Trajectory planning and training method and device, vehicle, equipment, chip and medium

Abstract

The present disclosure provides a trajectory planning and training method, device, vehicle, equipment, chip and medium, and relates to the application of artificial intelligence technology in the field of vehicles. The training method comprises: generating a first noise sample associated with a first driving scene according to first scene information of the first driving scene by using a noise generation strategy in a driving model; performing trajectory planning according to the first noise sample and the first scene information by using the driving model to obtain a planned trajectory; determining a target reward parameter according to the planned trajectory; wherein the target reward parameter is used to indicate the pros and cons of the planned trajectory; and training the driving model according to the target reward parameter. Thus, the paradigm is changed from "using static noise irrelevant to the scene" to "scene-aware controllable noise generation", which significantly improves the behavior rationality, individualized adaptation ability, generalization in complex dynamic traffic environment and deployment reliability of the driving model.

Description

Technical Field

[0001] This disclosure relates to the application of artificial intelligence technology in the vehicle field, and in particular to a trajectory planning and training method, apparatus, vehicle, device, chip, and medium. Background Technology

[0002] In related technologies, driving model training involves perturbing the action space or observation input using external random noise sources. However, using external random noise sources to generate input noise limits the adaptability of vehicle planning strategies in complex and dynamic traffic environments. Summary of the Invention

[0003] This disclosure proposes a trajectory planning and training method, apparatus, vehicle, device, chip, and medium to at least partially solve one of the technical problems in the related art.

[0004] One embodiment of this disclosure proposes a method for training a driving model, comprising: generating a first noise sample associated with the first driving scenario based on first scenario information of the first driving scenario using a noise generation strategy in the driving model; performing trajectory planning using the driving model based on the first noise sample and the first scenario information to obtain a planned trajectory; determining a target reward parameter based on the planned trajectory; wherein the target reward parameter is used to indicate the quality of the planned trajectory; and training the driving model based on the target reward parameter.

[0005] In summary, by introducing a noise generation strategy coupled with scene information of the driving scenario into the driving model, and dynamically modulating the statistical characteristics of the noise samples used as the initial input of the reverse denoising process, the trajectory generation process can be effectively guided by the semantics of the driving scenario while maintaining behavioral diversity, thereby outputting a planned trajectory highly adapted to the driving scenario. The target reward parameter calculated based on this planned trajectory can more accurately reflect the performance of the planning strategy in the driving model under the driving scenario. By using this target reward parameter to optimize the driving model in reverse, a paradigm shift from "using scene-independent static noise" to "scene-aware controllable noise generation" is achieved, significantly improving the behavioral rationality, personalized adaptability, generalization and deployment reliability of the driving model in complex dynamic traffic environments.

[0006] As one possible implementation, the step of generating a first noise sample based on the first scene information of the first driving scenario using a noise generation strategy in the driving model includes: using the noise generation strategy to determine noise distribution parameters adapted to the first driving scenario based on the first scene information; and generating the first noise sample associated with the first driving scenario based on the noise distribution parameters.

[0007] Therefore, by using the first scene information as input, the noise generation strategy adaptively infers the noise distribution parameters that are compatible with the first driving scene, and generates a first noise sample accordingly. This allows the statistical characteristics of the first noise sample in the latent space to implicitly encode the semantic prior of the first driving scene. Compared with the general noise with a fixed distribution (such as a standard Gaussian distribution) used in related technologies, the first noise sample generated by this disclosure can effectively guide the reverse denoising process of the driving model to perform diversified sampling in the behavioral subspace consistent with the scene semantics. This significantly improves the scene adaptability, behavioral rationality, and style consistency of the planned trajectory while maintaining the diversity of generation, thereby enhancing the generalization ability and deployment reliability of the driving model in complex dynamic environments.

[0008] As one possible implementation, generating the first noise sample associated with the first driving scenario based on the noise distribution parameters includes: constructing a target noise distribution corresponding to the first driving scenario based on the noise distribution parameters; wherein the target noise distribution is used to characterize the statistical characteristics of noise under the first driving scenario; and randomly sampling the target noise distribution to obtain the first noise sample associated with the first driving scenario.

[0009] Therefore, by explicitly constructing the target noise distribution corresponding to the first driving scenario based on the noise distribution parameters, and using this as the basis for random sampling, the distribution of the generated first noise sample in the latent space can match the semantic features of the first driving scenario. Compared with the general noise generation method that uses a fixed prior distribution (such as a standard Gaussian distribution), this disclosure can guide the reverse denoising process of the driving model to perform efficient and diversified sampling in a subspace that is consistent with the scene semantics and feasible behavior. This not only preserves the richness of the generated results, but also significantly improves the scene adaptability and behavioral rationality of the generated planned trajectory.

[0010] As one possible implementation, determining the target reward parameter based on the planned trajectory includes: constructing a driving simulation environment based on the first scenario information, and controlling the driving of the simulated vehicle in the driving simulation environment according to the planned trajectory; determining reward parameters for at least one driving style associated with the noise generation strategy based on the state information of the simulated vehicle during driving; wherein the reward parameters are used to characterize the degree of superiority or inferiority of the planned trajectory under the corresponding driving style; and determining the target reward parameter based on the reward parameters of the at least one driving style.

[0011] Therefore, by performing closed-loop playback and control of the planned trajectory in a driving simulation environment constructed based on the first scenario information, and combining reward evaluation mechanisms under different driving styles, the target reward parameters are dynamically determined, significantly improving the scenario relevance, behavioral diversity, and policy guidance accuracy of the reward signal. Different driving styles correspond to different optimization preferences (e.g., a safety-conservative driving style focuses on safe distance, while an efficiency-first driving style emphasizes traffic efficiency). Quantitatively evaluating the performance of the same planned trajectory under different driving styles reflects its comprehensive performance in different dimensions. Furthermore, fusing the reward parameters from different driving styles to obtain the target reward parameters not only avoids policy bias caused by a single reward function but also guides the driving model to learn robust planning capabilities that consider multiple objectives and adapt to various needs. In addition, since the driving simulation environment reproduces the context of the original first driving scenario, the obtained state feedback closely resembles the real operating logic, ensuring the credibility of the target reward parameter evaluation and the effectiveness of the training signal, ultimately improving the generalization performance and task adaptability of the driving model in complex driving scenarios.

[0012] As one possible implementation, training the driving model according to the target reward parameters includes: adjusting the parameters of the noise generation strategy in the driving model according to the target reward parameters, so that the planned trajectory output by the adjusted driving model obtains a better reward parameter value under the evaluation dimension of the corresponding driving style.

[0013] Therefore, by adjusting the parameters of the noise generation strategy in reverse through the target reward parameters, it gradually learns to adaptively output noise distribution parameters that match the corresponding driving style under different driving scenarios. The noise samples generated in this way implicitly encode the prior features of high-reward behavior in the latent space, effectively guiding the reverse denoising process to search for trajectory regions with better evaluation dimensions. Compared with the fixed or static noise priors in related technologies, this disclosure transforms noise generation from a passive, general random sampling mechanism to a learnable, scene-aware generation guidance mechanism driven by task rewards. This not only significantly improves the consistency of the planned trajectory under a specified driving style, but also enhances the adaptive ability and generalization performance of the driving model in complex and dynamic traffic environments.

[0014] As one possible implementation, the driving model has multiple noise generation strategies, and the target reward parameter corresponds one-to-one with the noise generation strategy, which is used to adjust the parameters of the corresponding noise generation strategy.

[0015] Among these, the driving styles associated with the multiple noise generation strategies are different; or, the driving styles associated with the multiple noise generation strategies are not completely the same.

[0016] Therefore, by training multiple noise generation strategies in the driving model separately, the driving model can learn and internalize various differentiated behavioral patterns or driving styles, thereby flexibly adapting to the personalized preferences and scenario needs of different users in trajectory planning, significantly improving the user experience and adaptability of the driving system.

[0017] As one possible implementation, before training the driving model according to the target reward parameter, the method further includes: inputting second scene information and second noise samples of multiple second driving scenarios into the driving model for trajectory planning to obtain the predicted trajectories of the multiple second driving scenarios output by the driving model, and the latent feature representations extracted by at least some network layers in the driving model for the multiple second driving scenarios; and pre-training the driving model according to the latent feature representations and predicted trajectories of the multiple second driving scenarios.

[0018] Therefore, by using relevant data from a second driving scenario for self-supervised or imitative pre-training of the driving model before reinforcement learning, the driving model can learn general scene understanding capabilities and robust trajectory generation priors during the pre-training stage. This two-stage paradigm of "pre-training first, then reinforcement optimization" significantly reduces the difficulty of subsequent policy search based on sparse reward signals, accelerates convergence, and avoids training instability or local optima problems caused by overly random initial policies.

[0019] As one possible implementation, the pre-training of the driving model based on the latent feature representations and predicted trajectories of the multiple second driving scenarios includes: sampling the latent feature representations of the multiple second driving scenarios to obtain feature representation pairs; determining a dispersion loss based on the correlation between two latent feature representations in the feature representation pair; determining a mimicry loss based on the difference between the labeled trajectories and predicted trajectories of the multiple second driving scenarios; and pre-training the driving model based on the dispersion loss and the mimicry loss.

[0020] In summary, by jointly optimizing the dispersion loss and the imitation loss, the dual objectives of feature decoupling and behavior alignment are simultaneously achieved during the pre-training stage: the imitation loss ensures that the planned trajectory output by the driving model closely resembles real driving behavior, guaranteeing its basic performance; while the dispersion loss, by constraining the correlation between potential feature representations under different driving scenarios, encourages the driving model to learn a low-redundancy, highly discriminative feature space, avoiding high coupling or information collapse between feature dimensions. Therefore, this not only improves the driving model's ability to represent diverse driving scenarios but also enhances its generalization potential when facing new driving scenarios.

