined based on the steering rate of change. The sum of the longitudinal and velocity costs, the lateral comfort cost, the longitudinal jerk cost, and the steering rate of change cost is determined as the comfort adjustment parameter.

[0106] The target driving trajectory is adjusted using comfort adjustment parameters to obtain the adjusted target driving trajectory.

[0107] Optionally, preset dynamic constraints include driving safety constraints.

[0108] Determine the distance between the vehicle and the obstacle.

[0109] Safety adjustment parameters are generated based on the difference between the distance and the preset safe distance threshold.

[0110] The target driving trajectory is adjusted using safety adjustment parameters to obtain the adjusted target driving trajectory.

[0111] All adjustment parameters mentioned in the embodiments of this application can be found in the following embodiments, and will not be described in detail here.

[0112] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0113] Based on the above, the method for generating vehicle driving trajectories provided in the embodiments of this application will be described. Figure 4 This is a flowchart of a method for generating a vehicle driving trajectory provided in another exemplary embodiment of this application. In this embodiment, the method is executed by a vehicle, such as... Figure 4 As shown, the method includes the following steps.

[0114] In this embodiment, a differentiable value function is trained based on offline data; a trajectory is generated based on a conditional diffusion model; a value gradient is introduced to guide the back-end optimization during the reverse denoising sampling process of the diffusion model; and an iterative linear quadratic regulator (iLQR) is further used to perform back-end optimization based on the diffusion output trajectory to obtain a final output trajectory that meets the dynamic constraints and trajectory quality requirements.

[0115] Step 400, offline training phase.

[0116] In this embodiment, an offline dataset is constructed. The offline dataset includes trajectory condition information, the corresponding trajectory or control sequence, and the reward value, reward value, or score value of the trajectory under the corresponding trajectory condition information. The trajectory condition information may include the current vehicle status, neighboring vehicle status, neighboring vehicle historical trajectory, map lane information, static obstacle information, navigation route information, and other scene constraint information.

[0117] Optionally, construct an offline dataset D (step 4000), which contains multiple sets of samples (C i , τ i R i ), where C i The trajectory condition information within the i-th sample group (including starting point constraints, neighboring vehicle history, map static features, and navigation routes, etc.); τ iR represents the entire trajectory or control sequence within the i-th sample group that corresponds to the conditional information; i Let be the reward / reward value of the trajectory within the i-th sample group under the conditional information. The reward / reward value can be obtained from the simulation evaluator, rule scorer, preference model or a combination thereof, where i is a positive integer.

[0118] Secondly, a differentiable value function is trained based on an offline dataset (step 4001). The differentiable value function is used to estimate the reward or preference rating corresponding to a trajectory under given conditions, and can calculate the gradient of the input trajectory to output value gradient information. The value gradient is used to characterize the adjustment direction of the trajectory towards higher value directions.

[0119] In the embodiments of this application, such as Figure 5 As shown, Figure 5 A flowchart illustrating the training process of a differentiable value function provided in an exemplary embodiment of this application is shown.

[0120] Step 500: Obtain offline samples.

[0121] Optionally, offline samples can be arbitrarily selected from the offline dataset D mentioned above. Illustratively, the offline sample is (C, τ, R), where C represents the trajectory condition information within the offline dataset D (including starting point constraints, neighbor vehicle history, map static features, and navigation routes); τ represents the entire trajectory or control sequence corresponding to the condition information in the offline dataset D; and R represents the reward / reward value of the trajectory under the condition information. This reward / reward value R can be obtained by a simulation evaluator, a rule scorer, a preference model, or a combination thereof. The reward / reward value R is evaluated by the simulation evaluator, the rule scorer, or a preset preference model.

[0122] Step 510: Jointly encode the trajectory condition information and the future trajectory.

[0123] Optionally, the trajectory condition information includes the vehicle's current state constraint X0 and the historical states of neighboring vehicles. The set of lane lines L, the set of static objects O, and the navigation route conditions G.

[0124] Constraining the vehicle's current state X0 and the historical states of neighboring vehicles. The lane polyline set L, the static object set O, and the navigation route conditions G are encoded to obtain the trajectory condition information C.

[0125] like Figure 6 As shown, Figure 6This diagram illustrates the trajectory condition information encoding provided in an exemplary embodiment of this application. The current state constraint 600 of the vehicle, the historical states 610 of neighboring vehicles, the lane polyline set 620, the static object set 630, and the navigation route conditions 640 are obtained and conditionally encoded or fused to obtain trajectory condition information. After subsequently predicting the future trajectory of the vehicle itself and the future trajectories of M neighboring vehicles, the future trajectory of the vehicle itself and the future trajectories of the M neighboring vehicles are encoded to obtain a future trajectory tensor, which is stored together with the trajectory condition information.

[0126] Step 520: Train the initial differentiable value function.

[0127] Establish an initial differentiable value function. The output of the initial differentiable value function is used to characterize the expected report of the sample noise trajectory under the sample trajectory condition information. It must satisfy the following expression, which can be found in Formula 1.

[0128] Formula 1: ; in, R refers to the initial differentiable value function, R refers to the reward or return value, and C refers to the sample trajectory condition information. It indicates the control sequence corresponding to the sample noise trajectory under the sample trajectory condition information. It refers to the network parameters of the value function, which is initially differentiable.

[0129] Optionally, the above-mentioned initial differentiable value function can be trained through supervised learning so that the loss function of the initial differentiable value function satisfies the following expression, as detailed in Formula 2.

