Performance enhancement method for automatic driving system based on expert hybrid architecture

Abstract

The invention belongs to the technical field of software engineering, particularly relates to an automatic driving system performance enhancement method based on an expert hybrid architecture, and aims to solve the core problems that an end-to-end automatic driving system is confronted with semantic fuzziness to cause unreliable decision, multi-task interference hinders optimization planning, too long reasoning delay increases driving risks and the like. According to the method, an ExpertAD framework is provided, task key features are amplified through a perception adapter (PA), and the relevance of scene context understanding is guaranteed; related driving tasks are dynamically activated through a sparse expert mixture (MoSE), and task interference is minimized; and in combination with a customized training loss function, collaborative optimization of planning effectiveness and reasoning efficiency is realized. Experiments show that compared with an existing method, the method has the advantages that the average collision rate is reduced by 20%, the reasoning delay is reduced by 25%, higher multi-skill planning capacity is achieved in rare scenes (such as accident handling and first-aid vehicle avoiding), and good generalization is achieved for unseen urban environments.

Description

Performance Enhancement Methods for Autonomous Driving Systems Based on Expert Hybrid Architecture Technical Field

[0001] This invention belongs to the field of autonomous driving technology, specifically relating to a method for enhancing the performance of autonomous driving systems. Background Technology

[0002] In recent years, end-to-end autonomous driving systems have made significant progress in environmental perception and general decision-making capabilities thanks to dedicated model architectures and large-scale, diverse training data, attracting widespread attention from academia and industry. Unlike traditional modular systems, end-to-end autonomous driving systems can handle diverse driving scenarios without relying on high-definition maps or extensive manual rule-making. Most vision-based end-to-end systems use multi-view camera input to generate 2D representations for downstream perception, prediction, and planning tasks; recent research has also integrated large language models (LLMs) and world models (WMs) to improve generalization and predictive capabilities.

[0003] However, existing end-to-end autonomous driving systems still face three core challenges: First, complex driving scenarios contain rich semantic information, but ambiguous or noisy semantics can affect decision reliability. For example, sensor noise or occlusion may lead to incomplete capture of key environmental features, resulting in gaps in context understanding. Second, the relevance of tasks varies in different driving scenarios. Fully activating all tasks can cause inter-task interference and hinder optimization planning. For example, map building is beneficial for curve planning, but has limited effect on straight-line driving. Third, excessive inference latency can slow down decision-making speed and increase the risk of unsafe driving behavior. Even systems that integrate LLMs or world models still need to retain traditional end-to-end models as a backup to ensure reliable execution.

[0004] Hybrid Expert (MoE) architectures, by activating only a subset of parameters to process input, have demonstrated potential in large language and vision models while reducing computational costs, showing promise in these areas. However, their application in autonomous driving remains limited: minute input variations in dynamic scenes can lead to unstable expert activations, and existing methods often apply MoE only in the planning module, lacking task-specific expert customization and frequently resulting in significant inference latency. Therefore, designing a MoE framework spanning the entire perception-planning process to address these challenges is of great significance. Summary of the Invention

[0005] The purpose of this invention is to provide a method for enhancing the performance of an autonomous driving system based on an expert hybrid architecture, which has a low average collision rate, low inference latency, and therefore good safety performance.

[0006] The present invention provides a method for enhancing the performance of autonomous driving systems based on an expert hybrid architecture. By constructing an expert-driven autonomous driving enhancement framework, ExpertAD (hereinafter referred to as the ExpertAD framework), it achieves a synergistic improvement in the planning effectiveness and reasoning efficiency of autonomous driving systems in complex driving scenarios.

[0007] The ExpertAD framework enhances the robustness and decision-making quality of the autonomous driving system without significantly increasing computational overhead by amplifying key task features to optimize scene understanding, dynamically selecting experts relevant to the current driving task to reduce multi-task interference, and combining a balanced training loss design to ensure overall model performance. Its core idea is to reconstruct and enhance the perception and prediction modules in the end-to-end autonomous driving system without changing the original planning module's function and structure. Specifically, it introduces a feature modeling mechanism with task perception capabilities and a dynamic expert selection mechanism while maintaining the original module's functional framework, thereby improving the system's overall performance in multi-task and complex scenarios.

