Adaptive end-to-end decision planning method for vehicle-mounted edge computing platform

Abstract

The application discloses a kind of adaptive end-to-end decision planning methods for vehicle-mounted edge computing platform, method includes: obtaining the driving parameter of ego vehicle, navigation instruction and image frame sequence, and extracting state feature vector;According to state feature vector, calculate environmental uncertainty using pre-trained SQR network model, and calculate model uncertainty according to multi-SQR network integration calculation model, and add to obtain total uncertainty;Compare environmental uncertainty, model uncertainty, total uncertainty with the division threshold obtained according to uncertainty distribution statistics, determine the risk quadrant to which the current vehicle state belongs;According to the risk quadrant of current state, select the offline trained small expert model, medium expert model or large expert model to output future reference trajectory and target speed.The application realizes adaptive, safe and efficient end-to-end decision planning for multi-scene driving behavior under limited vehicle-mounted computing power.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicles, and in particular to an adaptive end-to-end decision planning method and system for in-vehicle edge computing platforms. Background Technology

[0002] Motion planning and decision control are core modules of autonomous driving systems. Their task is to generate safe, comfortable, and efficient driving trajectories or control commands in real time in complex and dynamic traffic environments. With the development of deep learning, end-to-end decision planning methods have gradually emerged. These methods typically use camera images, vehicle status, and navigation information as inputs, and utilize neural networks to output future trajectories or low-level control variables, thereby reducing manual modeling and error accumulation in the traditional "perception-prediction-planning-control" pipeline. In academia and industry, solutions based on end-to-end imitation learning and deep reinforcement learning have demonstrated strong learning capabilities and usability in simulation evaluations such as CARLA, and have become an important direction for intelligent decision-making in autonomous driving.

[0003] However, existing end-to-end methods often rely on a single model to address all driving scenarios, and typically lack explicit characterization of scenario risks and model confidence. When faced with numerous cross-scenario, long-tail, and high-risk interactions on open roads, a single model often struggles to simultaneously ensure safety, generalization ability, and in-vehicle real-time performance. Meanwhile, the limited computing power and power consumption of in-vehicle embedded platforms further amplify the difficulty of deploying end-to-end models in practice.

[0004] Motion planning requires vehicles to generate safe and efficient trajectories in real time within complex and dynamic traffic conditions. End-to-end decision planning methods, which directly regress trajectories or control commands from perception data using deep neural networks, have made progress on benchmarks such as CARLA. Current mainstream solutions mainly include two categories: Imitation Learning (IL) and Deep Reinforcement Learning (DRL).

[0005] However, existing methods still have significant shortcomings in unified modeling across multiple scenarios and vehicle deployment. IL (Input-Like) strategies are prone to distribution shifts during mixed training across multiple scenarios, and the optimal behavior differs significantly across different scenarios. For example, highway scenarios emphasize speed and distance, while urban intersections emphasize slowing down and yielding, making it difficult for a single model to simultaneously learn shared features and scenario-specific patterns. Training gradients interfere with each other, resulting in poor stability and often converging to a conservative strategy. While DRL (Directed Regression) methods have some adaptability to out-of-distribution states, their reward design is complex, and their driving style is monotonous, making it difficult to dynamically adjust behavior according to scenario risks. Furthermore, the large number of end-to-end model parameters and inference overhead also limit their real-time application on computationally limited in-vehicle embedded platforms.

[0006] In summary, existing technologies lack an end-to-end framework capable of explicitly assessing scenario risks and adaptively adjusting model capacity and driving behavior, and have not yet simultaneously met the requirements for performance across multiple scenarios, policy stability, and onboard computing power constraints. Therefore, there is an urgent need for a new end-to-end decision-making and planning framework that can identify scenario risks, allocate model capacity on demand, and adapt to onboard computing power constraints. Summary of the Invention

[0007] The main objective of this invention is to provide an adaptive end-to-end decision planning method and system for vehicle edge computing platforms, enabling adaptive, safe, and efficient end-to-end decision planning for driving behaviors in multiple scenarios under limited vehicle computing power.