[0130] Formula 2: ; In Formula 2, R refers to the output reward or output return value of the initial differentiable value function, while R refers to the reward or return value corresponding to the sample noise trajectory under the sample trajectory condition information.

[0131] Step 530: Stop training the initial differentiable value function and obtain the differentiable value function.

[0132] Optionally, if the loss function satisfies Formula 2 above, then training of the initial differentiable value function should be stopped.

[0133] Step 540: Optimize the differentiable value function.

[0134] Optionally, the trained differentiable value function has the property of being differentiable with respect to the noise trajectory, and can output the value gradient corresponding to the noise trajectory, which can be used to guide the subsequent diffusion sampling process.

[0135] In the embodiments of this application, the value gradient is determined by the gradient output interface within the differentiable value function.

[0136] Optionally, after steps 500 to 540 above, a differentiable value function is obtained.

[0137] The value-guided diffusion sampling process mentioned in step 540 above can be found in the following content.

[0138] Optionally, the original denoising direction is output at each denoising time step t during backsampling. The original denoising direction score is corrected based on the value gradient.

[0139] Optionally, the noise trajectory and trajectory condition information C corresponding to the current denoising time step are input into a preset diffusion network, and the preset diffusion network outputs the original denoising direction. The original denoising direction can be expressed as .

[0140] In another alternative embodiment, a preset diffusion network can output a noise prediction. Or clean trajectory estimation And predict noise by analyzing the relationship. Or clean trajectory estimation The estimated conversion is based on the original denoising direction.

[0141] Represent the trajectory corresponding to the current denoising time step. ,Will Input the trained differentiable value function to obtain the value gradient. .

[0142] And based on the value gradient, a guided denoising direction is constructed, that is, the guided denoising direction is expressed as: .in, The guiding coefficients are implemented as constants or functions that vary with time steps.

[0143] Optionally, through the above guidance method, the original denoising direction term ensures that the output target driving trajectory is consistent with the historical driving trajectory distribution, and the value gradient term causes the output target driving trajectory to shift along the direction of increasing reward, thereby achieving target alignment.

[0144] In another alternative embodiment, to improve guidance stability and avoid being misled by differentiable value functions, one or more of the following methods may be used for processing.

[0145] The first type, for g t The amplitude is clipped or normalized to fall within a preset range to avoid excessive guidance that could cause sampling divergence.

[0146] The second type, when Q When (C,τ_t) is below the threshold, bootstrapping is enabled to reduce unnecessary disturbances.

[0147] The third method involves setting a guiding coefficient λ. t As t changes, diversity is maintained in the early stages and optimization is enhanced in the later stages (or vice versa) to improve generation quality and stability.

[0148] The fourth type will reward value Q r With cost / risk value Q cost Combination serves as a guiding factor for the direction of denoising, that is, the direction of denoising after guidance. To achieve the dual goals of "optimization + safety".

[0149] The fifth method involves guiding sampling to generate multiple candidate trajectories, and then using Q... Candidate trajectories are scored and the best one is output to reduce the risk of single-guide failure.

[0150] Next, a pre-defined diffusion network is constructed and trained. This network learns the distribution of future trajectories given specific conditions and, during the inference phase, starts from the initial noise state and gradually recovers candidate trajectories through multiple reverse denoising steps. The diffusion model can output noise predictions, clean trajectory estimates, or equivalent score information.

[0151] The process of training the pre-defined diffusion network is as follows: Figure 7 As shown, Figure 7 A flowchart illustrating the training process of a preset diffusion network provided in an exemplary embodiment of this application is shown. Figure 7 A flowchart illustrating the application of a pre-defined diffusion network is also shown.

[0152] Step 700, Training Phase.

[0153] Obtain the sample trajectory, which includes the actual trajectory traveled by the vehicle (step 710).

[0154] Set the random sampling time step t and the Gaussian noise to be added subsequently (step 720).

[0155] Gaussian noise is added to the real trajectory using a forward noise addition method to obtain a noisy trajectory (step 730).

[0156] Obtain the sample trajectory condition information of the vehicle, and perform condition encoding and / or fusion representation on the sample trajectory condition information to obtain condition encoding information (step 740).

[0157] The conditional coding information and the noisy trajectory are input into the initial diffusion network (step 750), and the noise prediction is output. Or clean trajectory estimation (Step 760).

[0158] Determine the true trajectory and noise prediction during subsequent training. Or clean trajectory estimation The mean square error between them (step 770), if the mean square error satisfies Stop training the initial diffusion network , The network parameters are defined, and the initial diffusion network in this prediction process is determined as the pre-trained diffusion network. For specific application process, please refer to step 410 below.

[0159] Step 780, Reasoning Application Stage.

[0160] Optionally, obtain historical driving trajectories. Perform noise processing on the historical driving trajectories to obtain an initial noisy trajectory (step 790).

[0161] Optionally, the trajectory condition information is obtained (step 791), and an iterative reverse denoising process is performed based on the initial noisy trajectory and the trajectory condition information (792), and the target driving trajectory is finally output (793).

[0162] It should be noted that the iterative reverse denoising process requires the application of a pre-trained diffusion network, as detailed below.

[0163] In this embodiment, the clean trajectory predicted by the preset diffusion network is first... Value gradient update obtained guided Then by guided Calculate the corresponding noise prediction ε And the predicted quantity ε As input to the solver, execute from arrive Fast reverse denoising update.