[0008] In this invention, the autonomous driving system is an end-to-end autonomous driving system based on visual input, which typically includes a perception module, a prediction module, and a planning module. The perception module is used to extract features and perform semantic understanding on environmental information from onboard sensors; the prediction module is used to predict the behavior of traffic participants and the vehicle's motion state; and the planning module is used to generate the vehicle's driving trajectory or control commands based on the perception and prediction results. In a preferred embodiment, the perception module includes a Bird's-Eye-View Encoder (BEV) for encoding multi-view camera images into a bird's-eye view representation, and a target tracking module and a panoramic segmentation module for performing environmental understanding tasks based on the BEV representation. These modules are common functional modules in existing end-to-end autonomous driving systems, and this invention does not limit their specific network structure or implementation.

[0009] This invention presents a performance enhancement method for autonomous driving systems based on an expert hybrid architecture. It reconstructs the perception and prediction modules of the autonomous driving system using a perception adapter (PA) and a sparse expert hybrid (MoSE) to achieve a synergistic improvement in planning effectiveness and inference efficiency. Specifically, the perception adapter (PA) and the sparse expert hybrid (MoSE) are inserted or integrated into the perception and prediction modules as functional enhancement modules to optimize feature selection, task association modeling, and inference processes without altering the basic functions of the original planning module. The method comprises the following three core steps:

[0010] (i) Utilizing a perception adapter to dynamically weight the features output by the BEV encoder, amplifying key task features, and generating a vehicle query based on fused scene understanding; specifically including:

[0011] (1) The importance of each channel of BEV features is evaluated by a learnable adapter based on task-related parameters (including but not limited to tracking weights and map weights) to generate corresponding importance scores. On this basis, the Top-K mechanism is used to dynamically filter the dimensions of task-related features, and the soft channel selection weights are obtained through a constrained optimization process, thereby forming a weighted representation for outputting task features (including but not limited to tracking features and map features).

[0012] (2) The feature dimensions are reweighted by the alignment layer, and nonlinear transformation is introduced by combining the multilayer perceptron layer to generate the aligned features after calibration, which retains the original spatial information and optimizes gradient propagation;

[0013] (3) After the alignment features are processed by the target tracking module and the panoramic segmentation module, they are stitched together with the learnable embedding to generate a vehicle query, providing a unified input for the subsequent prediction module;

[0014] (ii) Using a sparse expert hybrid to dynamically activate relevant experts based on the vehicle query, minimizing task interference, and generating motion queries; specifically including:

[0015] (1) Construct three types of 8 sparse experts, corresponding to the environmental expert group (tracking expert, map expert), the vehicle status expert group (speed expert, yaw angle expert, acceleration expert), and the navigation expert group (reference point expert, BEV expert, command expert), and each expert is configured with a dedicated sparse attention mechanism;

[0016] (2) Activate the Top-K relevant experts by dynamically selecting them based on the expert routing weights calculated by the router according to the vehicle query.

[0017] (3) Integrate the output of activation experts to generate motion queries and provide scene-adaptive prediction information for the planning module;

[0018] (iii) The combined training loss function is used to jointly optimize the model in order to balance the utilization of expert resources and task performance. The combined training loss function includes: perception module loss, prediction module loss, planning module loss and switching loss. The perception module loss, prediction module loss and planning module loss are used to perform end-to-end joint training of the model to improve the output accuracy of the corresponding task. The switching loss is used to apply a balance constraint to the allocation of expert weights, so that the weights of each expert are called tend to be balanced.