[0008] The technical solution adopted in this invention is: An adaptive end-to-end decision planning method for in-vehicle edge computing platforms is provided, including the following steps: The system acquires the vehicle's driving parameters, navigation commands, and the latest image frame sequence captured by the forward-facing camera, and extracts state feature vectors. Based on the state feature vector, the environmental uncertainty is calculated using a pre-trained SQR quantile regression network model, and the model uncertainty is calculated by integrating multiple SQR quantile regression networks. The total uncertainty is obtained by summing the results. By comparing environmental uncertainty, model uncertainty, total uncertainty, and classification thresholds obtained from uncertainty distribution statistics, the risk quadrant to which the current vehicle status belongs is determined. Based on the risk quadrant of the current state, select and activate the offline-trained small expert model, medium expert model, or large expert model to output the future reference trajectory and target velocity; among them, the small expert model is used for low-risk scenarios; the medium expert model is used for highly random interaction scenarios; and the large expert model is used for high-risk scenarios.

[0009] Following the above technical solution, when both environmental uncertainty and model uncertainty are less than the corresponding thresholds, the corresponding scenario is a low-risk, easy scenario in the lower left quadrant; when environmental uncertainty is less than the corresponding threshold and model uncertainty is greater than or equal to the corresponding threshold, the corresponding scenario is a structural blind spot scenario in the upper left quadrant; when environmental uncertainty is greater than or equal to the corresponding threshold and model uncertainty is less than the corresponding threshold, the corresponding scenario is a high-randomness interaction scenario in the lower right quadrant; and when both environmental uncertainty and model uncertainty are greater than the corresponding thresholds, the corresponding scenario is a high-risk, rare scenario in the upper right quadrant.

[0010] Following the above technical solution, when the total uncertainty continues to exceed the preset upper limit threshold for a certain time window, or when the environmental uncertainty alone exceeds the highest uncertainty threshold, the rule expert model takes over and generates a safety control command for safe stopping, deceleration and avoidance, or low-speed maintenance, thus safely downgrading from a learning strategy to a verifiable rule strategy.

[0011] Following the above technical solution, the training process of the SQR quantile regression network model is as follows: construct a deep reinforcement learning coach model based on PPO proximal policy optimization; use the coach model to sample vehicle trajectories in a simulation environment under multiple map, multiple weather, and multiple traffic density configurations; replay all collected vehicle trajectories and calculate the comprehensive expected labels to obtain a dataset with expected labels; train the pre-constructed SQR quantile regression network model based on the dataset with expected labels.

[0012] Following the above technical solution, the multi-SQR quantile regression network ensemble is specifically as follows: multiple independent SQR networks are trained using data resampling and random initialization to form a deep ensemble; when evaluating any state, multiple preset quantiles are calculated for each SQR network, and the total uncertainty is decomposed into environmental uncertainty and model uncertainty through cross-model and intra-model variance decomposition.

[0013] Following the above technical solution, a multi-SQR quantile regression network is used to label each state in the vehicle trajectory collected by the coach model with environmental uncertainty and model uncertainty. The state space is divided into four quadrants with environmental uncertainty as the horizontal axis and model uncertainty as the vertical axis. For the data in different quadrants, three different sizes of student models and a high-precision mechanistic motion planner are selected for data correction and training. Under the constraint of limited computing power on the vehicle, adaptive modeling of scenarios with different difficulty and risk levels is achieved, resulting in small expert models, medium expert models and large expert models.

[0014] Following the above technical solution, when trajectory prediction is performed using a small expert model, a medium expert model, or a large expert model, if the prediction time exceeds the inference delay threshold or an abnormal signal is detected, the current expert model prediction is interrupted and switched to other control strategies to ensure that the vehicle maintains a controllable and safe motion state when computing power degrades.

[0015] Following the above technical solution, the capacity of the three expert models is dynamically matched according to the scenario at the algorithm implementation level.