[0164] In other words, during each step of the reverse denoising process of the preset diffusion network, the trajectory estimation result at the current moment is input into the value function to calculate the corresponding value gradient. The original output of the preset diffusion network is then corrected according to the preset guidance coefficient to obtain the value-guided denoising direction or trajectory estimation result. Then, based on the guided result, the reverse update of the current step and subsequent steps continues until all sampling steps are completed, and the output trajectory of the diffusion model is obtained.

[0165] Optionally, the diffusion direction update formula after denoising is used to obtain the driving trajectory required for the previous time step, and the loop continues until t=0, at which point the target driving trajectory is output. In this embodiment, the diffusion direction update formula is implemented as an update formula set in the DDPM solver, DDIM solver, ODE solver, or SDE solver.

[0166] Step 410, Online Application Phase.

[0167] Optionally, obtain trajectory condition information (step 4100). Obtain historical driving trajectory, perform noise addition and initialization processing on the historical driving trajectory to obtain initial noise trajectory (step 4101).

[0168] Based on the initial noise trajectory and trajectory condition information, a value-guided reverse iterative denoising process is executed (step 4102). In this process, the differentiable value network and the preset diffusion network trained in the above steps are applied. For specific application details, please refer to the above embodiments.

[0169] Optionally, after the reverse iterative denoising process is completed, the target driving trajectory is output (step 4103). Subsequently, the target driving trajectory is optimized by combining iterative quadratic adjustment (step 4104), and finally the optimized planned trajectory or control sequence is output (step 4105).

[0170] The following combination Figure 8 A detailed introduction to the value-guided reverse iterative denoising process is provided.

[0171] In the embodiments of this application, such as Figure 8 As shown, Figure 8 The illustration shows a flowchart of a directional denoising sampling process using DPM-Solver++ provided in an exemplary embodiment of this application, which specifically includes the following steps.

[0172] S1: Condition construction and initialization.

[0173] Optionally, obtain trajectory condition information C, set the number of denoising steps K (K is a positive integer), and set the noise schedule parameters (step 800).

[0174] S2: Construct the discrete denoising time grid of DPM-Solver++.

[0175] Optionally, construct a discrete denoised time series of length K. (Step 800), where Corresponding to the high noise end, Corresponding to the low-noise end, and satisfying a monotonically decreasing relationship, > > > .

[0176] In one alternative implementation, the discrete denoising time grid can be uniformly partitioned, logarithmically partitioned, or otherwise monotonically decreasing; alternatively, a noise scale sequence can be used. The two are equivalent representations.

[0177] Historical driving trajectories are processed based on noise schedule parameters to obtain a noisy trajectory. An initialization operation is then performed on the noisy trajectory to obtain an initial noisy trajectory. (Step 810). Among them, It can be obtained by sampling from a standard Gaussian distribution, that is, N(0, I).

[0178] In an optional embodiment, It can also be obtained by adding noise to the heuristic trajectory or the output trajectory of the previous planning cycle, in order to improve the closed-loop stability.

[0179] S3: Reverse iteration.

[0180] Optionally, for each denoising step k∈{K,K-1,…,1}, perform steps S3.1 to S3.4 below.

[0181] S3.1: Denoising network predicts clean trajectories (Step 820).

[0182] Optionally, the current Time stamp The trajectory condition information c is input into a preset diffusion network to obtain the predicted trajectory. .

[0183] S3.2: Value gradient guided update (step 830).

[0184] Optionally, will Inputting a differentiable value network yields an expected value. Based on this expected value, one or more guided updates are performed on the clean trajectory to obtain the guided predicted trajectory, i.e., ,in, The guiding coefficient is used to control the intensity of value guidance.

[0185] Indicatively, the initial noisy trajectory is input into a differentiable value network to obtain the expected value, and one or more guided updates are performed on the clean trajectory based on the expected value to obtain the guided predicted trajectory.

[0186] In another alternative embodiment, to avoid overly strong guidance leading to unstable solutions, the gradient can be pruned or normalized. Illustratively, gradient pruning or normalization is performed using Equation 6 or Equation 7 below.

[0187] Formula Six: ; Formula 7: ; In formulas six and seven, and This is an optional hyperparameter.

[0188] S3.3: By guided Calculate noise prediction ε .

[0189] Optionally, ε can be obtained by calculating the noise estimate based on the diffusion schedule parameters. Schematic, ε is calculated using the following formula three. .

[0190] Formula 8: ; In Formula 8, The cumulative product coefficient, This indicates the direction of noise reduction after guidance. This is the initial noisy trajectory.

[0191] Convert the above noise predictions into fractions (step 840).

[0192] S3.4: According to ε Constructing the reverse drift term and performing a higher-order integral update yields... .

[0193] In this embodiment, the reverse denoising update adopts an ordinary differential equation form based on probability flow or an equivalent deterministic update form.

[0194] For the current state and noise prediction ε k First, according to the preset noise schedule, at the time step The coefficients at a given point (e.g., noise figure) are used to construct a reverse drift term, which is used to characterize the shift from... Towards The direction and magnitude of noise reduction during evolution.

[0195] Step 850, Construction of the reverse drift term.

[0196] Optionally, ε can be predicted from the noise first. k We obtain the denoising direction quantity equivalent to the reverse evolution.

[0197] For example, it is possible to use ε based on the following equivalence relation. k The direction of noise reduction can be obtained from Formula 9 below.

[0198] Formula Nine: ; In another alternative embodiment, the noise prediction ε is maintained. k This serves as the direct input for constructing the drift term. Subsequently, based on the noise schedule and the pre-defined diffusion network parameterization method, the drift function F( For details, please refer to Formula 10 below.