[0019] Furthermore:

[0020] In step (one):

[0021] (1) The feature selection process of the learning adapter is as follows: First, the BEV features output by the BEV encoder are normalized and pooled in the time dimension to obtain a frame-independent BEV representation that is independent of the frame sequence length. The BEV features are denoted as... Where H is the height of the BEV feature map, W is the width of the BEV feature map, and d is the number of channels. For any spatial location (i,j), its channel vector is denoted as... The colon ":" indicates that all values ​​can be taken for the channel dimension. For each perception task, a task-specific learnable parameter vector is introduced. This is used to characterize the degree of attention the task pays to each channel. Based on the parameter vector w and the frame-independent BEV representation, the initial task perception selection score s is calculated as follows:

[0022] ;

[0023] in," " indicates element-wise multiplication. Let represent the channel-level score vector, where the k-th component characterizes the relative importance of the k-th channel to the current task. Based on this, the soft channel selection weights are solved through a constrained optimization process. This allows for a smooth selection of channel importance and control over the selection scale. The optimization process can be expressed as:

[0024] ;

[0025] in, For scoring items, The regularization coefficient is . This is an entropy regularization term, used to improve the smoothness of selection weights and avoid overly sharp channel selection; A vector consisting entirely of 1s, "Indicates to Summing the elements; Choose constraint parameters for the channel (generally, (Take 128), used to control the number of effective channels that are assigned a larger weight.

[0026] The (Generally speaking, A value of 0.5 is used to adjust the smoothness of the soft channel selection weights, and its physical meaning is equivalent to the temperature parameter in the soft Top-K selection process. When When a larger value is taken, the magnitude of change in the exponential mapping in the optimization objective is smoothed, and the channel selection weights tend to be continuously distributed, thereby improving gradient stability and enhancing the differentiability of the training process; when As the weight gradually decreases and approaches zero, the optimization problem gradually degenerates into a linear programming form, and its solution approaches the discrete selection result corresponding to the traditional Top-K operation, so that only channels with high scores are assigned non-zero weights. Therefore, by adjusting... The value of Ω(α) can achieve a balance between continuously differentiable soft selection mechanisms and approximately discrete hard selection mechanisms, thereby ensuring the stability of the optimization process during model training and enhancing the ability to focus on key feature channels during inference. Ω(α) is used to constrain the distribution of soft channel selection weights to avoid excessive weight concentration or degeneration into non-smooth solutions. In a preferred embodiment, Ω(α) employs an entropy regularization term, defined as: ;

[0027] in, for The k-th component, where d is the number of channels.

[0028] By jointly introducing smooth control parameters By employing a regularization function Ω(α), this invention constructs a continuously differentiable soft Top-K channel selection mechanism, enabling end-to-end optimization of the key channel selection process in perceptual features using the standard backpropagation algorithm. Compared to traditional hard Top-K operations based on sorting and truncation, this mechanism avoids the non-differentiability problem caused by discrete selection and can adaptively adjust the sparsity of channel selection according to different driving task requirements and scenario complexity, thereby providing a more stable and task-relevant feature representation for subsequent prediction and planning modules.

[0029] (2) The feature calibration process of the alignment layer is as follows: the BEV features are multiplied element-wise with the selection weights output by the learning adapter to achieve feature amplification; after nonlinear transformation by the multilayer perceptron layer, they are connected with the residuals of the original BEV features to generate the alignment features. ;

[0030] ;

[0031] (3) The dimension of the vehicle query is B×L×d, where B is the batch size, L is the sequence length, and d is the feature dimension, which is composed of tracking query (output of target tracking module), map query (output of panoramic segmentation) and learnable embedding.

[0032] In step (two):

[0033] (1) The attention mechanism of the sparse experts is configured as follows: the environmental expert group adopts block-level sparse attention (block size m) to constrain attention to focus on local areas and reduce computational costs; the vehicle state expert group adopts sliding window attention (window size h) to quickly adapt to changes in vehicle state; and the navigation expert group adopts global Top-K attention to capture long-distance dependencies. In one implementation, each sparse expert adopts the following constrained multi-head attention computation form:

[0034] ;

[0035] in, These represent the query matrix, key matrix, and value matrix, respectively; L is the sequence length (or the number of tokens / positions). Let Q be the i-th row vector. and Let K and V represent the j-th row vectors, respectively. For the key / query vector dimension, This is the scaling factor; Indicates in the candidate index set Normalization is performed on the above.