[0021] Another embodiment of this disclosure proposes a trajectory planning method, including: acquiring scene information of the driving scenario in which the vehicle is located; using a target noise generation strategy in a driving model to generate noise features associated with the driving scenario based on the scene information; and using the driving model to perform trajectory planning based on the scene information and the noise features to obtain the planned trajectory of the vehicle.

[0022] In summary, by using the scene information of the current driving scenario as input, the target noise generation strategy in the driving model generates noise features that semantically match the driving scenario. This allows the trajectory planning process to generate diverse features based on a potential prior aligned with the scene context during the inference phase. Compared to traditional methods using fixed or general noise distributions, the noise features generated in this disclosure effectively guide the driving model to sample within a feasible subspace consistent with the current scene's behavioral semantics, thereby outputting a planned trajectory that is highly coordinated in terms of driving style, driving intent, and environmental constraints. This not only improves the rationality and scene adaptability of the generated trajectory but also enhances the driving model's generalization ability and deployment reliability in complex and dynamic traffic environments.

[0023] As one possible implementation, the method for determining the target noise generation strategy includes any one of the following: in response to a style selection operation, selecting a target driving style from a variety of displayed driving styles, and using the noise generation strategy associated with the target driving style in the driving model as the target noise generation strategy; determining the actual driving style of the vehicle based on the vehicle's historical driving data, and using the noise generation strategy associated with the actual driving style in the driving model as the target noise generation strategy; in response to a strategy selection operation, selecting the target strategy identifier from a variety of displayed strategy identifiers, and using the noise generation strategy in the driving model that matches the target strategy identifier as the target noise generation strategy; wherein the strategy identifier is displayed in correspondence with the associated driving style.

[0024] In summary, by providing a flexible and configurable selection mechanism for target noise generation strategies, personalized and adaptive customization of driving behavior is achieved: when a user actively selects a target driving style, the driving system invokes the noise generation strategy associated with that style, ensuring the planned trajectory aligns with the user's subjective preferences; when the system automatically identifies the actual driving style based on historical vehicle data, it achieves seamless adaptation, making driving behavior more closely resemble the driver's habits and enhancing the human-vehicle collaborative experience; through the corresponding display of strategy identifiers and driving styles, users can switch noise generation strategies as needed, enhancing the controllability and debugging flexibility of the driving system. Therefore, the same driving model can dynamically output diverse and stylized trajectory planning results without retraining, thus catering to the differentiated needs of different users, tasks, or road conditions, significantly improving the practicality and deployment generalization capabilities of the driving system.

[0025] Another embodiment of this disclosure proposes a training apparatus for a driving model, comprising: a generation module, configured to generate a first noise sample associated with the first driving scenario based on first scenario information of the first driving scenario using a noise generation strategy in the driving model; a planning module, configured to perform trajectory planning using the driving model based on the first noise sample and the first scenario information to obtain a planned trajectory; a determination module, configured to determine a target reward parameter based on the planned trajectory; wherein the target reward parameter is used to indicate the degree of quality of the planned trajectory; and a training module, configured to train the driving model based on the target reward parameter.

[0026] Therefore, a noise generation strategy coupled with scene information of the driving scenario is introduced into the driving model. The statistical characteristics of the noise samples used as the initial input of the reverse denoising process are dynamically modulated, so that the trajectory generation process can be effectively guided by the semantics of the driving scenario while maintaining behavioral diversity, thereby outputting a planned trajectory that is highly adapted to the driving scenario. The target reward parameter calculated based on this planned trajectory can more accurately reflect the performance of the planning strategy in the driving model under the driving scenario. By using this target reward parameter to optimize the driving model in reverse, a paradigm shift from "using scene-independent static noise" to "scene-aware controllable noise generation" is achieved, which significantly improves the behavioral rationality, personalized adaptation capability, generalization and deployment reliability of the driving model in complex dynamic traffic environments.

[0027] As one possible implementation, the generation module is configured to: use the noise generation strategy to determine noise distribution parameters adapted to the first driving scenario based on the first scenario information; and generate the first noise sample associated with the first driving scenario based on the noise distribution parameters.

[0028] Therefore, by using the first scene information as input, the noise generation strategy adaptively infers the noise distribution parameters that are compatible with the first driving scene, and generates a first noise sample accordingly. This allows the statistical characteristics of the first noise sample in the latent space to implicitly encode the semantic prior of the first driving scene. Compared with the general noise with a fixed distribution (such as a standard Gaussian distribution) used in related technologies, the first noise sample generated by this disclosure can effectively guide the reverse denoising process of the driving model to perform diversified sampling in the behavioral subspace consistent with the scene semantics. This significantly improves the scene adaptability, behavioral rationality, and style consistency of the planned trajectory while maintaining the diversity of generation, thereby enhancing the generalization ability and deployment reliability of the driving model in complex dynamic environments.

[0029] As one possible implementation, the generation module is configured to: construct a target noise distribution corresponding to the first driving scenario based on the noise distribution parameters; wherein the target noise distribution is used to characterize the statistical characteristics of noise under the first driving scenario; and randomly sample the target noise distribution to obtain the first noise sample associated with the first driving scenario.

[0030] Therefore, by explicitly constructing the target noise distribution corresponding to the first driving scenario based on the noise distribution parameters, and using this as the basis for random sampling, the distribution of the generated first noise sample in the latent space can match the semantic features of the first driving scenario. Compared with the general noise generation method that uses a fixed prior distribution (such as a standard Gaussian distribution), this disclosure can guide the reverse denoising process of the driving model to perform efficient and diversified sampling in a subspace that is consistent with the scene semantics and feasible behavior. This not only preserves the richness of the generated results, but also significantly improves the scene adaptability and behavioral rationality of the generated planned trajectory.

[0031] In another aspect, this disclosure provides a trajectory planning device, comprising: an acquisition module for acquiring scene information of the driving scenario in which the vehicle is located; a generation module for generating noise features associated with the driving scenario based on the scene information using a target noise generation strategy in a driving model; and a planning module for performing trajectory planning using the driving model based on the scene information and the noise features to obtain the planned trajectory of the vehicle.

[0032] In summary, by using the scene information of the current driving scenario as input, the target noise generation strategy in the driving model generates noise features that semantically match the driving scenario. This allows the trajectory planning process to generate diverse features based on a potential prior aligned with the scene context during the inference phase. Compared to traditional methods using fixed or general noise distributions, the noise features generated in this disclosure effectively guide the driving model to sample within a feasible subspace consistent with the current scene's behavioral semantics, thereby outputting a planned trajectory that is highly coordinated in terms of driving style, driving intent, and environmental constraints. This not only improves the rationality and scene adaptability of the generated trajectory but also enhances the driving model's generalization ability and deployment reliability in complex and dynamic traffic environments.

[0033] As one possible implementation, the target noise generation strategy is determined using any of the following modules: a first processing module, configured to, in response to a style selection operation, select a target driving style from a variety of displayed driving styles, and use the noise generation strategy associated with the target driving style in the driving model as the target noise generation strategy; a second processing module, configured to, based on the vehicle's historical driving data, determine the vehicle's actual driving style, and use the noise generation strategy associated with the actual driving style in the driving model as the target noise generation strategy; and a third processing module, configured to, in response to a strategy selection operation, select the target strategy identifier from a variety of displayed strategy identifiers, and use the noise generation strategy in the driving model that matches the target strategy identifier as the target noise generation strategy; wherein the strategy identifier is displayed in correspondence with the associated driving style.

[0034] In summary, by providing a flexible and configurable selection mechanism for target noise generation strategies, personalized and adaptive customization of driving behavior is achieved: when a user actively selects a target driving style, the driving system invokes the noise generation strategy associated with that style, ensuring the planned trajectory aligns with the user's subjective preferences; when the system automatically identifies the actual driving style based on historical vehicle data, it achieves seamless adaptation, making driving behavior more closely resemble the driver's habits and enhancing the human-vehicle collaborative experience; through the corresponding display of strategy identifiers and driving styles, users can switch noise generation strategies as needed, enhancing the controllability and debugging flexibility of the driving system. Therefore, the same driving model can dynamically output diverse and stylized trajectory planning results without retraining, thus catering to the differentiated needs of different users, tasks, or road conditions, significantly improving the practicality and deployment generalization capabilities of the driving system.

[0035] This disclosure also provides an embodiment of a vehicle, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to: implement a training method for a driving model as described in the foregoing aspect, or implement a trajectory planning method as described in the foregoing aspect.

[0036] In another aspect of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements either the driving model training method described in the foregoing aspect or the trajectory planning method described in the foregoing aspect.

[0037] Another aspect of this disclosure provides a chip including an interface circuit and a processing circuit coupled to each other. The interface circuit is used to input or output signals, and the processing circuit is configured to perform a training method for a driving model as described in the preceding aspect, or to perform a trajectory planning method as described in the preceding other aspect.

[0038] In another aspect of this disclosure, a non-transitory computer-readable storage medium is provided, on which computer program instructions are stored, which, when executed by a processor, implement the training method for the driving model as described in the foregoing aspect, or, when executed, implement the trajectory planning method as described in the foregoing aspect.

[0039] Another aspect of this disclosure provides a computer program product having a computer program stored thereon, which, when executed by a processor, implements the driving model training method as described in the foregoing aspect, or, when executed, implements the trajectory planning method as described in the foregoing aspect.

[0040] Additional aspects and advantages of this disclosure will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this disclosure. Attached Figure Description

[0041] The above and / or additional aspects and advantages of this disclosure will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, in which: Figure 1 A flowchart illustrating a method for training a driving model provided as an exemplary embodiment of this disclosure; Figure 2 A flowchart illustrating another method for training a driving model provided as an exemplary embodiment of this disclosure; Figure 3 A schematic flowchart illustrating another method for training a driving model provided as an exemplary embodiment of this disclosure; Figure 4A flowchart illustrating another method for training a driving model provided as an exemplary embodiment of the present disclosure; Figure 5 A flowchart illustrating a trajectory planning method provided for an exemplary embodiment of this disclosure; Figure 6 A schematic diagram of the structure of a training device for a driving model provided as an exemplary embodiment of the present disclosure; Figure 7 A schematic diagram of the structure of a trajectory planning device provided for an exemplary embodiment of the present disclosure; Figure 8 A schematic diagram of the structure of an electronic device provided for an exemplary embodiment of the present disclosure; Figure 9 A block diagram illustrating a vehicle according to an exemplary embodiment; Figure 10 This is a schematic diagram of the structure of a chip proposed according to an exemplary embodiment of the present disclosure. Detailed Implementation

[0042] Embodiments of this disclosure are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this disclosure, and should not be construed as limiting this disclosure.