[0199] Formula 10: ; In Formula 10, F( The term represents the reverse evolution drift term consistent with the noise schedule, and includes at least the result of the current state x(t) and the noise prediction. The denoising direction is determined jointly; c represents trajectory condition information.

[0200] Step 860, update higher-order numerical integrals.

[0201] Optionally, the step size of the interval [t_(k-1),t_k] is defined as: .

[0202] The drift equation is updated in one step using higher-order numerical integration to obtain... .

[0203] The aforementioned higher-order numerical integrals can be selected in first-order, second-order, or third-order form, corresponding to different trade-offs between precision and computational complexity.

[0204] Optionally, the first-order update method can perform a single display integration based on the current point drift term, as detailed in Formula 3 above.

[0205] Optionally, the second-order update method involves selecting the midpoint of the interval. Alternatively, a prediction-correction method can be used, combining two drift assessments for determination, as detailed in Formula 4 above.

[0206] Optionally, the third-order update method involves introducing multiple sampling points (such as the starting point, midpoint, and ending point) within the interval to perform multi-point combination of drift. For details, please refer to Formula 5 above.

[0207] Step 870: Output the next step status.

[0208] After completing the above higher-order integrals, the updated state is obtained. Then proceed to the next solution step k←k-1, and continue to execute steps S3.1~S3.4 until K noise reduction updates are completed to output the target driving trajectory.

[0209] S4: Output target driving trajectory (same as above) Figure 8 Step 880 is shown.

[0210] Optionally, after completing K solution steps, the low-noise output is obtained as the target driving trajectory.

[0211] Indicatively, the denoising steps are set to t=3, and the directional iterative denoising process is (x3→x2→x1→x0): The initial noisy trajectory x3 is input into a differentiable value network to obtain the first expected value, and the first expected value is used to update the initial noisy trajectory and output the first intermediate trajectory; the first intermediate trajectory is substituted into the diffusion reverse update formula to obtain the first noisy trajectory required for x2, completing the first denoising update. The first noisy trajectory output in the previous denoising process is input into a differentiable value network to obtain the second expected value, and the second expected value is used to update the first noisy trajectory and output the second intermediate trajectory; the second intermediate trajectory is substituted into the diffusion reverse update formula to obtain the second noisy trajectory required for x1, completing the second denoising update. The second noisy trajectory output in the previous denoising process is input into a differentiable value network to obtain the third expected value, and the third expected value is used to update the second noisy trajectory and output the third intermediate trajectory; the third intermediate trajectory is substituted into the diffusion reverse update formula to obtain the third noisy trajectory required for x0, completing the third denoising update. At this point, t is 0, completing the denoising process, and the third noisy trajectory output by the third denoising update is taken as the target driving trajectory.

[0212] In a preferred embodiment, the backsampling process employs DPM-Solver++, DDIM, DDPM, ODE solver, SDE solver, or equivalent implementations thereof. Value guidance can be applied to any one or more equivalent forms of the predicted trajectory, noise estimation, score function, or backshift term.

[0213] In another optional embodiment, to improve guidance stability, this application may also employ strategies such as gradient pruning, gradient normalization, gating guidance, time step weight scheduling, multi-value function combination, and candidate trajectory resampling optimization to suppress sampling divergence, trajectory jitter, or semantic offset problems caused by excessive guidance or guidance failure.

[0214] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0215] In conjunction with the above embodiments, after the target driving trajectory is output by the preset diffusion network, this application embodiment further includes an iLQR backend optimization module. This module uses the target driving trajectory as the initial optimization value or reference trajectory to perform local refinement processing on the trajectory. Under the premise of satisfying the initial state constraints and system dynamics constraints, the iLQR backend optimization jointly optimizes objectives such as trajectory reference proximity, control cost, comfort, curvature, rate of change of curvature, terminal target, safe distance, and obstacle avoidance, thereby outputting the optimized final trajectory or control sequence. Specifically, the iLQR backend optimization can further correct problems that may exist in the trajectory directly output by the diffusion model, such as insufficient smoothness, local instability, excessive curvature, abrupt steering changes, and non-executable dynamics, making the final result more suitable as a planning trajectory, control reference trajectory, or control sequence in an autonomous driving system.

[0216] In the embodiments of this application, such as Figure 9 As shown, Figure 9 A schematic diagram of an iLQR backend optimization structure provided in an exemplary embodiment of this application is shown.

[0217] After completing the above steps, the low-noise output is obtained as the target driving trajectory. ,in, and These can be regarded as symbols of the same object under different expressions, all representing the future trajectory (or control sequence) output by the preset diffusion network.

[0218] To further improve the performance of the target driving trajectory in terms of dynamic feasibility, constraint satisfaction, smoothness, comfort, and control executability, this embodiment sets up a back-end optimization module based on iLQR after the diffusion denoising output to locally refine the target driving trajectory and output the optimized target trajectory. The optimized target trajectory can be found in Formula 11 below.

[0219] Formula 11: ; In Formula 11, For trajectory condition information; The current state constraint includes at least the current state of the vehicle itself and the current state of at least one neighboring vehicle. Optimize the iLQR operator.

[0220] In one implementation, the output target driving trajectory of the preset diffusion network is... This serves as the initial reference trajectory for iLQR (step 900), thereby avoiding the dependence of traditional trajectory optimization methods on random or coarse heuristic initial values, and improving convergence speed and optimization stability.