[0036] The candidate index set Based on the sparsity mode settings of different expert groups:

[0037] Block-level sparse attention (Environmental expert group):

[0038] ;

[0039] Where m is the block size (generally, Take 8), This represents the floor function, used to divide the index into the corresponding local block.

[0040] Sliding window attention (Automotive status expert group):

[0041] ;

[0042] Where h is the window radius (generally, Taking 16 means that h neighboring positions to the left and right of the current index i are selected to participate in the attention calculation.

[0043] Global Attention (Navigation Experts Group):

[0044] ;

[0045] in, Indicates from the entire index The set consisting of the k highest-scoring indices is selected from the set, where k is the preset number of selections (generally, Take 8).

[0046] (2) The expert selection process for the router is as follows: introducing a learnable parameter matrix. Mapping vehicle queries to expert logarithmic probabilities Gaussian noise is added during the training phase. Controlling noise levels; through functions Select the Top-K experts, where K is usually set to 4 by default.

[0047] The routing score function Defined as:

[0048] ;

[0049] Where x represents the input vehicle query vector; This is the learnable weight matrix used to calculate the expert route score; This is the noise modulation weight matrix; Random noise that follows a standard normal distribution is used to introduce random perturbations during the training phase; For smooth activation functions; This is a smoothing bias term used to prevent the routing score from degrading to zero.

[0050] The routing score function Defined as:

[0051] ;

[0052] in, Used to normalize the routing score into a probability distribution This represents the set of indices corresponding to the K experts with the highest probability values.

[0053] (3) The motion query The generation process is as follows: Based on the routing score output by the router, the exclusive query of the activation expert is weighted and fused, integrating vehicle features, environmental features, and navigation features to form a global feature representation adapted to the current driving scenario.

[0054] ;

[0055] in, Indicates the characteristics of a vehicle query; This represents the routing weight corresponding to the i-th selected expert; This represents the feature representation output by the i-th expert.

[0056] In step (three):

[0057] The switching loss To balance the query load distribution among experts, the actual query load of each expert is matched with the expected routing probability by penalizing experts who receive excessively high routing probabilities; the combined training loss function Used for end-to-end joint training of models to improve the output accuracy of corresponding tasks.

[0058] The switching loss is defined as:

[0059] ;

[0060] Where N is the total number of experts; This represents the relative frequency at which the i-th expert is selected in a training batch; This represents the average routing probability corresponding to the expert, used to measure the load occupancy of that expert.

[0061] The combined training loss function is defined as:

[0062] ;

[0063] in, The loss of the perception module is used to constrain the output accuracy of perception tasks such as object detection, object tracking, or semantic segmentation. The loss of the prediction module is used to constrain the accuracy of traffic participant trajectory prediction or vehicle motion prediction; The planning module loss is used to constrain the safety, feasibility, and comfort of the generated trajectory; This is for switching losses. , , , These are weighting coefficients, generally speaking. =1, =1.0, =2.0, =1.0.