[0043] It should be noted that the acquisition, storage, use, and processing of data in this disclosed technical solution comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0044] It should also be noted that all data processed in this disclosure is data that has been explicitly authorized by users or relevant parties, and has been de-identified or anonymized before collection and use, and does not contain any personally identifiable information or user privacy content; all data is used only for trajectory planning and model training purposes, ensuring that data security and user privacy rights are fully protected while achieving technical effects.

[0045] In view of at least one of the problems existing in the above-mentioned related technologies, this disclosure proposes a trajectory planning and training method, apparatus, vehicle, device, chip and medium.

[0046] The trajectory planning and training methods, apparatus, vehicles, devices, chips, and media of embodiments of the present disclosure are described below with reference to the accompanying drawings.

[0047] Figure 1 This is a schematic flowchart illustrating a method for training a driving model, provided as an exemplary embodiment of the present disclosure.

[0048] It should be noted that the driving model training method of this disclosure can be applied to a driving model training device. In some possible embodiments, the driving model training device can be configured in an electronic device so that the electronic device can perform the driving model training function. Additionally, in some possible embodiments, the driving model training device can also be software within an electronic device.

[0049] Among them, electronic devices include, but are not limited to: terminals, servers (or cloud, servers, etc.). Terminals can be cars with communication functions, smart cars, mobile phones, wearable devices, computers, tablets (Pads), computers with wireless transceiver functions, supercomputers, etc.

[0050] like Figure 1 As shown, the training method for this driving model may include the following steps S101 to S104: Step S101: Using the noise generation strategy in the driving model, generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario.

[0051] The driving model is an artificial intelligence (AI) model to be trained, and this disclosure does not impose specific limitations on the model structure of the driving model. For example, the driving model may include a diffusion model, which includes a forward denoising process and a backward denoising process. The forward denoising process is used to progressively add noise to the target driving data or its feature representation to form diffusion state representations at different diffusion time steps. The backward denoising process is used to take noise samples as input and combine them with scene information in the driving scenario for conditional guidance to progressively denoise the noise samples, thereby generating target driving data or candidate driving behaviors that match the driving scenario. The noise samples are used as the initial input or denoised state of the backward generation process of the diffusion model, and their randomness is used to support the generation of multiple possible target driving data (such as trajectories) under the same or similar scene information.

[0052] The number of noise generation strategies in the driving model can be one or more, and this embodiment does not limit this. Each noise generation strategy is associated with at least one driving style. Different noise generation strategies are associated with different driving styles, or the driving styles associated with different noise generation strategies are not completely the same, so that the driving model can learn different driving styles.

[0053] Driving styles include, but are not limited to: safe and conservative, efficiency-first, comfortable and smooth, energy-saving and economical, and courteous and friendly driving styles.

[0054] Different driving styles are associated with different reward functions, and each reward function is specifically designed based on its associated driving style. For example, for a safe and conservative driving style, the corresponding reward function mainly penalizes behaviors such as collision risk, insufficient following distance, rapid acceleration, rapid deceleration, and sharp turns, thereby guiding the vehicle to maintain a safe distance and adopt a more stable driving style. For an efficiency-first driving style, the corresponding reward function mainly encourages the vehicle to continue along the target path and maintain a speed close to the desired speed, while penalizing unnecessary deceleration or stopping behaviors to improve overall traffic efficiency. For a comfortable and smooth driving style, the corresponding reward function mainly constrains the vehicle's acceleration, jerk, and steering change range, reducing abrupt longitudinal and lateral maneuvers, thereby improving ride comfort and smoothness. For an energy-efficient driving style, the corresponding reward function mainly penalizes energy consumption levels, frequent acceleration and deceleration behaviors, and excessive control input, prompting the vehicle to adopt a smoother and more energy-efficient driving strategy. For a courteous and friendly driving style, the corresponding reward function mainly encourages the vehicle to perform reasonable yielding in interactive scenarios, while penalizing driving behaviors that approach other road users or interfere with surrounding vehicles, to reflect a more friendly and socially friendly driving style.

[0055] The vehicle can be a hybrid vehicle, a non-hybrid vehicle, an electric vehicle, a fuel cell vehicle, or other types of vehicle. The vehicle can be a driver-assisted vehicle, a semi-driver-assisted vehicle, or a driver-free vehicle.

[0056] Among them, assisted driving refers to the technology that uses sensors, algorithms and artificial intelligence to perceive the environment, make decisions, plan and execute control commands of the vehicle in order to assist the driver to drive more safely and efficiently.

[0057] The first driving scenario can be any driving scenario in which a vehicle is located, and the scenario information of the first driving scenario (referred to as the first scenario information in this disclosure) includes at least one of the visual data, text data and point cloud data in the first driving scenario.

[0058] The visual data in the first driving scenario includes, but is not limited to, images and videos. This visual data can be collected by the vehicle's onboard cameras. For example, the onboard cameras can be cameras located on the exterior of the vehicle and can have surround-view functionality. There can be at least one onboard camera; however, to improve the comprehensiveness and accuracy of image information collection, multiple onboard cameras can be used, distributed in different locations within the vehicle as needed.

[0059] The text data in the first driving scenario refers to text information related to vehicle driving, including but not limited to at least one of the following: vehicle driving instructions, vehicle driving task information, vehicle motion status information, vehicle historical driving trajectory, and weather information (or meteorological information) of the area where the vehicle is located.

[0060] Navigation instructions are used to direct the vehicle's driving plan at the current moment. For example, navigation instructions can be commands output by a navigation system (such as map software) based on the destination location at the current moment. These instructions can be used to instruct the driving model on the macro-level operations it needs to perform, including but not limited to: left turn, right turn, left lane change, right lane change, straight ahead, and driving along the lane. In other words, navigation instructions include, but are not limited to: left turn instructions, right turn instructions, left lane change instructions, right lane change instructions, straight ahead instructions, and driving along the lane instructions.

[0061] Among them, driving task information refers to text information related to the driving task that the vehicle needs to complete, which clarifies the goal and requirements of the vehicle's current driving.

[0062] Among them, motion status information can be collected through relevant vehicle sensors, including but not limited to: driving speed, driving acceleration, steering angle, braking status, etc.

[0063] Weather information (or meteorological information) can be used to characterize the degree of impact of the vehicle's environment on perception and driving. For example, rain and snow may cause slippery roads, reduced visibility, or decreased sensor performance; high temperatures or strong light may cause overexposure of vehicle cameras; and smog will significantly weaken the detection capabilities of vehicle radar and vehicle cameras.

[0064] Understandably, driving commands provide the driving model with clear current operational instructions, allowing it to clearly understand the specific actions the vehicle should perform; driving task information clarifies the overall journey objective, enabling the driving model to plan its trajectory globally around the task; motion status information provides real-time feedback on vehicle dynamics, helping the driving model adjust its planning based on the current state; historical driving trajectories contain the vehicle's past driving patterns and habits, providing a reference for the driving model; and integrating weather information (or meteorological information) into scenario understanding helps the driving model anticipate environmental risks and adjust its trajectory planning strategy (hereinafter referred to as planning strategy), improving the robustness and safety of the driving system in adverse weather conditions. By comprehensively considering this textual data, the driving model can understand driving needs and the driving environment from multiple dimensions and levels, thereby planning more realistic, accurate, and reliable driving trajectories, effectively improving the vehicle's driving performance and safety in complex driving scenarios.

[0065] In the first driving scenario, point cloud data can be collected by the vehicle's onboard radar, which includes, but is not limited to, lidar, millimeter-wave radar, and ultrasonic radar.

[0066] The first noise sample is a noise sample associated with the first driving scenario. That is, the first noise sample is a noise sample generated by a noise generation strategy based on the first scenario information of the first driving scenario.

[0067] In this embodiment of the disclosure, a noise generation policy can be introduced in the noise generation stage. The noise generation policy is used to generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario.

[0068] Step S102: The driving model is used to perform trajectory planning based on the first noise sample and the first scene information to obtain the planned trajectory.

[0069] For example, a planning strategy in a driving model can be used to plan the trajectory based on the given first scene information and first noise sample to obtain the planned trajectory.

[0070] Step S103: Determine the target reward parameter based on the planned trajectory; wherein, the target reward parameter is used to indicate the quality of the planned trajectory.

[0071] The target reward parameter, which can be the value of a reward function, indicates the quality of the planned trajectory. A higher target reward parameter indicates a better planned trajectory, such as a safe, smooth, efficient trajectory that meets traffic rules and driving comfort requirements; a lower target reward parameter indicates a worse planned trajectory, such as a risky, uneven, inefficient, or non-compliant trajectory.

[0072] In this embodiment of the disclosure, the target reward parameter can be calculated based on the planned trajectory output by the driving model. For example, the target reward parameter can be calculated based on the planned trajectory and the reward function under the driving style associated with the noise generation strategy.

[0073] Step S104: Train the driving model according to the target reward parameters.

[0074] As one possible approach, reinforcement learning algorithms can be used to train the driving model based on the target reward parameters.