[0016] The present invention also provides an adaptive end-to-end decision planning system for vehicle-mounted edge computing platforms, comprising: The data acquisition module is used to acquire the vehicle's driving parameters, navigation instructions, and the latest image frame sequence captured by the forward-facing camera, and to extract state feature vectors. The uncertainty estimation module is used to calculate environmental uncertainty based on the state feature vector using a pre-built SQR quantile regression network model, and to calculate model uncertainty by integrating multiple SQR quantile regression networks, and then sum them to obtain the total uncertainty. The environmental uncertainty, model uncertainty, and total uncertainty are compared with the classification threshold obtained by uncertainty distribution statistics to determine the risk quadrant to which the current vehicle state belongs. The lightweight expert routing module is used to select and activate an offline-trained small expert model, medium expert model, or large expert model to output the future reference trajectory and target velocity based on the risk quadrant to which the current state belongs. Among them, the small expert model is used for low-risk scenarios; the medium expert model is used to handle highly random interaction scenarios; and the large expert model is used for high-risk scenarios.

[0017] Following the above technical solution, the lightweight expert routing module is also used to take over when the total uncertainty continues to exceed the preset upper limit threshold for a certain time window, or when the environmental uncertainty alone exceeds the highest uncertainty threshold. In this case, the rule expert model takes over and generates a safety control command for safe stopping, deceleration and avoidance, or low-speed maintenance, thus safely downgrading from a learning strategy to a verifiable rule strategy.

[0018] The present invention also provides a computer storage medium storing a computer program executable by a processor, the computer program being used to implement the adaptive end-to-end decision planning method for vehicle-mounted edge computing platforms described in the above technical solution.

[0019] The beneficial effects of this invention are as follows: This invention is a lightweight uncertainty assessment scheme based on network distillation, enabling the vehicle end to output environmental uncertainty and model uncertainty with only minimal parameters and latency increments. It achieves quantitative assessment of environmental randomness and model cognitive blind spots, providing a reliable basis for multi-expert dynamic routing. It realizes a multi-expert collaborative mechanism where "small expert models handle daily low-risk scenarios, medium models handle highly random interactive scenarios, and large models handle high-risk rare scenarios." The vehicle end relies only on its own vehicle's driving parameters, navigation commands, and image frame sequences. Through the two key innovations of "risk identification and expert model collaborative decision-making" and "lightweight deployment of risk assessment distillation," it achieves adaptive, safe, and efficient end-to-end decision planning for multi-scenario driving behavior under limited onboard computing power.

[0020] Furthermore, this invention constructs a multi-capacity expert system that can be invoked on demand, balancing performance and real-time performance under conditions of limited onboard computing power. Since the vast majority of vehicle states are in the low-risk quadrant during operation, this invention can activate only small or medium-sized models for the majority of the time, thereby significantly reducing average inference latency and computing power consumption. In truly complex or dangerous scenarios, large experts can be activated in a timely manner to ensure the safety and accuracy of decision-making. This design of allocating model capacity on demand enables the end-to-end planning system to maintain both high performance and high real-time performance on resource-constrained edge platforms.

[0021] Furthermore, uncertainty distillation and rule-based fallback are introduced to achieve a verifiable safety assurance mechanism. This invention distills the uncertainty assessment capabilities of multi-model integration into a lightweight prediction head embedded in the online inference process, enabling the vehicle-mounted system to possess continuous risk perception capabilities while maintaining extremely low computational overhead. When real-time uncertainty exceeds a threshold or the inference cycle becomes abnormal, the system automatically switches to rule experts or traditional stable control strategies, achieving a safe degradation from "learning-based decision-making" to "verifiable rule-based strategies." This mechanism effectively avoids uncontrollable decision-making behavior that deep models may exhibit in off-distribution scenarios, significantly improving the overall system's safety robustness.