[0221] In this embodiment, to avoid confusion with the noisy trajectory variables and denoised output variables in the aforementioned diffusion sampling process, the back-end optimization stage reconstructs and redefines the single-step state and control in the discrete time domain of the trajectory (step 910). Let the planning time domain length be N, then the optimized trajectory is expressed as... ,or, .

[0222] in, Represents the trajectory in the time domain. The system state at discrete moments ; Indicates the first Control input at discrete time points ; This indicates the trajectory output by the preset diffusion network. The corresponding number One reference state; This represents the first [missing information] corresponding to the target driving trajectory output by the preset diffusion network. A reference control; if the preset diffusion network does not directly output the control sequence, it can be estimated through differential, inverse dynamics, single-track model fitting or other methods; These represent the state and control after iLQR optimization, respectively; the initial state satisfies the vehicle state in the current state constraints (i.e.: (Step 920).

[0223] In an optional embodiment, the state vector includes at least position and heading information, and may also include velocity, acceleration, curvature-related quantities, etc.

[0224] Schematic, state vector Or, state vector The control vector can be optionally represented as ,in, For longitudinal acceleration, This refers to the front wheel steering angle or equivalent steering control quantity.

[0225] In this embodiment of the application, a dynamic model is constructed (step 930).

[0226] In this embodiment, the iLQR backend optimization uses a discrete nonlinear dynamics model to constrain trajectory evolution. The dynamics model can be any one of a monorail vehicle model, a bicycle model, a discrete dynamics model of a robotic arm, a kinematics model of a mobile robot, or other controlled object models.

[0227] Schematic example, using a monorail / bicycle model, the discrete dynamics can be expressed as: .

[0228] In one exemplary implementation, if the state adopts Control adopts Then the optional dynamic expression is: , , as well as .in, The time interval for the trajectory discrete step is... This refers to the vehicle's wheelbase or equivalent geometric parameters.

[0229] In another alternative embodiment, in the implementation characterized by "heading angle sine and cosine", the dynamics can also be rewritten as about The equivalent update form is used to improve numerical continuity and downstream control executability.

[0230] In this embodiment of the application, the iLQR cost function is constructed (step 940).

[0231] Optionally, the goal of backend optimization is to ensure that, while satisfying the dynamic and initial constraints, the optimized trajectory retains as much of the prior advantage of the pre-defined diffusion network output trajectory as possible, and further satisfies smoothness, comfort, curvature constraints, and the task objective. To this end, the following optimization problem is constructed: Among them, the total cost function It can be expressed as Formula Twelve below.

[0232] Formula 12: ; In formula twelve, For the cost of the terminal; As a stage cost; The cost of reference trajectory tracking; For reference, control costs.

[0233] In order to retain the global prior and multimodal advantages of the diffusion model-generated trajectory, one implementation penalizes the deviation between the optimized trajectory and the diffusion output trajectory, defining a reference state proximity cost. ,in, It is a positive semi-definite or positive definite weight matrix.

[0234] Accordingly, the reference control proximity cost can be defined as ,in, To control the deviation weight matrix.

[0235] In one alternative implementation, and It can vary with the time step, thus assigning different weights to the near-end time domain and the far-end time domain.

[0236] Optionally, to avoid excessive control leading to actuator saturation, tracking difficulties, or a decline in the riding experience, a control cost can be set. or ,in, These are non-negative weight parameters.

[0237] Optional, comfort cost. In autonomous driving, mobile robotics, or human-machine co-driving scenarios, comfort is typically closely related to acceleration, jerk, and the smoothness of steering changes. Therefore, in an optional embodiment, a comfort cost term is introduced, which includes the following four costs.

[0238] 1) Cost of longitudinal acceleration: .

[0239] 2) Vertical jerk cost, the discrete approximation of the vertical jerk is defined as: Then the cost is .

[0240] 3) Cost of Steering Change Rate. The discrete approximation of the steering change rate is defined as follows: The corresponding cost is .

[0241] 4) Lateral comfort cost. In one implementation, lateral comfort can be approximated by lateral acceleration. For a vehicle system, lateral acceleration can be approximated as... ,in, For the first The curvature of the trajectory at any given moment. Then the cost of lateral comfort can be defined as... .

[0242] Therefore, the total cost of comfort can be written as .

[0243] Optionally, to improve trajectory tracking, reduce the risk of sharp turns, and enhance vehicle control feasibility in narrow roads, curves, or complex scenarios, a curvature cost is introduced in a preferred embodiment. Trajectory curvature It can be approximated by changes in heading and arc length.

[0244] For example, ,in, .

[0245] Optionally, when using a single-track model, the curvature can also be approximated by the steering angle. The corresponding curvature cost can be defined as .

[0246] Furthermore, to suppress curvature abrupt changes, a curvature change rate cost can be introduced. Therefore, the total curvature-related cost can be expressed as: .

[0247] Optionally, to enhance the trajectory's ability to achieve its target at the planned endpoint, a terminal cost can be set. ,in, For the target terminal status, This is the terminal weight matrix.

[0248] In one alternative implementation, if the task is to track the endpoints of a preset diffusion network output, then it can be set as follows: This ensures that the optimization result remains consistent with the diffusion output at the endpoint.

[0249] Optional, the cost of obstacles and safe distance.

[0250] Combined with current state constraints The status and condition information of neighboring vehicles Static obstacles, lane boundaries, map features, etc., can be further considered by introducing a safety cost term. Let the first... Time from the car and the first The distance between the obstacle targets is The safety threshold is Then, the form of soft punishment can be defined as: .