[0075] In summary, compared to related technologies that use statically distributed external random noise sources to perturb the action space or observation input, this disclosure introduces a noise generation strategy coupled with the driving scene into the driving model to generate noise samples as the initial input for the reverse denoising process. This noise generation strategy does not simply sample from a standard Gaussian distribution, but dynamically adjusts its statistical characteristics (e.g., mean, standard deviation, or latent spatial distribution) based on the scene information of the current driving scene. This guides the reverse process of the driving model towards searching a behavioral subspace semantically consistent with the current driving scene while maintaining generation diversity. Based on this, the driving model, guided by the aforementioned scene information, performs reverse denoising on the noise samples generated through scene perception, generating a planned trajectory matching the driving scene. Since the distribution characteristics of the initial noise implicitly encode prior preferences related to the driving scene, the generated planned trajectory not only satisfies safety and dynamic constraints but also exhibits diverse behaviors with consistent styles under the same driving scene.

[0076] In this disclosure, the target reward parameter calculated based on the generated planned trajectory can truly reflect the strategy performance of the driving model in a specific driving scenario. By using the target reward parameter as a supervision signal, the noise generation strategy or driving model can be trained or fine-tuned end-to-end, thereby driving the noise generation strategy to learn the ability to dynamically adjust the noise distribution parameters according to the driving scenario, so as to stimulate behavioral diversity consistent with its semantics and reward preferences in different driving scenarios.

[0077] The driving model training method of this disclosure introduces a noise generation strategy coupled with scene information of the driving scenario into the driving model. It dynamically modulates the statistical characteristics of the noise samples used as the initial input for the reverse denoising process, enabling the trajectory generation process to be effectively guided by the semantics of the driving scenario while maintaining behavioral diversity. This results in the output of a planned trajectory highly adapted to the driving scenario. The target reward parameter calculated based on this planned trajectory more accurately reflects the performance of the planning strategy in the driving model under the driving scenario. By using this target reward parameter to reverse-optimize the driving model, a paradigm shift from "using scene-independent static noise" to "scene-aware controllable noise generation" is achieved. This significantly improves the behavioral rationality, personalized adaptability, generalization, and deployment reliability of the driving model in complex dynamic traffic environments.

[0078] As one possible implementation method, Figure 2 A flowchart illustrating another method for training a driving model provided as an exemplary embodiment of this disclosure.

[0079] It should be noted that the training method of the driving model can be executed alone, or it can be executed together with any embodiment of this disclosure or any possible implementation in the embodiment, or it can be executed together with any technical solution in the related technology. The embodiments of this disclosure do not limit this.

[0080] like Figure 2 As shown, the training method for this driving model may include the following steps S201 to S205: Step S201: Using the noise generation strategy in the driving model, determine the noise distribution parameters that are adapted to the first driving scenario based on the first scenario information of the first driving scenario.

[0081] It should be noted that the explanations of the driving model, noise generation strategy, first driving scenario, and first scenario information in the foregoing embodiments also apply to this embodiment, and will not be repeated here.

[0082] The noise generation strategy is used to learn the correspondence between the input scene information and the noise distribution parameters. The noise distribution parameters include, but are not limited to, at least one of the following: noise mean, noise standard deviation, noise disturbance direction, and noise intensity.

[0083] For example, the first scene information is labeled as s, and the noise generation strategy is pi. w Assuming the noise distribution parameters include the noise mean and noise standard deviation adapted to the first driving scenario, then: (1) in, It is the noise mean output by the noise generation strategy. It is the noise standard deviation output by the noise generation strategy, used to control noise generation.

[0084] Step S202: Generate a first noise sample associated with the first driving scenario based on the noise distribution parameters.

[0085] In this embodiment of the disclosure, noise sampling can be performed based on noise distribution parameters to obtain a first noise sample associated with a first driving scenario. For example, a target noise distribution corresponding to the first driving scenario can be constructed based on the noise distribution parameters, and the target noise distribution can be randomly sampled to obtain a first noise sample associated with the first driving scenario.

[0086] The target noise distribution includes, but is not limited to, any parameterizable distribution such as Gaussian, Laplace, or mixed distribution, used to characterize the statistical properties of noise in the first driving scenario.

[0087] For example, the first noise sample is labeled as Then we have: .

[0088] Understandably, by explicitly constructing a target noise distribution corresponding to the first driving scenario based on noise distribution parameters, and using this as the basis for random sampling, the distribution of the generated first noise samples in the latent space can match the semantic features of the first driving scenario. Compared with the general noise generation method that uses a fixed prior distribution (such as a standard Gaussian distribution), this disclosure can guide the reverse denoising process of the driving model to perform efficient and diversified sampling in a subspace that is consistent with the scene semantics and behaviorally feasible. This not only preserves the richness of the generated results, but also significantly improves the scene adaptability and behavioral rationality of the generated planned trajectory.

[0089] Step S203: The driving model is used to plan the trajectory based on the first noise sample and the first scene information to obtain the planned trajectory.

[0090] Step S204: Determine the target reward parameter based on the planned trajectory; wherein, the target reward parameter is used to indicate the quality of the planned trajectory.

[0091] Step S205: Train the driving model according to the target reward parameters.

[0092] It should be noted that the explanations of steps S203 to S205 can be found in the relevant descriptions in any embodiment of this disclosure, and will not be repeated here.

[0093] The driving model training method of this disclosure uses first scene information as input, adaptively infers noise distribution parameters that are compatible with the first driving scene by a noise generation strategy, and generates a first noise sample accordingly. The statistical characteristics of the first noise sample in the latent space can implicitly encode the semantic prior of the first driving scene. Compared with the general noise with a fixed distribution (such as a standard Gaussian distribution) used in related technologies, the first noise sample generated by this disclosure can effectively guide the reverse denoising process of the driving model to perform diversified sampling in the behavioral subspace consistent with the scene semantics. This significantly improves the scene adaptability, behavioral rationality and style consistency of the planned trajectory while maintaining the diversity of generation, thereby enhancing the generalization ability and deployment reliability of the driving model in complex dynamic environments.

[0094] As one possible implementation method, Figure 3 A flowchart illustrating another method for training a driving model provided as an exemplary embodiment of this disclosure.

[0095] It should be noted that the training method of the driving model can be executed alone, or it can be executed together with any embodiment of this disclosure or any possible implementation in the embodiment, or it can be executed together with any technical solution in the related technology. The embodiments of this disclosure do not limit this.

[0096] like Figure 3 As shown, the training method for this driving model may include the following steps S301 to S306: Step S301: Using the noise generation strategy in the driving model, generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario; wherein, the noise generation strategy is associated with at least one driving style.

[0097] Step S302: The driving model is used to plan the trajectory based on the first noise sample and the first scene information to obtain the planned trajectory.

[0098] It should be noted that the explanations of steps S301 to S302 can be found in the relevant descriptions in any embodiment of this disclosure, and will not be repeated here.

[0099] Step S303: Based on the first scenario information, construct a driving simulation environment, and in the driving simulation environment, control the driving of the simulated vehicle according to the planned trajectory.

[0100] In this embodiment of the disclosure, a simulated driving environment can be constructed based on the first scenario information, and the simulated vehicle can be controlled to drive within the simulated driving environment according to the planned trajectory output by the driving model.

[0101] Step S304: Based on the state information of the simulated vehicle during driving, determine the reward parameter for at least one driving style associated with the noise generation strategy; wherein the reward parameter is used to characterize the quality of the planned trajectory under the corresponding driving style.

[0102] The real-time status information of the simulated vehicle includes, but is not limited to: the kinematic state of the simulated vehicle, such as its position coordinates, speed, acceleration, heading angle, steering angle, yaw rate, braking status, and throttle opening; the environmental interaction state of the simulated vehicle, such as the relative position information, relative speed, collision time, and minimum predicted collision distance between the simulated vehicle and surrounding obstacles; the path and traffic compliance state, such as the lateral deviation of the simulated vehicle from the lane centerline, whether it deviates from the lane, and whether it violates traffic rules (such as running a red light or crossing the line); the ride comfort state, such as the longitudinal acceleration rate of change, lateral acceleration rate of change, steering angle rate of change, and suspension displacement; and the energy and task efficiency state, such as energy consumption per unit distance and trip completion time.

[0103] The reward parameter for each driving style can be the value of the reward function associated with that driving style, which is used to characterize the quality of the planned trajectory under that driving style.

[0104] For example, reward parameters for at least one driving style associated with the noise generation strategy can be calculated based on the real-time state information of the simulated vehicle during its driving process and the reward function corresponding to at least one driving style associated with the noise generation strategy.

[0105] Step S305: Determine the target reward parameter based on the reward parameter of at least one driving style.

[0106] The target reward parameter is used to indicate the quality of the planned trajectory.

[0107] For example, if the number of driving styles associated with the noise generation strategy is only one, the reward parameter under the unique driving style can be used as the target reward parameter.

[0108] For example, when there are multiple driving styles associated with the noise generation strategy, the cumulative value, mean, or weighted sum of the reward parameters under multiple driving styles can be used as the target reward parameter.

[0109] Step S306: Train the driving model according to the target reward parameters.

[0110] As one possible approach, the parameters of the noise generation strategy in the driving model can be iteratively adjusted based on the target reward parameters until the planned trajectory generated by the driving model converges to a better reward parameter value under the evaluation dimension of the driving style associated with the noise generation strategy.

[0111] For example, a reinforcement learning algorithm can be used to adjust the parameters of the noise generation strategy in the driving model according to the target reward parameters, so that the planned trajectory output by the adjusted driving model can obtain a better reward parameter value under the evaluation dimension of the corresponding driving style.

[0112] For example, the noise generation strategy can be adjusted based on the target reward parameter Reward. Reinforcement learning is performed to learn behavioral patterns that conform to specific reward preferences, i.e., behavioral patterns under the driving style associated with the noise generation strategy.

[0113] In summary, by adjusting the parameters of the noise generation strategy in reverse using the target reward parameters, the strategy learns to adaptively output noise distribution parameters that match the corresponding driving style under different driving scenarios. The generated noise samples implicitly encode prior features of high-reward behavior in the latent space, effectively guiding the reverse denoising process towards a trajectory region with a better evaluation dimension. Compared to fixed or static noise priors in related technologies, this disclosure transforms noise generation from a passive, general random sampling mechanism to a learnable, scenario-aware generation guidance mechanism driven by task rewards. This not only significantly improves the consistency of the planned trajectory under a given driving style but also enhances the driving model's adaptability and generalization performance in complex and dynamic traffic environments.