[0022] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of an adaptive end-to-end decision planning method for an in-vehicle edge computing platform according to an embodiment of the present invention; Figure 2 This is a flowchart illustrating the selection of an expert model based on uncertainty in an embodiment of the present invention; Figure 3A This is a flowchart of the model construction and lightweight distillation process according to an embodiment of the present invention; Figure 3B yes Figure 3A A flowchart illustrating the specific process of constructing a scenario uncertainty assessment model based on multi-SQR network integration; Figure 4 This is a four-quadrant dataset divided according to environmental uncertainty and model uncertainty when training an expert model in an embodiment of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0026] It should be noted that the illustrations provided in the embodiments of the present invention are only schematic representations of the basic concept of the present invention. Therefore, the illustrations only show the components related to the present invention and are not drawn according to the number, shape and size of the components in actual implementation. In actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0027] In this invention, it should also be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application. Furthermore, the terms "first" and "second" are used only for descriptive and distinguishing purposes and should not be construed as indicating or implying relative importance.

[0028] Furthermore, it should be noted that the features of the various embodiments of the present invention can be combined or integrated in whole or in part, and as those skilled in the art will understand, they can interact and operate in different ways. Each embodiment can be implemented independently of each other or in association with one another.

[0029] To address the challenge of a single end-to-end deep learning model simultaneously achieving "full-scenario adaptability, driving safety, and in-vehicle real-time performance," this invention proposes a highly efficient and reliable end-to-end decision-making and planning system for in-vehicle edge computing platforms. The overall approach can be summarized in two steps. First, the system learns to "judge the danger of the current scenario and the degree of uncertainty it faces." We obtain a high-level coaching model through deep reinforcement learning and conduct uncertainty risk assessments of its future performance based on multiquantile regression ensembles, categorizing scenarios into four risk levels. Second, models of different "capability levels" are assigned specific roles. Small, medium, and large expert models are trained for the four risk scenarios respectively. During vehicle operation, the appropriate expert outputs the trajectory and speed based on real-time risk assessment results. This allows for rapid decision-making in most low-risk, everyday scenarios by utilizing a lightweight expert, while switching to a stronger expert in high-risk, complex scenarios, thereby significantly reducing the average computing power overhead in the vehicle without compromising safety margins.

[0030] To ensure that the aforementioned risk assessment can run in real time on vehicle-mounted platforms with limited computing power, this invention further proposes a lightweight uncertainty assessment scheme based on network distillation. During the training phase, deep ensemble of multiple models is used to obtain high-precision risk "truth values." During the deployment phase, this ensemble capability is distilled into a lightweight prediction head attached to the backbone network. This allows the vehicle to output aleatoric and epistemic uncertainties with minimal parameter and latency increments, providing a reliable basis for multi-expert dynamic routing. Through these two key innovations—"risk identification and expert collaborative decision-making" and "lightweight deployment of risk assessment distillation"—this invention achieves adaptive, safe, and efficient end-to-end decision planning for multi-scenario driving behavior under limited onboard computing power.

[0031] like Figure 1 As shown, the adaptive end-to-end decision planning method for vehicle-mounted edge computing platforms according to an embodiment of the present invention includes the following steps: S1. Obtain the vehicle's driving parameters, navigation commands, and the latest image frame sequence captured by the forward-facing camera, and extract the state feature vector; S2. Based on the state feature vector, calculate the environmental uncertainty using the pre-trained SQR quantile regression network model, and calculate the model uncertainty using the multi-SQR quantile regression network ensemble, and add them together to obtain the total uncertainty. S3. Compare environmental uncertainty, model uncertainty, and total uncertainty with the classification thresholds obtained from uncertainty distribution statistics to determine the risk quadrant to which the current vehicle status belongs. S4. Based on the risk quadrant of the current state, select and activate the offline-trained small expert model, medium expert model, or large expert model to output the future reference trajectory and target velocity. Among them, the small expert model covers most low-risk daily scenarios and scenarios with rare structures but simple environments; the medium expert model handles highly random interactive scenarios; and the large expert model is used to deal with high-risk and rare scenarios with high environmental uncertainty and high model uncertainty.