[0064] Figure 1 illustrates the basic process of this invention. Addressing the challenges of insufficient semantic understanding, significant multi-task interference, and high inference latency in complex scenarios, end-to-end autonomous driving systems propose a performance enhancement method based on an expert hybrid architecture. This method constructs the ExpertAD autonomous driving enhancement framework, which, without altering the original planning module's functions and interfaces, specifically reconstructs and enhances the perception and prediction modules: The perception adapter dynamically weights and performs soft Top-K channel selection on the bird's-eye view features output by the BEV encoder, amplifying key semantics and suppressing redundant information, thereby improving the stability and interpretability of scene context understanding; a sparse expert hybrid selects routes among three expert groups (environment, vehicle state, and navigation) and activates only the Top-K relevant experts, reducing irrelevant task coupling and computational overhead, forming motion queries tailored to the current driving scenario to support more reliable prediction and planning; and end-to-end joint optimization of the model is achieved by introducing a combined training loss function, including switching loss, to synergistically improve expert load balancing and task performance. Compared to existing end-to-end autonomous driving methods, this invention maintains the integrability of systems engineering while more accurately focusing on key features and key tasks in multi-task, long-tail, and rare risk scenarios, reducing task interference and inference delay, thereby improving the robustness and safety of planning and decision-making, and exhibiting better generalization ability and stability in different cities and unseen environments. Attached Figure Description

[0065] Figure 1 shows the overall architecture of the ExpertAD framework.

[0066] Figure 2 shows the collision rate versus delay trade-off for different models.

[0067] Figure 3 is a schematic diagram comparing the detection and trajectory adjustment effects of UniAD (top) and ExpertAD (bottom). Detailed Implementation

[0068] The following describes the specific implementation of this invention using the nuScenes dataset, the Bench2Drive benchmark set, and three mainstream end-to-end autonomous driving systems: UniAD, VAD, and VADv2. The main process is as follows:

[0069] (I) Framework Integration and Parameter Configuration: The ExpertAD framework is integrated into UniAD, VAD, and VADv2, resulting in Expert-UniAD, Expert-VAD, and Expert-VADv2, respectively. The original planning modules and hyperparameters of each baseline model are retained; only the perception and prediction modules are replaced with a hybrid of a perception adapter and a sparse expert module.

[0070] 1. Perception Adapter: Set the candidate range for the number of dominant feature channels T to [32, 64, 128, 256], and determine the optimal value through grid search (default T=128);

[0071] 2. Sparse expert hybrid: The number of experts is fixed at 8, the routing selection has Top-K=4 experts, the window size of sliding window attention is w=16, and the block size of block-level sparse attention is m=8;

[0072] 3. Training loss: The switching loss weight coefficient λ=0.1, using the AdamW optimizer, and the training epochs are 20.

[0073] (II) Experimental Environment and Dataset Setup

[0074] 1. Hardware environment: 8 NVIDIA Tesla A100 GPUs, each with 40GB of video memory;

[0075] 2. Dataset:

[0076] (1) Open-loop evaluation: The nuScenes dataset, containing 1000 real scenes, was used. Only 6 camera images were used as visual input. The evaluation metrics were L2 error and collision rate.

[0077] (2) Closed-loop evaluation: The Bench2Drive benchmark test set is used, which contains 2 million labeled training frames (covering 12 towns and 23 weather conditions). The evaluation set contains 220 routes and 44 interaction scenarios. The evaluation metrics are driving score (DS), route success rate (SR), and route completion rate (RC).

[0078] 3. Efficiency metrics: inference latency (average time for a single forward propagation), computational complexity (GFLOPs), and model parameter size (Params).

[0079] (III) Experimental Results and Analysis

[0080] 1. Overall performance improvement:

[0081] (1) Open-loop evaluation: Compared with UniAD, Expert-UniAD reduces the collision rate by 23%, reduces L2 error by 14%, and increases inference speed by 1.2 times; Expert-VAD and Expert-VADv2 reduce the collision rate by 21% and 17% respectively, reduce L2 error by 9% and 15% respectively, and increase inference speed by 1.4 times and 1.3 times respectively;

[0082] (2) Closed-loop evaluation: The average driving score (DS) of the three integrated models increased by 16%, the average route success rate (SR) increased by 22%, and the average route completion rate (RC) increased by 14%.

[0083] (3) Efficiency improvement: The model parameter size increases only slightly, the computational complexity (GFLOPs) is significantly reduced, and the average inference latency is reduced by 25%.