[0114] In any embodiment of this disclosure, there may be multiple noise generation strategies in the driving model. Each noise generation strategy may correspond to a target reward parameter, that is, there is a one-to-one correspondence between the target reward parameter and the noise generation strategy. Each target reward parameter is used to adjust the parameters of its corresponding noise generation strategy.

[0115] In summary, by training multiple noise generation strategies in the driving model, the driving model can learn and internalize various differentiated behavioral patterns or driving styles, thereby flexibly adapting to the personalized preferences and scenario needs of different users in trajectory planning, significantly improving the user experience and adaptability of the driving system.

[0116] The driving model training method of this disclosure improves the scene relevance, behavioral diversity, and strategy guidance accuracy of the reward signal by performing closed-loop playback and control of the planned trajectory in a driving simulation environment constructed based on the first scene information, and dynamically determining the target reward parameters by combining reward evaluation mechanisms under different driving styles. Different driving styles correspond to different optimization preferences (e.g., a safety-conservative driving style focuses on safe distance, while an efficiency-first driving style emphasizes traffic efficiency). By quantitatively evaluating the performance of the same planned trajectory under different driving styles, its comprehensive performance in different dimensions can be reflected. Furthermore, by fusing the reward parameters under different driving styles to obtain the target reward parameters, not only is the strategy bias caused by a single reward function avoided, but the driving model can also learn robust planning capabilities that take into account multiple objectives and adapt to multiple needs. In addition, since the driving simulation environment reproduces the context of the original first driving scene, the obtained state feedback is highly close to the real operating logic, thereby ensuring the credibility of the target reward parameter evaluation and the effectiveness of the training signal, ultimately improving the generalization performance and task adaptability of the driving model in complex driving scenarios.

[0117] It should be noted that the above embodiments are the post-training stage of the driving model, and this disclosure also proposes a pre-training stage of the driving model.

[0118] Figure 4 A flowchart illustrating another method for training a driving model provided as an exemplary embodiment of this disclosure.

[0119] It should be noted that the training method of the driving model can be executed alone, or it can be executed together with any embodiment of this disclosure or any possible implementation in the embodiment, or it can be executed together with any technical solution in the related technology. The embodiments of this disclosure do not limit this.

[0120] like Figure 4 As shown, in Figures 1 to 3 Based on any embodiment, the training method for the driving model may further include the following steps S401 to S402: Step S401: Input the second scene information and second noise samples of multiple second driving scenarios into the driving model for trajectory planning, so as to obtain the predicted trajectories of multiple second driving scenarios output by the driving model, and the potential feature representations extracted by at least some network layers in the driving model for multiple second driving scenarios.

[0121] Among them, multiple second driving scenarios are different driving scenarios, and each second driving scenario has an associated labeled trajectory. For example, the labeled trajectory can be the actual driving trajectory of the vehicle in the second driving scenario.

[0122] The explanation of the second scene information can be found in the description of the first scene information in the aforementioned embodiments, as the principle is similar and will not be repeated here.

[0123] The second noise sample is different from the first noise sample. For example, the second noise sample is, for instance, standard Gaussian noise in related technologies.

[0124] The aforementioned at least some network layers may be designated network layers in the driving model, or they may be random network layers; this disclosure does not limit this. Furthermore, this disclosure does not limit the number of network layers included in at least some network layers.

[0125] In this embodiment of the disclosure, second scene information and second noise samples of multiple second driving scenarios can be input into the driving model for trajectory planning to obtain the predicted trajectory of each second driving scenario output by the driving model, as well as the potential feature representation extracted by at least some network layers in the driving model for each second driving scenario.

[0126] Step S402: Pre-train the driving model based on the potential feature representations and predicted trajectories of multiple second driving scenarios.

[0127] In this embodiment of the disclosure, a dispersive loss can be calculated based on the potential feature representations of multiple second driving scenarios, and an imitation loss can be calculated based on the difference between the predicted trajectories of multiple second driving scenarios and the corresponding labeled trajectories. Thus, this disclosure can combine the dispersive loss and the imitation loss to pre-train the driving model.

[0128] Among them, the imitation loss is positively correlated with the difference between the predicted trajectory and the labeled trajectory for each second driving scenario.

[0129] As an example, the dispersion loss can be calculated as follows: Sample the latent feature representations of multiple second driving scenarios to obtain feature representation pairs, and determine the dispersion loss based on the correlation between the two latent feature representations in the feature representation pair. For example, the covariance matrix can be determined based on the two latent feature representations in the feature pair; where the element Cov in the m-th row and n-th column of the covariance matrix... mn , used to characterize the degree of linear correlation between the m-th feature point in one latent feature representation and the n-th feature point in another latent feature representation; the dispersion loss is determined based on the sum of squares of the elements in the covariance matrix.

[0130] For example, if we label the dispersion loss as dispersive loss, then we have: (2) For example, taking the driving model as a diffusion model (Diffusion Transformer, or DiT for short), multiple second scene information from a batch can be input into DiT, and the latent feature representations output from the first c blocks (modules) of DiT can be extracted. Then, any pair of latent feature representations can be sampled within the batch to obtain a feature representation pair (or feature pair). The dispersion loss of the feature representation pair is calculated using formula (2). The specific number of c blocks can be flexibly set according to actual needs.

[0131] In summary, compared to methods relying solely on L2 regularization or simple contrastive learning, the dispersion loss based on the covariance matrix can more comprehensively characterize the global dependencies between feature dimensions. This effectively prevents the driving model from encoding almost identical latent feature representations for different driving scenarios, thereby losing its ability to perceive and express scenario differences. Consequently, the latent feature representations learned by the driving model not only possess semantic discriminative power but also have a good geometric structure, providing a more stable, interpretable, and transferable feature foundation for subsequent trajectory planning. This significantly improves the adaptability and robustness of the driving system in complex, open road environments.

[0132] As an example, a pre-training method could be as follows: calculate the target loss based on the dispersion loss and the imitation loss, and pre-train the driving model based on the target loss to minimize the target loss.

[0133] Among them, the target loss and the dispersion loss are positively correlated, and the target loss and the imitation loss are also positively correlated.

[0134] It should be noted that the above example only uses the termination condition of pre-training as the objective of minimizing loss, but this disclosure is not limited to this. For example, the termination condition may also include: the number of training rounds reaches a set number of rounds, the training duration reaches a set duration, etc.

[0135] In summary, by jointly optimizing the dispersion loss and the imitation loss, the dual objectives of feature decoupling and behavior alignment are simultaneously achieved during the pre-training stage: the imitation loss ensures that the planned trajectory output by the driving model closely resembles real driving behavior, guaranteeing its basic performance; while the dispersion loss, by constraining the correlation between potential feature representations under different driving scenarios, encourages the driving model to learn a low-redundancy, highly discriminative feature space, avoiding high coupling or information collapse between feature dimensions. Therefore, this not only improves the driving model's ability to represent diverse driving scenarios but also enhances its generalization potential when facing new driving scenarios.

[0136] It should be noted that in related technologies, driving models are mainly built using imitation learning methods. However, imitation learning methods can only learn the inherent and relatively singular driving styles and rigid, inflexible driving behaviors inherent in the driving dataset. Faced with complex and ever-changing real-world driving scenarios, this limitation makes it difficult for driving models to make reasonable decisions quickly and accurately, thus failing to fully meet the stringent safety and adaptability requirements of driving systems.

[0137] To address the aforementioned issues, the training phase of the driving model provided in this disclosure includes: 1. Pre-training stage: Taking the driving model as a diffusion model as an example, the latent feature representations of the output of the first c blocks of the diffusion model are extracted, and a pair of latent feature representations are randomly sampled within the batch. The dispersion loss is calculated. This dispersion loss is used as an auxiliary supervision signal in addition to the imitation loss to enhance the multimodal behavior capability of the driving model.

[0138] For example, during the training phase of the driving model, the set or overall space of driving scenario information is labeled as S, where S encompasses all possible driving scenarios, including all possible combinations of scenario elements such as various road conditions, traffic flow, and weather conditions; a specific instance in S is s, representing driving information for a specific driving scenario. For example, at a certain moment, the combination of information such as the vehicle's road position, the position and speed of surrounding vehicles, and the status of traffic lights constitutes a specific s; the planned trajectory output by the driving model is a; X is a theoretical random variable that follows a standard normal distribution N(0,I), used to describe the overall distribution characteristics of noise in the driving model; x is a specific sample value of the random variable X. During the training and inference process of the driving model, multiple samples can be taken according to the distribution of X, and each specific value x obtained will be used as one of the inputs of the driving model to participate in operations such as trajectory generation.

[0139] Where pi is the diffusion strategy Pi(a|s,x) of the driving model: given scene information s and input x, the probability distribution of generating the planned trajectory a.

[0140] 2. Post-training phase: A policy is introduced during the noise generation phase to output the noise and its standard deviation. Subsequently, based on the new noise with standard deviation Sampling is performed to obtain noise samples, which are then input into the downstream DiT to generate the planned trajectory. With the backbone network of the driving model frozen, the reward is calculated based on the reward function associated with the driving style of the policy, and the reward is then applied to the [specific application / process]. Reinforcement learning is used to train individuals to learn behavioral patterns that align with specific reward preferences.

[0141] Different policies are associated with different driving styles, which can solve the problem that driving models based on imitation learning in related technologies cannot generate controllable multiple driving styles and behaviors.

[0142] The driving model training method of this disclosure pre-trains the driving model using relevant data from a second driving scenario before reinforcement learning. This allows the driving model to learn general scenario understanding capabilities and robust trajectory generation priors during the pre-training stage. This two-stage paradigm of "pre-training first, then reinforcement optimization" significantly reduces the difficulty of subsequent policy search based on sparse reward signals, accelerates convergence, and avoids training instability or local optima problems caused by overly random initial policies.