[0032] This invention employs a multi-SQR network ensemble (SQR: Simultaneous Quantile Regression) to achieve an uncertainty assessment method combining "multi-model ensemble + quantile regression," used to model the probability distribution of the future performance of an end-to-end autonomous driving coach model under different states. Uncertainty is categorized into two types: aleatoric uncertainty, stemming from environmental randomness and irreducible noise; and epistemic uncertainty, arising from insufficient training data or model limitations.

[0033] During the vehicle-mounted deployment phase, the main focus is on each state. Output four quantities: 1) Median estimate of future composite score This indicates the typical behavior of the strategy in this state; 2) Aleatoric environmental uncertainty To measure performance fluctuations caused by environmental noise; 3) Epistemic model uncertainty To measure the reducibility risk caused by insufficient model knowledge; 4) Total uncertainty This is the sum of the first two types of uncertainty.

[0034] When total uncertainty A significant increase typically corresponds to highly stochastic scenarios such as complex intersections or high-traffic interactions; even if the model is confident, its future performance may fluctuate greatly. Conversely, when model uncertainty increases... A significant increase indicates that the data for that state is scarce or belongs to a novel scenario, and can be used as a trigger signal for degradation control or strategy switching.

[0035] like Figure 3A , 3B As shown, in order to construct the SQR quantile regression network model, firstly, a DRL coach model (DRL: Deep Reinforcement Learning) is trained to provide a high-quality demonstration for uncertainty assessment and subsequent imitation learning.

[0036] The specific process is as follows: 1) Simulation Interactive Training: A training environment is constructed in CARLA that includes multiple city scenarios, road network structures, and dynamic traffic participants. At each step of the simulation, the scenario state is output, and the coach model outputs continuous low-level control commands (steering, acceleration, or braking). The environment returns a reward and a termination flag. The coach model does not rely on manual data or rule demonstrations, but uses the PPO algorithm to iteratively update the strategy and value network through repeated interactions until the metrics converge.

[0037] 2) Input / Output Design: The state input consists of a bird's-eye view (BEV) semantic image and the vehicle's measurement vectors to reduce learning difficulty and improve generalization ability. The BEV semantic image is a multi-channel bird's-eye view, including the drivable area, lane lines, planned route, traffic signals, and the temporal positions of surrounding vehicles and pedestrians. The measurement vectors include vehicle states not explicitly represented in the BEV, such as steering angle, throttle, braking, gear position, and lateral and longitudinal speeds. The policy network encodes the two inputs and outputs the probability distribution parameters of steering and longitudinal acceleration. The action is modeled using a bounded Beta distribution, which naturally satisfies the physical constraints of the control variables and facilitates the calculation of entropy and KL divergence (KL: Kullback-Leibler Divergence).

[0038] 3) Rewards and Goal-Oriented Exploration: Guiding the strategy to learn safe, smooth, and efficient driving. Dense rewards include at least speed tracking rewards, path deviation rewards, attitude alignment rewards, and steering smoothness penalties. Termination events such as collisions, running red lights or stopping signs, deviating from the route, and prolonged congestion correspond to significant penalties. In addition, event-related exploration losses are introduced. For trajectories containing failure events, the KL divergence between the current action distribution and the prior event distribution (e.g., collision corresponds to deceleration prior, congestion corresponds to acceleration prior) is calculated several steps before failure and used as an additional optimization term, thereby encouraging the strategy to explore in the direction of avoiding failure during high-risk phases.

[0039] 4) Parallel Sampling and Convergence Consolidation: Simulation instances are launched in parallel under multiple map, weather, and traffic density configurations, with trajectories sampled synchronously using the same strategy. After each round of sampling, the strategy and value network are updated multiple times, and learning rate scheduling and gradient pruning are combined to improve training stability. Training stops when the success rate of the validation set, driving score, and other indicators stably reach the threshold, and the reinforcement learning coach model is consolidated.