[0251] In an alternative implementation, security costs can also be constructed using barrier potential fields, exponential potential energy, or barrier functions.

[0252] Optional, the first The total cost of each step can be uniformly written as .Right now: .

[0253] The final total cost function can be written as In one embodiment of this application, the iLQR method is used to iteratively solve the optimization problem with nonlinear dynamic constraints.

[0254] like Figure 10 As shown, Figure 10 This illustration shows a schematic diagram of the trajectory optimization cost term and constraint term provided in an exemplary embodiment of this application. The iLQR optimization module includes constraint term 1000 and cost term 1100.

[0255] Among them, constraint 1000 includes starting point consistency constraint 1001, dynamic constraint 1002, and local feasible region constraint 1003.

[0256] Among them, the starting point consistency constraint 1001 refers to fixing the initial state of the vehicle's optimized trajectory as follows: Local feasible region constraint 1003 includes lane boundaries, heading boundaries, and driving boundaries.

[0257] Cost item 1100 includes reference trajectory proximity cost 1101, control cost 1102, comfort cost 1103, curvature cost 1104, safety distance cost 1105, and terminal cost 1106. It should be noted that each item in the cost item can be referred to the above content, and in actual application, it can be weighted and combined according to task requirements.

[0258] Step 950, Initialization.

[0259] Output trajectory using diffusion model As the initial reference trajectory, the initialization process is as follows: .

[0260] Step 960, linearize the dynamics.

[0261] In the In the next iteration, around the current trajectory The nonlinear dynamics are linearized to the first order, as illustrated. .

[0262] in, .

[0263] Step 970, the cost function is quadratic approximation.

[0264] A second-order approximation of the total cost function is made near the current trajectory to construct a locally linear quadratic optimal control subproblem.

[0265] Step 980, reverse the recursion.

[0266] By using dynamic programming and backward recursion, the control update law at each time step is obtained, that is, implemented as follows: ,in, For feedforward term, This is the feedback gain matrix.

[0267] In one alternative implementation, to improve numerical stability, the control Hessian is regularized, that is... ,in, is the regularization coefficient.

[0268] Step 990, forward scrolling and line search.

[0269] Using line search coefficients Perform a forward simulation update, as an illustration. .

[0270] Step 991, convergence determination.

[0271] The iteration stops when any of the following conditions are met.

[0272] 1) The total cost reduction between two adjacent iterations is less than the threshold.

[0273] 2) The state or control update amount is less than the threshold; the maximum number of iterations has been reached.

[0274] Optionally, after stopping the iteration, the optimized trajectory is obtained, illustratively. .

[0275] As mentioned above, the current state constraint is denoted as: ,in, Let be the current state vector of the vehicle. For the first The current state vector of each neighboring vehicle.

[0276] During the backend optimization phase, the current state constraints are reflected in at least the following aspects.

[0277] 1) Consistency constraint at the starting point.

[0278] Optionally, the initial state of the vehicle's optimized trajectory is fixed as follows: .

[0279] 2) Basis for constructing safety constraints / costs for adjacent vehicles.

[0280] Optional, based on and condition information The historical or predicted trajectories of neighboring vehicles are used to construct the relative motion relationships between neighboring vehicles, which are used to define collision avoidance costs, safety distance costs, or optional explicit constraints.

[0281] 3) The construction of the local feasible region is based on... The map information is used to determine the current feasible lane, feasible heading range, local speed boundary, etc., thereby constraining iLQR optimization to only be carried out within the reasonable state neighborhood.

[0282] After completing the above steps, output the final trajectory obtained after backend optimization. and will Output as a planned trajectory, control reference trajectory, or control sequence.

[0283] In one implementation, the output includes an optimized sequence of states:

[0284] In another implementation, the output also includes a control sequence. .

[0285] Optional, It can also be used as a priori trajectory for the next planning cycle, part of the diffusion conditions, or a back-sampled initialization trajectory to improve the temporal consistency and closed-loop stability in the continuous planning process.

[0286] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0287] Please see Figure 11 The diagram illustrates a structural block diagram of a vehicle trajectory generation apparatus provided in another exemplary embodiment of this application. The apparatus is executed by a vehicle and includes the following components.

[0288] The acquisition module 1100 is used to acquire the historical driving trajectory and trajectory condition information of the vehicle, wherein the trajectory condition information is used to indicate the driving status of the vehicle and its neighboring vehicles respectively. Processing module 1101 is used to perform noise addition processing on the historical driving trajectory to obtain an initial noise trajectory; The acquisition module 1100 is further configured to acquire the first expected value corresponding to the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory; The denoising module 1102 is used to perform a reverse iterative denoising process based on the intermediate noise trajectory and the trajectory condition information to obtain the target driving trajectory.

[0289] In some embodiments, the acquisition module 1100 is further configured to input the intermediate noise trajectory into a preset diffusion network to obtain a noise prediction, wherein the noise prediction is used to indicate the noise information contained in the intermediate noise trajectory; The acquisition module 1100 is further configured to perform a reverse iterative denoising process based on the intermediate noise trajectory, the noise prediction, and the trajectory condition information to obtain the target driving trajectory.

[0290] In some embodiments, the acquisition module 1100 is further configured to determine the drift function corresponding to the intermediate noise trajectory, wherein the drift function is used to characterize the denoising direction evolving from the current time step to the next time step; The acquisition module 1100 is further configured to determine the target driving trajectory through the drift function during the reverse iterative denoising process.