[0143] The above are various embodiments corresponding to the training method of the driving model. This disclosure also proposes an application method of the driving model (i.e., trajectory planning method).

[0144] As one possible implementation method, Figure 5 This is a flowchart illustrating a trajectory planning method provided for an exemplary embodiment of the present disclosure.

[0145] It should be noted that the trajectory planning method can be executed alone, or it can be executed together with any embodiment of this disclosure or any possible implementation in the embodiment, or it can be executed together with any technical solution in related technologies. This disclosure does not limit this.

[0146] like Figure 5 As shown, the trajectory planning method may include the following steps S501 to S503: Step S501: Obtain scene information of the driving scenario in which the vehicle is located.

[0147] The scene information of the driving scenario in which the vehicle is located may include at least one of visual data, text data, and point cloud data of the driving scenario. It should be noted that the explanation of visual data, text data, and point cloud data can be found in the relevant descriptions in any of the above embodiments, and will not be repeated here.

[0148] Step S502: Using the target noise generation strategy in the driving model, noise features associated with the driving scenario are generated based on the scene information.

[0149] The driving model is a trained AI model. For example, the driving model may employ the aforementioned... Figures 1 to 4 The training method proposed in any embodiment is used for training.

[0150] The target noise generation strategy can be any noise generation strategy in the driving model, the default noise generation strategy, the noise generation strategy selected by the user, etc., and this embodiment does not limit it.

[0151] For example, multiple driving styles can be displayed to the user through a user interface (UI) or a human-machine interface (HMI), and in response to the style selection operation triggered by the user through touch, voice or shortcut commands, the target driving style can be selected from the multiple driving styles displayed, and the noise generation strategy associated with the target driving style in the driving model can be used as the target noise generation strategy.

[0152] The number of target driving styles selected by the user can be one or more, and this embodiment does not limit this.

[0153] For example, the actual driving style of a vehicle can be determined based on its historical driving data, and the noise generation strategy associated with the actual driving style in the driving model can be used as the target noise generation strategy.

[0154] Historical driving data includes, but is not limited to: driving speed distribution, acceleration distribution, following distance, lane change frequency, etc.

[0155] For example, a variety of policy identifiers can be displayed to the user through the vehicle's user interface (UI) or human-machine interface (HMI), and in response to the policy selection operation triggered by the user through touch, voice or shortcut commands, a target policy identifier can be selected from the displayed policy identifiers, and the noise generation strategy in the driving model that matches the target policy identifier can be used as the target noise generation strategy.

[0156] Among them, the strategy identifier is used to uniquely identify the noise generation strategy.

[0157] The strategy identifier is displayed in relation to its associated driving style, so that users can intuitively understand which driving style each noise generation strategy corresponds to.

[0158] In summary, by providing a flexible and configurable selection mechanism for target noise generation strategies, personalized and adaptive customization of driving behavior is achieved: when a user actively selects a target driving style, the driving system invokes the noise generation strategy associated with that style, ensuring the planned trajectory aligns with the user's subjective preferences; when the system automatically identifies the actual driving style based on historical vehicle data, it achieves seamless adaptation, making driving behavior more closely resemble the driver's habits and enhancing the human-vehicle collaborative experience; through the corresponding display of strategy identifiers and driving styles, users can switch noise generation strategies as needed, enhancing the controllability and debugging flexibility of the driving system. Therefore, the same driving model can dynamically output diverse and stylized trajectory planning results without retraining, thus catering to the differentiated needs of different users, tasks, or road conditions, significantly improving the practicality and deployment generalization capabilities of the driving system.

[0159] In this embodiment of the disclosure, a target noise generation strategy in the driving model can be used to generate noise features of the actual driving scenario of the vehicle based on scene information.

[0160] Step S503: The driving model is used to perform trajectory planning based on scene information and noise characteristics to obtain the planned trajectory of the vehicle.

[0161] In this embodiment of the disclosure, a driving model can be used to perform trajectory planning based on given scene information and noise characteristics to obtain the planned trajectory of the vehicle, thereby enabling driving control of the vehicle based on the planned trajectory.

[0162] As an application scenario, taking the application of driving models (including noise generation strategies and trajectory planning strategies) to vehicle trajectory planning as an example, personalized trajectory generation can be achieved in the following ways: (1) Explicit preference selection: The vehicle's UI or HMI presents the driver with a variety of driving styles (such as: safe and conservative, efficiency-first, comfortable and smooth, energy-saving and economical, courteous and friendly driving styles). The driver can select the target driving style he prefers from the various driving styles through touch, voice or shortcut commands. In the trajectory planning stage, the noise generation strategy associated with the target driving style selected by the driver can be called in the driving model. Based on the multi-source scene information of the driving scene where the vehicle is located (including but not limited to: multi-view images collected by the vehicle camera, real-time navigation instructions of the vehicle, driving task information, motion status information, weather information, etc.), noise features matching the driving scene are generated. The trajectory planning strategy in the driving model is called to perform trajectory planning based on scene information and noise features, and output the planned trajectory that conforms to the target driving style. In this way, the planned trajectory can not only meet the safety and traffic rule constraints, but also dynamically adapt to the personalized preferences or habits of different drivers, thus improving the driver's driving experience.

[0163] (2) Implicit style learning: The driving system automatically analyzes or models the driver's actual driving style based on the vehicle's historical driving data. Thus, in the trajectory planning stage, it can call the noise generation strategy associated with the actual driving style in the driving model, generate noise features that match the driving scene based on the multi-source scene information of the driving scene where the vehicle is located (including but not limited to: multi-view images collected by the vehicle camera, real-time navigation instructions of the vehicle, driving task information, motion status information, weather information, etc.), and call the trajectory planning strategy in the driving model to perform trajectory planning based on scene information and noise features, so as to obtain a planned trajectory that matches the driver's actual driving style, making the driving behavior closer to the driver's personal habits and improving the human-vehicle collaborative experience.

[0164] (3) Strategy Identification Switching: The vehicle's UI or HMI displays multiple strategy identifiers (such as icons, numbers, or labels) to the driver. Each strategy identifier corresponds to its associated driving style. The driver can switch strategy identifiers as needed to achieve flexible control over the planning behavior. For example, the driver can determine their preferred driving style from the displayed content and select the target strategy identifier associated with their preferred driving style from multiple strategy identifiers through touch, voice, or shortcut commands. In the trajectory planning stage, the driver can call the noise generation strategy in the driving model that matches the target strategy identifier. Based on the multi-source scene information of the driving scenario (including but not limited to: multi-view images collected by the vehicle camera, real-time navigation instructions of the vehicle, driving task information, motion status information, weather information, etc.), noise features matching the driving scenario are generated. The driver can then call the trajectory planning strategy in the driving model to perform trajectory planning based on the scene information and noise features to obtain a planned trajectory that matches their preferred driving style. This allows the driver to switch noise generation strategies as needed, enhancing the controllability and debugging flexibility of the driving system.

[0165] It should be noted that the trajectory planning method provided in this disclosure can be applied not only to the field of vehicles, but also to the field of robotics, to perform trajectory planning for robots and improve their autonomous navigation capabilities and task execution efficiency in complex environments.

[0166] The trajectory planning method of this disclosure takes the scene information of the current driving scenario as input and generates noise features that semantically match the driving scenario using a target noise generation strategy in the driving model. This allows the trajectory planning process to generate diverse features based on a potential prior aligned with the scene context during the inference phase. Compared to traditional methods that use fixed or general noise distributions, the noise features generated by this disclosure can effectively guide the driving model to sample within a feasible subspace consistent with the current scene's behavioral semantics, thereby outputting a planned trajectory that is highly coordinated in terms of driving style, driving intent, and environmental constraints. This not only improves the rationality and scene adaptability of the generated trajectory but also enhances the generalization ability and deployment reliability of the driving model in complex and dynamic traffic environments.

[0167] To implement the above embodiments, this disclosure also proposes a training device for a driving model.

[0168] Figure 6 This is a schematic diagram of the structure of a training device for a driving model provided as an exemplary embodiment of the present disclosure.

[0169] like Figure 6 As shown, the training device 600 for the driving model may include: a generation module 610, a planning module 620, a determination module 630, and a training module 640.

[0170] The generation module 610 is used to generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario using the noise generation strategy in the driving model. The planning module 620 is used to perform trajectory planning based on the first noise sample and the first scene information using the driving model to obtain the planned trajectory. The determination module 630 is used to determine the target reward parameters based on the planned trajectory; wherein, the target reward parameters are used to indicate the quality of the planned trajectory. Training module 640 is used to train the driving model based on the target reward parameters.

[0171] In one implementation of this disclosure, the generation module 610 is configured to: use a noise generation strategy to determine noise distribution parameters adapted to the first driving scenario based on the first scenario information; and generate a first noise sample associated with the first driving scenario based on the noise distribution parameters.

[0172] In one implementation of this disclosure, the generation module 610 is configured to: construct a target noise distribution corresponding to the first driving scenario based on noise distribution parameters; wherein the target noise distribution is used to characterize the statistical characteristics of noise under the first driving scenario; and randomly sample the target noise distribution to obtain a first noise sample associated with the first driving scenario.

[0173] In one implementation of this disclosure, the determining module 630 is configured to: construct a driving simulation environment based on first scenario information, and control the driving of the simulated vehicle according to a planned trajectory in the driving simulation environment; determine a reward parameter for at least one driving style associated with the noise generation strategy based on the state information of the simulated vehicle during driving; wherein the reward parameter is used to characterize the degree of superiority or inferiority of the planned trajectory under the corresponding driving style; and determine a target reward parameter based on the reward parameter of at least one driving style.

[0174] In one implementation of this disclosure, the training module 640 is used to: adjust the parameters of the noise generation strategy in the driving model according to the target reward parameters, so that the planned trajectory output by the adjusted driving model obtains a better reward parameter value under the evaluation dimension of the corresponding driving style.