[0040] Next, an expert dataset is constructed based on the coaching model. This dataset is used to train small, medium, and large expert models. This invention treats the trained and converged coaching policy as a black-box decision-maker. By continuously sampling trajectories under multiple CARLA maps, traffic densities, and weather configurations, a supervised learning dataset for "state → future comprehensive score" is constructed offline. Then, a conditional quantile regression network (SQR) is introduced, and quantitative estimation of uncertainty is achieved through multi-SQR quantile regression network model ensemble and variance decomposition.

[0041] The PPO policy that has already converged and been trained on the CARLA simulation platform will be used. Fixed parameters (PPO: Proximal Policy Optimization) are used as a fixed black-box driving policy. This is applied under various urban scenarios, road network structures, traffic participant behavior parameters, and weather conditions. Continuously sample a large number of driving trajectories. Project trajectory-level metrics to the state level for each time step in the trajectory. Calculate the sample value of the "Future Composite Score" from the start to the end of this state. Construct state-level random variables The state at each time step is represented. With corresponding tags Composition of supervised learning sample pairs High-risk segments are then resampled or reweighted to obtain a supervised dataset for training. .

[0042] For each track collected Calculate a comprehensive performance / risk scalar This metric uniformly considers discounted rewards as well as penalties for termination events such as collisions, running red lights, and prolonged blockages. In one embodiment, the CARLA simulation environment outputs a reward at each step. The reward already includes dense terms such as speed deviation, path following error, attitude deviation, and ride comfort penalties. Simultaneously, the simulation environment also marks termination events such as collisions, running red lights / stop signs, and prolonged congestion. For time intervals from... The starting trajectory Define the trajectory-level comprehensive scoring function as follows:

[0043] in As a discount factor, These indicate whether events such as collisions, running red lights / stop signs, and prolonged traffic jams have occurred within the trajectory. For additional penalty weights corresponding to the events, the coefficients from the PPO training process are reused.

[0044] The specific steps for constructing the dataset are as follows: ① Parallel startup of environmental instances: Configure multiple different maps (such as Town01, Town02, etc.), different weather conditions, and different traffic participant densities and behavioral parameters in the CARLA simulation platform. Start multiple simulation instances in parallel.

[0045] ② Coaching strategy execution and data recording: In each simulation instance, a fixed coaching strategy is used. Starting from a randomized starting point and target point, run for several rounds. For each time step... Record the following information: Multi-channel BEV bird's-eye view semantic image This includes the drivable area, lane lines, planned routes, traffic signals, and the spatial distribution of other vehicles and pedestrians; vehicle measurement vectors. This includes current steering angle, throttle, brakes, gear, lateral and longitudinal speed, etc.; current instant rewards. Whether to terminate and the type of termination event (collision, running a red light / stop sign, being blocked, etc.).

[0046] ③ Trajectory playback and label calculation: Play back each complete trajectory and calculate the label for each time step in the trajectory. Calculate the corresponding future comprehensive score label according to the aforementioned formula. .

[0047] To make the uncertainty assessment network focus more on high-risk and rare scenarios, the samples are filtered and reweighted after trajectory acquisition. For trajectories containing collisions or serious violation termination events, the state time step samples in the last few seconds before termination (the last 3 to 5 seconds) are oversampled or given higher weight in the loss function; for simple cruise segments with long periods of no events, appropriate downsampling is performed to reduce the excessive bias of data distribution towards low-risk scenarios. The final supervised dataset can be represented as follows: in This is a sample of the future comprehensive score for the corresponding state.

[0048] SQR quantile regression network design: a conditional quantile regression network with input consistent with the coaching model. Given state features and quantile levels The network outputs quantile estimates of future scores. It is trained using the simultaneous quantile regression loss. The SQR uncertainty assessment network architecture consists of three parts: state input encoding, quantile conditional encoding, regression head, and output.