[0291] In some embodiments, the processing module 1101 is configured to perform a display integration operation on the drift function to obtain an integration result, and determine the integration result as the target driving trajectory.

[0292] In some embodiments, the acquisition module 1100 is further configured to determine the first drift value corresponding to the drift function at the first time point within the current denoising time step, and to determine the second drift value corresponding to the drift function at the second time point within the current denoising time step. The acquisition module 1100 is further configured to perform drift prediction on the first drift value and the second drift value using a preset correction method, obtain a prediction result, and determine the prediction result as the target driving trajectory.

[0293] In some embodiments, the acquisition module 1100 is further configured to adjust the target driving trajectory according to preset dynamic constraints to obtain the adjusted target driving trajectory.

[0294] In some embodiments, the preset dynamic constraints include reference trajectory proximity conditions; The acquisition module 1100 is further configured to determine the reference driving trajectory of the vehicle based on the current state of the vehicle and the trajectory condition information. The acquisition module 1100 is further configured to determine the deviation between the reference driving estimate and the target driving trajectory; The processing module 1101 is further configured to generate a control deviation weight matrix using the deviation value, adjust the target driving trajectory according to the control deviation weight matrix, and obtain the adjusted target driving trajectory.

[0295] In some embodiments, the preset dynamic constraints include control constraints; The acquisition module 1100 is also used to acquire the longitudinal acceleration and front wheel angle of the vehicle; The acquisition module 1100 is also used to generate control adjustment parameters based on a preset first non-negative weight parameter, a preset second non-negative weight parameter, the longitudinal acceleration, and the front wheel angle; The processing module 1101 is used to adjust the target driving trajectory using the control adjustment parameters to obtain the adjusted target driving trajectory.

[0296] In some embodiments, the preset dynamic constraints include driving comfort constraints; The acquisition module 1100 is also used to generate comfort adjustment parameters based on the vehicle's longitudinal acceleration, lateral acceleration, longitudinal jerk and steering change rate; The processing module 1101 is used to adjust the target driving trajectory using the comfort adjustment parameters to obtain the adjusted target driving trajectory.

[0297] In some embodiments, the preset dynamic constraints include driving safety constraints; The acquisition module 1100 is also used to determine the distance between the vehicle and the obstacle. The acquisition module 1100 is also used to generate safety adjustment parameters based on the difference between the distance and the preset safe distance threshold. The processing module 1101 is used to adjust the target driving trajectory using the comfort adjustment parameters to obtain the adjusted target driving trajectory.

[0298] In some embodiments, the acquisition module 1100 is further configured to input the target driving trajectory into an iterative linear quadratic regulator, and optimize the target driving trajectory through the iterative linear quadratic regulator to obtain an optimized target driving trajectory.

[0299] In this embodiment, historical driving trajectories are denoised to obtain an initial noise trajectory. A first expected value corresponding to the initial noise trajectory under trajectory condition information is determined, and the initial noise trajectory is updated using the first expected value to obtain an intermediate noise trajectory. The intermediate noise trajectory is then applied to the subsequent reverse iterative denoising process to obtain the target driving trajectory. In this scheme, the first expected value is embedded in each subsequent reverse iterative denoising step, allowing the intermediate noise trajectory to evolve with the noise level. This achieves the effect of maintaining the distribution diversity of the initial noise trajectory in the early stages and enhancing the convergence of high-value areas in the later stages, ensuring that the target driving trajectory can cope with complex traffic scenarios and further improving the driving experience of the navigation path.

[0300] Figure 12 This illustration shows a structural block diagram of a computer device 1200 provided in an exemplary embodiment of this application. The computer device 1200 can be a portable mobile terminal, such as a smartphone, tablet computer, MP3 player (Moving Picture Experts Group Audio Layer III), MP4 player (Moving Picture Experts Group Audio Layer IV), laptop computer, or desktop computer. The computer device 1200 may also be referred to as a user device, portable terminal, laptop terminal, desktop terminal, or other names. Optionally, the computer device 1200 can also be implemented as a mobile device, such as a vehicle-mounted terminal or other portable smart terminal.

[0301] Typically, computer device 1200 includes a processor 1201 and a memory 1202.

[0302] Processor 1201 may include one or more processing cores, such as a 12-core processor. Processor 1201 may be implemented using at least one hardware form selected from DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), and PLA (Programmable Logic Array). Processor 1201 may also include a main processor and a coprocessor. The main processor, also known as a CPU (Central Processing Unit), is used to process data in the wake-up state; the coprocessor is a low-power processor used to process data in the standby state. In some embodiments, processor 1201 may integrate a GPU (Graphics Processing Unit), which is responsible for rendering and drawing the content to be displayed on the screen. In some embodiments, processor 1201 may also include an AI (Artificial Intelligence) processor, which is used to handle computational operations related to machine learning.

[0303] The memory 1202 may include one or more computer-readable storage media, which may be non-transitory. The memory 1202 may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments, the non-transitory computer-readable storage media in the memory 1202 are used to store at least one instruction, which is executed by the processor 1201 to implement the model training method or behavior encoding method provided in the method embodiments of this application.

[0304] This application also provides a computer-readable storage medium storing at least one instruction, at least one program, code set, or instruction set, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the vehicle trajectory generation method provided in the above method embodiments.

[0305] This application provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the vehicle trajectory generation method provided in the above-described method embodiments.