[0084] 2. Module validity verification:

[0085] (1) Perception adapter module: When the feature channel T=128, the introduction of the perception module improves the driving score by 17%, the route success rate by 30%, and the route completion rate by 11%. Too large a channel selection T (such as 256) will introduce feature redundancy, resulting in a decrease in performance.

[0086] (2) Sparse expert hybrid module: The performance is optimal when Top-4 experts are activated. Compared with full activation (Top-8), the route completion rate is increased by 5% and the inference latency is reduced by 18%, which verifies the role of dynamic expert selection in reducing interference and computational cost.

[0087] 3. Scene adaptability and generalization:

[0088] (1) Performance in rare scenarios: In emergency braking (such as pedestrian crossing, parking and cutting in), lane changing (such as highway cutting in, intersection turning), and traffic sign recognition (such as stop signs, traffic lights), the performance of the ExpertAD framework is significantly improved; in overtaking (such as accident scene, construction area) and yielding (such as yielding to emergency vehicles), a rule-based backup system is still required.

[0089] Cross-city generalization: Cross-city testing on the nuScenes dataset (trained in Boston and tested in Singapore, or vice versa) showed that ExpertAD reduced the collision rate by an average of 15% and the L2 error by an average of 10%, validating the framework's ability to adapt to unseen urban environments.

[0090] (2) Case analysis: In a complex scenario involving traffic administrators, UniAD failed to detect the traffic administrator on the right front and the planned trajectory deviated towards the administrator; while Expert-UniAD amplified the characteristics of traffic participants through the perception adapter, activated the environmental expert and navigation expert through the sparse expert hybrid, adjusted the route in real time, and safely completed the lane change, verifying the framework's scenario understanding and dynamic planning capabilities.

[0091] Figure 3 is a schematic diagram comparing the detection and trajectory adjustment effects of UniAD (top) and ExpertAD (bottom).

[0092] Experimental results show that the ExpertAD framework proposed in this invention can effectively improve the planning effectiveness and inference efficiency of end-to-end autonomous driving systems. While reducing collision risks and inference delays, it enhances scenario adaptability and generalization, providing technical support for the safe deployment of high-level autonomous driving systems.

Claims

1. A method for enhancing the performance of an autonomous driving system based on an expert hybrid architecture, wherein the autonomous driving system is an end-to-end autonomous driving system based on visual input, comprising a perception module, a prediction module, and a planning module; wherein, The perception module is used to extract features and understand semantics from environmental information from vehicle sensors, the prediction module is used to predict the behavior of traffic participants and the motion state of the vehicle, and the planning module is used to generate the vehicle's driving trajectory or control commands based on the perception and prediction results; the perception module includes a BEV encoder for encoding multi-view camera images into a bird's-eye view representation, and a target tracking module and a panoramic segmentation module for performing environmental understanding tasks based on the BEV representation; the feature is that by constructing an expert hybrid-driven autonomous driving framework (ExpertAD) based on the functional enhancement module, the synergistic improvement of planning effectiveness and reasoning efficiency is achieved; the functional enhancement module includes a perception adapter (PA) and a sparse expert hybrid (MoSE), which are inserted or integrated into the perception module and the prediction module to optimize feature selection, task association modeling and reasoning process without changing the basic functions of the original planning module; specifically, it includes the following three steps: (i) using the perception adapter to dynamically weight the features output by the BEV encoder, amplify the key features of the task, and generate a vehicle query that integrates scene understanding; specifically, it includes: (1) using a learnable adapter to adjust the BEV according to the task-related parameters. The importance of each feature channel is evaluated and a corresponding importance score is generated. On this basis, the Top-K mechanism is used to dynamically filter the feature dimensions related to the task, and the soft channel selection weight is obtained through the constrained optimization process, thereby forming a weighted representation for outputting task features. (2) The feature dimensions are reweighted through the alignment layer, and nonlinear transformation is introduced by combining the multilayer perceptron layer to generate the calibrated alignment features, retain the original spatial information and optimize gradient propagation. (3) After the alignment features are processed by the target tracking module and the panoramic segmentation module, they are spliced ​​with the learnable embedding to generate the vehicle query, providing a unified input for the subsequent prediction module. (II) The sparse expert hybrid is used to dynamically activate relevant experts according to the vehicle query, minimize task interference, and generate motion query. Specifically, it includes: (1) Constructing three types of 8 sparse experts, namely, the environmental expert group, including tracking experts and map experts; the vehicle status expert group, including speed experts, yaw angle experts, and acceleration experts; and the navigation expert group, including reference point experts and BEV experts. Experts and instruction experts; each expert is configured with a dedicated sparse attention mechanism; (2) the router dynamically selects the Top-K relevant experts to activate based on the expert routing weights calculated by the vehicle query; (3) the outputs of the activated experts are fused to generate motion queries, providing scene-adaptive prediction information for the planning module; (III) the combined training loss function is used to jointly optimize the model in order to balance the utilization rate of expert resources and task performance; the combined training loss function includes: perception module loss, prediction module loss, planning module loss and switching loss; among them, perception module loss, prediction module loss and planning module loss are used to jointly train the model end-to-end in order to improve the output accuracy of the corresponding task;Switching loss is used to impose an equilibrium constraint on the allocation of expert weights, prompting the weights of each expert to tend towards equilibrium.