[0049] ①State Input Encoding: To reduce implementation complexity and ensure feature space consistency, this invention adopts a state encoding structure similar to or the same as the PPO coaching model. This includes a BEV image branch and a measurement vector branch. The BEV image branch inputs a multi-channel BEV bird's-eye view semantic image into a convolutional neural network or a residual network (ResNet), which then passes through several convolutional layers, pooling layers, and fully connected layers to extract visual feature vectors. The measurement vector branch inputs measurements such as the current steering angle, throttle, brake, gear position, lateral speed, and longitudinal speed into a fully connected network of several layers to obtain the measurement feature vector. Finally, the state features are concatenated, combining the two feature sets mentioned above to obtain the complete state feature vector. .

[0050] ② Quantile conditional encoding: In order to learn different quantiles simultaneously in the same network, quantile levels are... Perform conditional encoding. First, convert the scalar... Mapped to quantile embedding vectors through one or two layers of fully connected networks. Then, the state features are concatenated with the quantile embedding vector to obtain the joint feature vector. This design enables the network to adapt to different... Instead of training an independent head for each quantile, predict the corresponding conditional quantile.

[0051] ③ Regression Head and Output: Joint Feature Vector Input a fully connected network of several layers and output a scalar prediction. .in Indicates parameters SQR uncertainty assessment network, For state Lower quantile level The corresponding future overall score estimate.

[0052] The loss function of SQR is the quantile regression pinball loss, which applies to a single sample. and target quantile level The optimal solution to this loss corresponds to the conditional distribution of the th Quantiles. The pinball loss for quantile regression is defined as:

[0053] The specific training process is as follows: for each mini-batch, from the dataset... Extraction Sample For each sample, a quantile level is independently sampled from a uniform distribution. Calculate the average loss of this mini-batch:

[0054] By continuously sampling different And optimize the above-mentioned losses, network Can be throughout Learning conditional quantile functions on intervals .

[0055] Multi-model ensemble and uncertainty decomposition, using Multiple independent SQR networks were trained using data resampling and random initialization to form a deep ensemble. In evaluating any state At that time, calculate for each network separately. Quantiles were used to decompose the total uncertainty into aleatoric and epistemic components through cross-model and intra-model variance decomposition. The specific steps are as follows. From the original supervised dataset... The method of sampling with replacement is used to construct training subset For each subset Each SQR network is trained independently using different randomly initialized weights and different optimized hyperparameters. Achieving multi-model ensemble Integration scale The value should be between 10 and 25 to balance the stability of the estimate and the computational cost.

[0056] The specific calculation method for the uncertainty index is as follows: During the evaluation phase, for any state to be evaluated This invention sequentially calculates several key quantiles on each model. For example, for each model , median (approximate conditional median) lower quantile upper quantile The median estimate of the ensemble is obtained by averaging the median predictions from each model within the ensemble.

[0057] This invention is based on the concept of variance decomposition, which breaks down the total uncertainty into... and It consists of two parts, as defined below.

[0058] ①Aleatoric uncertainty: For each model The internal "class standard deviation" estimate is defined as:

[0059] At the integration level, all models will be integrated. The squared average yields an estimate of the aleatoric uncertainty:

[0060] This reflects the inherent fluctuations in future scores caused by environmental randomness and observation noise, given model parameters.

[0061] ②Epistemic uncertainty: Epistemic uncertainty stems from the uncertainty of model parameters and insufficient training data coverage, and is reflected in the degree of consistency among different models in predicting the median. This invention adopts the following definition:

[0062] When training data is scarce near a certain state or there are out-of-distribution samples, different models... The predicted discrepancy for quantiles will increase significantly. The corresponding increase.

[0063] ③Total uncertainty: In this invention, the total uncertainty can be expressed as the sum of aleatoric and epistemic uncertainties. .

[0064] This invention, building upon the aforementioned reinforcement learning coach and its uncertainty assessment model, no longer trains only a single end-to-end decision-making network, but instead constructs a multi-model collaborative framework. This framework first utilizes the aforementioned multi-SQR ensemble uncertainty assessment network to assign a label to each state in the coach trajectory. Uncertainty and Uncertainty label, then with Uncertainty is represented by the horizontal axis, and... Uncertainty is used as the vertical axis, and the state space is divided into four quadrants. Three different scales of student models and a high-precision mechanistic motion planner are selected for data correction and training for data in different quadrants, so as to achieve adaptive modeling of scenarios with different difficulty and risk levels under the constraint of limited computing power on the vehicle.