[0306] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk. The above descriptions are merely optional embodiments of this application and are not intended to limit the application. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for generating a vehicle's driving trajectory, characterized in that, The method includes: The historical driving trajectory and trajectory condition information of the vehicle are obtained, and the trajectory condition information is used to indicate the vehicle status corresponding to the vehicle and neighboring vehicles respectively. The historical driving trajectory is subjected to noise processing to obtain an initial noise trajectory; Obtain the first expected value corresponding to the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory; Based on the intermediate noise trajectory and the trajectory condition information, a reverse iterative denoising process is performed to obtain the target driving trajectory.

2. The method according to claim 1, characterized in that, The step of performing a reverse iterative denoising process based on the intermediate noise trajectory and the trajectory condition information to obtain the target driving trajectory includes: The intermediate noise trajectory is input into a preset diffusion network to obtain a noise prediction, which is used to indicate the noise information contained in the intermediate noise trajectory. Based on the intermediate noise trajectory, the noise prediction, and the trajectory condition information, a reverse iterative denoising process is performed to obtain the target driving trajectory.

3. The method according to claim 2, characterized in that, The process of performing a reverse iterative denoising process based on the intermediate noise trajectory, the noise prediction, and the trajectory condition information to obtain the target driving trajectory includes: Determine the drift function corresponding to the intermediate noise trajectory, the drift function being used to characterize the denoising direction evolving from the current time step to the next time step; In the reverse iterative denoising process, the target driving trajectory is determined by the drift function.

4. The method according to claim 3, characterized in that, Determining the target driving trajectory through the drift function includes: Perform an explicit integration operation on the drift function to obtain the integration result, and determine the integration result as the target driving trajectory.

5. The method according to claim 3, characterized in that, Determining the target driving trajectory through the drift function includes: Determine the first drift value corresponding to the drift function at the first time point within the current denoising time step, and determine the second drift value corresponding to the drift function at the second time point within the current denoising time step; A preset correction method is used to perform drift prediction on the first drift value and the second drift value to obtain the prediction result, and the prediction result is determined as the target driving trajectory.

6. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The target driving trajectory is adjusted according to the preset dynamic constraints to obtain the adjusted target driving trajectory.

7. The method according to claim 6, characterized in that, The preset dynamic constraints include reference trajectory proximity conditions; The step of adjusting the target driving trajectory according to preset dynamic constraints to obtain the adjusted target driving trajectory includes: Based on the current state of the vehicle and the trajectory condition information, determine the reference driving trajectory of the vehicle; Determine the deviation between the reference driving estimate and the target driving trajectory; The deviation value is used to generate a control deviation weight matrix, and the target driving trajectory is adjusted according to the control deviation weight matrix to obtain the adjusted target driving trajectory.

8. The method according to claim 6, characterized in that, The preset dynamic constraints include control constraints; The step of adjusting the target driving trajectory according to preset dynamic constraints to obtain the adjusted target driving trajectory includes: Obtain the longitudinal acceleration and front wheel steering angle of the vehicle; Based on the preset first non-negative weight parameter, the preset second non-negative weight parameter, the longitudinal acceleration, and the front wheel angle, control adjustment parameters are generated. The target driving trajectory is adjusted using the control and adjustment parameters to obtain the adjusted target driving trajectory.

9. The method according to claim 6, characterized in that, The preset dynamic constraints include driving comfort constraints; The step of adjusting the target driving trajectory according to preset dynamic constraints to obtain the adjusted target driving trajectory includes: Based on the vehicle's longitudinal acceleration, lateral acceleration, longitudinal jerk, and steering change rate, comfort adjustment parameters are generated. The target driving trajectory is adjusted using the aforementioned comfort adjustment parameters to obtain the adjusted target driving trajectory.

10. The method according to claim 6, characterized in that, The preset dynamic constraints include driving safety constraints; The step of adjusting the target driving trajectory according to preset dynamic constraints to obtain the adjusted target driving trajectory includes: Determine the distance between the vehicle and the obstacle. Based on the difference between the distance and the preset safe distance threshold, a safety adjustment parameter is generated; The target driving trajectory is adjusted using the aforementioned comfort adjustment parameters to obtain the adjusted target driving trajectory.

11. The method according to any one of claims 1 to 5, characterized in that, The method further includes: The target driving trajectory is input into an iterative linear quadratic regulator, which optimizes the target driving trajectory to obtain the optimized target driving trajectory.

12. A device for generating a vehicle driving trajectory, characterized in that, The device includes: The acquisition module is used to acquire the historical driving trajectory and trajectory condition information of the vehicle, wherein the trajectory condition information is used to indicate the driving status of the vehicle and its neighboring vehicles respectively; The processing module is used to perform noise addition processing on the historical driving trajectory to obtain an initial noise trajectory; The acquisition module is further configured to acquire the first expected value corresponding to the initial noise trajectory under the trajectory condition information, and update the initial noise trajectory based on the first expected value to obtain the intermediate noise trajectory; The denoising module is used to perform a reverse iterative denoising process based on the intermediate noise trajectory and the trajectory condition information to obtain the target driving trajectory.

13. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing at least one program, which is loaded and executed by the processor to implement the method for generating vehicle driving trajectories as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that, The storage medium stores at least one program segment, which is loaded and executed by a processor to implement the method for generating vehicle driving trajectories as described in any one of claims 1 to 11.

15. A computer program product or computer program, characterized in that, The computer program product or computer program includes computer instructions stored in a computer-readable storage medium, a processor of a computer device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to cause the computer device to perform the method for generating a vehicle driving trajectory as described in any one of claims 1 to 11.

Images