2. The method for enhancing the performance of an autonomous driving system according to claim 1, characterized in that, In step (1): (1) The feature selection process of the learning adapter is as follows: First, the BEV features output by the BEV encoder are normalized and pooled in the time dimension to obtain a frame-independent BEV representation that is independent of the frame sequence length; the BEV features are denoted as Where H is the height of the BEV feature map, W is the width of the BEV feature map, and d is the number of channels; for any spatial location (i,j), its channel vector is denoted as... The colon ":" indicates that all values ​​can be taken for the channel dimension. For each perception task, a task-specific learnable parameter vector is introduced. The parameter vector w is used to characterize the degree of attention the task pays to each channel; the initial task perception selection score s is calculated based on the frame-independent BEV representation of the parameter vector w, specifically: ;in," " indicates element-wise multiplication. Let represent the channel-level score vector, where the k-th component characterizes the relative importance of the k-th channel to the current task. Based on this, the soft channel selection weights are solved through a constrained optimization process. This allows for a smooth selection of channel importance and control over the selection scale; the optimization process is expressed as: ;in, For scoring items, The regularization coefficient is . This is an entropy regularization term, used to improve the smoothness of selection weights and avoid overly sharp channel selection; A vector consisting entirely of 1s, "Indicates to Summing the elements; Select constraint parameters for the channels to control the number of effective channels assigned larger weights; the Used to adjust the smoothness of soft channel selection weights, when When a larger value is taken, the magnitude of change in the exponential mapping in the optimization objective is smoothed, and the channel selection weights tend to be continuous, thereby improving gradient stability and enhancing the differentiability of the training process; when As the weight gradually decreases and approaches zero, the optimization problem gradually degenerates into a linear programming form, and its solution approaches the discrete selection result corresponding to the traditional Top-K operation, so that only channels with high scores are assigned non-zero weights; therefore, by adjusting... The value of Ω(α) can achieve a balance between continuously differentiable soft selection mechanisms and approximately discrete hard selection mechanisms, thereby ensuring the stability of the optimization process during model training and enhancing the ability to focus on key feature channels during inference. Ω(α) is used to constrain the distribution of soft channel selection weights to avoid excessive weight concentration or degeneration into non-smooth solutions. Ω(α) employs an entropy regularization term, defined as: ;in, for The k-th component, where d is the number of channels; smoothing control parameters are introduced jointly. The regularization function Ω(α) constructs a continuously differentiable soft Top-K channel selection mechanism, enabling the selection process of key channels in the perception features to be optimized end-to-end through the standard backpropagation algorithm; (2) The feature calibration process of the alignment layer is as follows: the BEV features are multiplied element-wise with the selection weights output by the learning adapter to achieve feature amplification; after nonlinear transformation by the multilayer perceptron layer, they are connected with the residual of the original BEV features to generate the alignment features. ； (3) The dimension of the vehicle query is B×L×d, where B is the batch size, L is the sequence length, and d is the feature dimension, which is composed of tracking query, map query and learnable embedding.