[0175] In one implementation of this disclosure, the number of noise generation strategies in the driving model is multiple, and the target reward parameter corresponds one-to-one with the noise generation strategy, which is used to adjust the parameters of the corresponding noise generation strategy. Among these, different noise generation strategies are associated with different driving styles; or, The driving styles associated with different noise generation strategies are not entirely the same.

[0176] In one implementation of this disclosure, the planning module 620 is further configured to: input second scene information and second noise samples of multiple second driving scenarios into the driving model for trajectory planning, so as to obtain the predicted trajectories of multiple second driving scenarios output by the driving model, and the potential feature representations extracted by at least some network layers in the driving model for multiple second driving scenarios. The training module 640 is also used to: pre-train the driving model based on the potential feature representations and predicted trajectories of multiple second driving scenarios.

[0177] In one implementation of this disclosure, the training module 640 is configured to: sample the latent feature representations of multiple second driving scenarios to obtain feature representation pairs; determine the dispersion loss based on the correlation between the two latent feature representations in the feature representation pairs; determine the imitation loss based on the difference between the labeled trajectories and predicted trajectories of multiple second driving scenarios; and pre-train the driving model based on the dispersion loss and the imitation loss.

[0178] In one implementation of this disclosure, the training module 640 is configured to: determine a covariance matrix based on two latent feature representations in a feature pair; wherein the element in the m-th row and n-th column of the covariance matrix is ​​used to characterize the degree of linear correlation between the m-th feature point in one latent feature representation and the n-th feature point in the other latent feature representation; and determine the dispersion loss based on the sum of squares of the elements in the covariance matrix.

[0179] It should be noted that the foregoing explanation of the training method embodiment for any driving model also applies to the training device for the driving model of that embodiment, and will not be repeated here.

[0180] In the training device for the driving model of this embodiment, a noise generation strategy coupled with scene information of the driving scenario is introduced into the driving model. The statistical characteristics of the noise samples, which are the initial inputs of the reverse denoising process, are dynamically modulated. This allows the trajectory generation process to be effectively guided by the semantics of the driving scenario while maintaining behavioral diversity, thereby outputting a planned trajectory that is highly adapted to the driving scenario. The target reward parameter calculated based on this planned trajectory can more accurately reflect the performance of the planning strategy in the driving model under the driving scenario. By using this target reward parameter to optimize the driving model in reverse, a paradigm shift from "using static noise that is independent of the scenario" to "scene-aware and controllable noise generation" is achieved, which significantly improves the behavioral rationality, personalized adaptation capability, generalization and deployment reliability of the driving model in complex dynamic traffic environments.

[0181] To achieve the above embodiments, this disclosure also proposes a trajectory planning device.

[0182] Figure 7 This is a schematic diagram of the structure of a trajectory planning device provided for an exemplary embodiment of the present disclosure.

[0183] like Figure 7 As shown, the trajectory planning device 700 may include: an acquisition module 710, a generation module 720, and a planning module 730.

[0184] Among them, the acquisition module 710 is used to acquire scene information of the driving scenario in which the vehicle is located; The generation module 720 is used to generate noise features associated with the driving scene based on scene information by adopting the target noise generation strategy in the driving model. The planning module 730 is used to perform trajectory planning based on scene information and noise characteristics using a driving model to obtain the planned trajectory of the vehicle.

[0185] In one implementation of this disclosure, the target noise generation strategy is determined using any of the following modules: The first processing module is used to respond to the style selection operation, select the target driving style from the multiple driving styles displayed, and use the noise generation strategy associated with the target driving style in the driving model as the target noise generation strategy. The second processing module is used to determine the actual driving style of the vehicle based on the vehicle's historical driving data, and to use the noise generation strategy in the driving model that is associated with the actual driving style as the target noise generation strategy. The third processing module is used to respond to the strategy selection operation, select the target strategy identifier from the various displayed strategy identifiers, and take the noise generation strategy in the driving model that matches the target strategy identifier as the target noise generation strategy; wherein, the strategy identifier is displayed in correspondence with the associated driving style.

[0186] It should be noted that the foregoing explanation of the trajectory planning method embodiment also applies to the trajectory planning device of this embodiment, and will not be repeated here.

[0187] In the trajectory planning device of this disclosure, scene information of the current driving scenario of the vehicle is used as input, and a target noise generation strategy in the driving model generates noise features that semantically match the driving scenario. This enables the trajectory planning process to generate diverse features based on a potential prior aligned with the scene context during the inference phase. Compared to traditional methods that use fixed or general noise distributions, the noise features generated by this disclosure can effectively guide the driving model to sample within a feasible subspace consistent with the current scene's behavioral semantics, thereby outputting a planned trajectory that is highly coordinated in terms of driving style, driving intention, and environmental constraints. This not only improves the rationality and scene adaptability of the generated trajectory but also enhances the generalization ability and deployment reliability of the driving model in complex and dynamic traffic environments.

[0188] To implement the above embodiments, this disclosure also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the training method or trajectory planning method of the driving model as described in any of the foregoing embodiments.

[0189] Figure 8 This is a schematic diagram of the structure of an electronic device provided for an exemplary embodiment of the present disclosure. For example, the electronic device 800 may be a server, a terminal, a vehicle, etc.

[0190] Reference Figure 8 The electronic device 800 may include one or more of the following components: processing component 802, memory 804, power component 806, multimedia component 808, audio component 810, input / output (I / O) interface 812, sensor component 814, and communication component 816.

[0191] Processing component 802 typically controls the overall operation of electronic device 800, such as operations associated with display, telephone calls, data communication, camera operation, and recording operations. Processing component 802 may include one or more processors 820 to execute instructions to complete all or part of the steps of the methods described above. Furthermore, processing component 802 may include one or more modules to facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.

[0192] The memory 804 is configured to store various types of data to support the operation of the electronic device 800. The memory 804 can be implemented by any type of volatile or non-volatile storage device or a combination thereof.

[0193] Power component 806 provides power to various components of electronic device 800. Power component 806 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to electronic device 800.

[0194] The multimedia component 808 includes a screen that provides an output interface between the electronic device 800 and the user. In some embodiments, the multimedia component 808 includes a front-facing camera and / or a rear-facing camera. When the electronic device 800 is in an operating mode, such as a shooting mode or a video mode, the front-facing camera and / or the rear-facing camera can receive external multimedia data. Each front-facing camera and rear-facing camera can be a fixed optical lens system or have focal length and optical zoom capabilities.

[0195] Audio component 810 is configured to output and / or input audio signals. For example, audio component 810 includes a microphone (MIC) configured to receive external audio signals when electronic device 800 is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 804 or transmitted via communication component 816. In some embodiments, audio component 810 also includes a speaker for outputting audio signals.

[0196] I / O interface 812 provides an interface between processing component 802 and peripheral interface modules, such as keyboards, click wheels, buttons, etc.

[0197] Sensor assembly 814 includes one or more sensors for providing state assessments of various aspects of electronic device 800. For example, sensor assembly 814 can detect the on / off state of electronic device 800, the relative positioning of components, changes in the position of electronic device 800 or a component of electronic device 800, the presence or absence of user contact with electronic device 800, the orientation or acceleration / deceleration of electronic device 800, temperature changes of electronic device 800, the presence of nearby objects, etc.

[0198] Communication component 816 is configured to facilitate wired or wireless communication between electronic device 800 and other devices. Electronic device 800 can access wireless networks based on communication standards, such as WiFi, 4G, or 5G, or combinations thereof. In one exemplary embodiment, communication component 816 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 816 also includes a Near Field Communication (NFC) module to facilitate short-range communication.

[0199] In an exemplary embodiment, the electronic device 800 may be implemented by one or more chips for performing the methods described above.

[0200] In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions is also provided, such as a memory 804 including instructions, which can be executed by a processor 820 of an electronic device 800 to perform the above-described method. For example, the non-transitory computer-readable storage medium may be a read-only memory (ROM), a random access memory (RAM), a compact disc read-only memory (CD-ROM), magnetic tape, floppy disk, and optical data storage device, etc.

[0201] To implement the above embodiments, this disclosure also proposes a vehicle, including: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to implement the trajectory planning method as described in any of the foregoing embodiments.

[0202] Figure 9 This is a block diagram illustrating a vehicle 900 according to an exemplary embodiment. (Refer to...) Figure 9 The vehicle 900 may include various subsystems, such as an infotainment system 99, a perception system 920, a decision control system 930, a drive system 940, and a computing platform 950. The vehicle 900 may also include more or fewer subsystems, and each subsystem may include multiple components. Furthermore, each subsystem and component of the vehicle 900 can be interconnected via wired or wireless means.

[0203] In some embodiments, the infotainment system 99 may include a communication system, an entertainment system, and a navigation system, etc.

[0204] The perception system 920 may include several sensors for sensing information about the environment surrounding the vehicle 900. For example, the perception system 920 may include a global positioning system (which may be GPS, BeiDou, or other positioning systems), an inertial measurement unit (IMU), lidar, millimeter-wave radar, ultrasonic radar, and a camera device.

[0205] The decision control system 930 may include a computing system, a vehicle controller, a steering system, a throttle, and a braking system.

[0206] The drive system 940 may include components that provide powered motion to the vehicle 900. In one embodiment, the drive system 940 may include an engine, an energy source, a transmission system, and wheels. The engine may be one or a combination of internal combustion engines, electric motors, and compressed air engines. The engine is capable of converting energy provided by the energy source into mechanical energy.

[0207] Some or all of the functions of the vehicle 900 are controlled by a computing platform 950. The computing platform 950 may include at least one processor 951 and a memory 952, the processor 951 being able to execute instructions 953 stored in the memory 952.

[0208] Processor 951 can be any conventional processor. Processors may also include graphics processing units (GPUs), field-programmable gate arrays (FPGAs), systems on chips (SoCs), application-specific integrated circuits (ASICs), or combinations thereof.