3. The method for enhancing the performance of an autonomous driving system according to claim 2, characterized in that, In step (ii): (1) the attention mechanism of the sparse experts is configured as follows: the environmental expert group adopts block-level sparse attention constraint to focus attention on local areas to reduce computational costs, and the block size is set to m; the vehicle state expert group adopts sliding window attention to quickly adapt to changes in the vehicle state, and the window size is set to h; the navigation expert group adopts global Top-K attention to capture long-distance dependencies; each sparse expert adopts the following constrained multi-head attention calculation form: ;in, These represent the query matrix, key matrix, and value matrix, respectively; L is the sequence length or the number of tokens / positions. Let Q be the i-th row vector. and Let K and V represent the j-th row vectors, respectively. For the key / query vector dimension, This is the scaling factor; Indicates in the candidate index set Normalization is performed on the above; the candidate index set Based on the sparsity patterns of different expert groups, the following settings are used: For the environmental expert group, the block-level sparse attention is: Where m is the block size, This represents the floor function, used to divide indices into corresponding local blocks; the vehicle status expert group's sliding window attention is: Where h is the window radius, indicating that h neighboring positions to the left and right of the current index i are selected for attention calculation; Navigation expert group, global. Attention is: ;in, Indicates from the entire index The set consists of the k indexes with the highest scores, where k is the preset number of selections.

4. The method for enhancing the performance of an autonomous driving system according to claim 3, characterized in that, In step (ii): the process of dynamically selecting the Top-K relevant experts for activation by calculating the expert routing weights based on vehicle queries through the router is as follows: introducing a learnable parameter matrix. Mapping vehicle queries to expert logarithmic probabilities ； Add Gaussian noise during training Controlling noise levels; through functions Select Top-K experts; the routing score function Defined as: Where x represents the input vehicle query vector; This is the learnable weight matrix used to calculate the expert route score; This is the noise modulation weight matrix; Random noise that follows a standard normal distribution is used to introduce random perturbations during the training phase; For smooth activation functions; A smoothing bias term is used to prevent the routing score from degrading to zero; the routing score function Defined as: ;in, Used to normalize the routing score into a probability distribution This represents the set of indices corresponding to the K experts with the highest probability values.

5. The method for enhancing the performance of an autonomous driving system according to claim 4, characterized in that, In step (two): the motion query The generation process is as follows: Based on the routing score output by the router, the exclusive query of the activation expert is weighted and fused, integrating vehicle features, environmental features, and navigation features to form a global feature representation adapted to the current driving scenario. ;in, Indicates the characteristics of a vehicle query; This represents the routing weight corresponding to the i-th selected expert; This represents the feature representation output by the i-th expert.

6. The method for enhancing the performance of an autonomous driving system according to claim 5, characterized in that, In step (iii): the switching loss To balance the query load distribution among experts, the actual query load of each expert is matched with the expected routing probability by penalizing experts who receive excessively high routing probabilities; the combined training loss function Used for end-to-end joint training of the model to improve the output accuracy of the corresponding task; the switching loss is defined as: Where N is the total number of experts; This represents the relative frequency at which the i-th expert is selected in a training batch; This represents the average routing probability corresponding to the expert, used to measure the expert's load occupancy; the combined training loss function is defined as: ;in, The loss of the perception module is used to constrain the output accuracy of perception tasks such as object detection, object tracking, or semantic segmentation. The loss of the prediction module is used to constrain the accuracy of traffic participant trajectory prediction or vehicle motion prediction; The planning module loss is used to constrain the safety, feasibility, and comfort of the generated trajectory; For switching losses; ， ， ， These are the weighting coefficients.

Images