[0209] The memory 952 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0210] In addition to instruction 953, memory 952 can also store data, such as road maps, route information, vehicle position, direction, speed, and other data. The data stored in memory 952 can be used by computing platform 950.

[0211] In this embodiment of the disclosure, processor 951 may execute instruction 953 to complete all or part of the steps of any of the above method embodiments.

[0212] To implement the above embodiments, this disclosure also proposes a chip, wherein the chip includes an interface circuit and a processing circuit coupled to each other. The interface circuit is used to input or output signals, and the processing circuit is configured to execute the training method or trajectory planning method of the driving model provided in any of the foregoing embodiments.

[0213] Figure 10 This is a schematic diagram of the structure of a chip according to an exemplary embodiment of this disclosure. See also... Figure 10 The diagram shown is a schematic representation of the structure of chip 1000, but it is not limited to this.

[0214] Chip 1000 includes processing circuit 1001, which is configured to execute any of the above trajectory planning methods or driving model training methods.

[0215] In some embodiments, the chip 1000 further includes one or more interface circuits 1002. As one possible implementation, the interface circuit 1002 is connected to the memory 1003. The interface circuit 1002 can be used to receive signals from the memory 1003 or other devices, and can also be used to send signals to the memory 1003 or other devices. For example, the interface circuit 1002 can read instructions stored in the memory 1003 and send those instructions to the processing circuit 1001.

[0216] In some embodiments, the interface circuit 1002 performs at least one of the communication steps such as sending and / or receiving in the above method, and the processing circuit 1001 performs other steps.

[0217] In some embodiments, the terms interface circuit, interface, transceiver pin, transceiver, etc., can be used interchangeably.

[0218] In some embodiments, chip 1000 further includes one or more memories 1003 for storing instructions. Optionally, all or part of the memories 1003 may be located outside of chip 1000.

[0219] To implement the above embodiments, this disclosure also proposes a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the driving model training method or trajectory planning method as described in any of the foregoing method embodiments.

[0220] To implement the above embodiments, this disclosure also proposes a computer program product having a computer program stored thereon, which, when executed by a processor, implements the training method or trajectory planning method of the driving model as described in any of the foregoing method embodiments.

[0221] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0222] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this disclosure, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0223] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing custom logic functions or processes, and the scope of preferred embodiments of this disclosure includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of this disclosure pertain.

[0224] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and compact disc read-only memory (CDROM). Alternatively, the computer-readable medium may be paper or other suitable media on which the program can be printed, since the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0225] It should be understood that various parts of this disclosure can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0226] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0227] Furthermore, the functional units in the various embodiments of this disclosure can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0228] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of the present disclosure have been shown and described above, it is to be understood that the above embodiments are exemplary and should not be construed as limiting the present disclosure. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present disclosure.

Claims

1. A method for training a driving model, characterized in that, include: The noise generation strategy in the driving model is used to generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario; The driving model is used to perform trajectory planning based on the first noise sample and the first scene information to obtain the planned trajectory. Based on the planned trajectory, a target reward parameter is determined; wherein the target reward parameter is used to indicate the quality of the planned trajectory. The driving model is trained based on the target reward parameters.

2. The method according to claim 1, characterized in that, The noise generation strategy employed in the driving model generates a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario, including: The noise generation strategy is used to determine noise distribution parameters that are adapted to the first driving scenario based on the first scenario information. Based on the noise distribution parameters, a first noise sample associated with the first driving scenario is generated.

3. The method according to claim 2, characterized in that, The step of generating the first noise sample associated with the first driving scenario based on the noise distribution parameters includes: Based on the noise distribution parameters, a target noise distribution corresponding to the first driving scenario is constructed; wherein, the target noise distribution is used to characterize the statistical characteristics of noise under the first driving scenario; Randomly sample the target noise distribution to obtain the first noise sample associated with the first driving scenario.

4. The method according to claim 1, characterized in that, The step of determining the target reward parameters based on the planned trajectory includes: Based on the first scenario information, a driving simulation environment is constructed, and in the driving simulation environment, the driving control of the simulated vehicle is performed according to the planned trajectory; Based on the state information of the simulated vehicle during driving, a reward parameter for at least one driving style associated with the noise generation strategy is determined; wherein, the reward parameter is used to characterize the quality of the planned trajectory under the corresponding driving style; The target reward parameter is determined based on the reward parameters of the at least one driving style.

5. The method according to claim 4, characterized in that, Training the driving model based on the target reward parameters includes: Based on the target reward parameters, the noise generation strategy in the driving model is adjusted so that the planned trajectory output by the adjusted driving model obtains a better reward parameter value under the evaluation dimension of the corresponding driving style.

6. The method according to claim 5, characterized in that, The driving model has multiple noise generation strategies; the target reward parameter corresponds one-to-one with the noise generation strategy and is used to adjust the parameters of the corresponding noise generation strategy. Among them, the various noise generation strategies are associated with different driving styles; or, The driving styles associated with the various noise generation strategies are not entirely the same.

7. The method according to any one of claims 1-6, characterized in that, Before training the driving model based on the target reward parameters, the method further includes: Second scene information and second noise samples from multiple second driving scenarios are input into a driving model for trajectory planning to obtain the predicted trajectories of the multiple second driving scenarios output by the driving model, as well as the latent feature representations extracted by at least some network layers in the driving model for the multiple second driving scenarios. The driving model is pre-trained based on the potential feature representations and predicted trajectories of the multiple second driving scenarios.

8. The method according to claim 7, characterized in that, The step of pre-training the driving model based on the latent feature representations and predicted trajectories of the multiple second driving scenarios includes: The potential feature representations of the multiple second driving scenarios are sampled to obtain feature representation pairs; The dispersion loss is determined based on the degree of correlation between the two potential feature representations in the feature representation pair; The imitation loss is determined based on the difference between the labeled trajectory and the predicted trajectory of the multiple second driving scenarios; The driving model is pre-trained based on the dispersion loss and the imitation loss.

9. A trajectory planning method, characterized in that, include: Obtain scene information about the driving scenario in which the vehicle is located; The target noise generation strategy in the driving model is adopted to generate noise features associated with the driving scenario based on the scenario information; The driving model is used to perform trajectory planning based on the scene information and the noise characteristics to obtain the planned trajectory of the vehicle.

10. The method according to claim 9, characterized in that, The method for determining the target noise generation strategy includes any one of the following: In response to the style selection operation, a target driving style is selected from the multiple driving styles displayed, and the noise generation strategy associated with the target driving style in the driving model is used as the target noise generation strategy. Based on the vehicle's historical driving data, the actual driving style of the vehicle is determined, and the noise generation strategy associated with the actual driving style in the driving model is used as the target noise generation strategy. In response to the strategy selection operation, the target strategy identifier is selected from a variety of displayed strategy identifiers, and the noise generation strategy in the driving model that matches the target strategy identifier is used as the target noise generation strategy; wherein the strategy identifier is displayed in correspondence with the associated driving style.

11. A training device for a driving model, characterized in that, include: The generation module is used to generate a first noise sample associated with the first driving scenario based on the first scenario information of the first driving scenario using the noise generation strategy in the driving model. The planning module is used to perform trajectory planning based on the first noise sample and the first scene information using the driving model to obtain the planned trajectory; A determination module is used to determine a target reward parameter based on the planned trajectory; wherein the target reward parameter is used to indicate the degree of quality of the planned trajectory; The training module is used to train the driving model based on the target reward parameters.

12. The apparatus according to claim 11, characterized in that, The generation module is used for: The noise generation strategy is used to determine noise distribution parameters that are adapted to the first driving scenario based on the first scenario information. Based on the noise distribution parameters, a first noise sample associated with the first driving scenario is generated.

13. The apparatus according to claim 12, characterized in that, The generation module is used for: Based on the noise distribution parameters, a target noise distribution corresponding to the first driving scenario is constructed; wherein, the target noise distribution is used to characterize the statistical characteristics of noise under the first driving scenario; Randomly sample the target noise distribution to obtain the first noise sample associated with the first driving scenario.

14. A trajectory planning device, characterized in that, include: The acquisition module is used to acquire scene information of the driving scenario in which the vehicle is located; The generation module is used to generate noise features associated with the driving scene based on the scene information by employing the target noise generation strategy in the driving model. The planning module is used to perform trajectory planning based on the scene information and noise characteristics using the driving model to obtain the planned trajectory of the vehicle.

15. The apparatus according to claim 14, characterized in that, The target noise generation strategy is determined using any of the following modules: The first processing module is used to respond to the style selection operation, select a target driving style from the multiple driving styles displayed, and use the noise generation strategy associated with the target driving style in the driving model as the target noise generation strategy. The second processing module is used to determine the actual driving style of the vehicle based on the vehicle's historical driving data, and to use the noise generation strategy associated with the actual driving style in the driving model as the target noise generation strategy. The third processing module is used to respond to the strategy selection operation, select the target strategy identifier from the multiple displayed strategy identifiers, and take the noise generation strategy in the driving model that matches the target strategy identifier as the target noise generation strategy. The strategy identifier is displayed in correspondence with the associated driving style.

16. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the steps of the method as claimed in any one of claims 1 to 8, and / or implements the steps of the method as claimed in any one of claims 9 to 10.

17. A vehicle, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured as follows: The steps of implementing the method as described in any one of claims 9 to 10.

18. A non-transitory computer-readable storage medium having computer program instructions stored thereon, characterized in that, When executed by a processor, the program instructions implement the steps of the method according to any one of claims 1 to 8, and / or implement the steps of the method according to any one of claims 9 to 10.

19. A computer program product, characterized in that, The method includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 8, and / or implements the steps of the method according to any one of claims 9 to 10.

20. A chip, characterized in that, The chip includes an interface circuit and a processing circuit coupled to each other. The interface circuit is used to input or output signals, and the processing circuit is used to implement the method of any one of claims 1 to 8, and / or to implement the steps of the method of any one of claims 9 to 10.

Images