[0065] Under CARLA's multi-map, multi-traffic-density, and multi-weather configurations, the aforementioned PPO coaching strategy is used. Taking a bird's-eye view semantic raster image and the vehicle's state as input, the output is a sequence of local reference trajectory points and / or discrete driving intention labels for several future time points, resulting in large-scale coaching decision-making and planning trajectory data. After obtaining the original coaching trajectory, uncertainty is labeled for each state in the trajectory. For the state representation at any time step in the trajectory (multi-channel BEV bird's-eye view semantic image, vehicle measurement vector, and navigation commands), the aforementioned multi-SQR quantile regression deep ensemble model is input to obtain the median estimate of the future comprehensive score corresponding to that state, as well as Aleatoric and Epistemic uncertainty labels. These two uncertainty scalars are denoted as... (Aleatoric uncertainty) is used to quantify performance changes caused by irreducible noise such as environmental randomness and fluctuations in the behavior of other traffic participants. (Epistemic uncertainty) is used to quantify reducible uncertainty caused by factors such as insufficient training data coverage and limited model expressive power.

[0066] This invention is based on the training set and The statistical distribution is analyzed, and a preset threshold or quantile is selected as the dividing boundary, for example, the median of Aleatoric uncertainty is selected respectively. and median of Epistemic uncertainty .by For the horizontal axis, With the vertical axis as the center, The system is divided into four quadrants: ① Lower left quadrant (low Aleatoric, low Epistemic): This quadrant corresponds to "easy scenarios" with low environmental noise and sufficient model knowledge, such as straight-line cruising with clear rules and stable traffic participant behavior, and simple left / right turns. ② Upper left quadrant (low Aleatoric, high Epistemic): This quadrant corresponds to "structural blind spot scenarios" where the environment itself is relatively stable but training data is scarce or the strategy has not been fully learned, such as rare intersection topologies. ③ Lower right quadrant (high Aleatoric, low Epistemic): This quadrant corresponds to "highly random interactive scenarios" where training data coverage is sufficient but inherent noise is high due to random behavior of other vehicles, pedestrian crossings, and short-term visual obstruction. ④ Upper right quadrant (high Aleatoric, high Epistemic): This quadrant corresponds to "high-risk rare scenarios" where there is both strong environmental randomness and significant model knowledge deficiency, such as rare complex intersections, multi-vehicle lane contention under extremely high traffic conditions, and road sections with construction zones.

[0067] like Figure 4 As shown, this invention classifies all state samples according to their The quadrant in which it is located is divided into four subsets, denoted as follows: , , and The quadrant label is explicitly recorded in the sample tuple, providing a basis for subsequent multi-expert training and inference routing.

[0068] For the four quadrants, this invention employs differentiated data processing strategies. For the lower left quadrant... and the lower right quadrant Since the model uncertainties are all low, the reference trajectory or high-level intent output by the PPO coach can be directly regarded as a reliable supervision signal without additional correction. For the upper left quadrant... and the upper right quadrant Although aleatoric uncertainty is low in the upper left quadrant, epistemic uncertainty is high, indicating that the coaching strategy has a high risk of cognitive inadequacy or overfitting in this type of state. To avoid directly applying potentially biased or unrobust coaching demonstrations to student training, this invention introduces a high-precision mechanistic motion planner to perform offline correction on the data in these two quadrants based on global privileged information.

[0069] The high-precision mechanistic motion planner includes a vehicle dynamics model, vehicle kinematic constraints, a collision detection module, and a global cost function. It also has access to global privileged information within the simulation environment, such as high-precision maps, complete planned routes, precise locations and speeds of all traffic participants, and even short-term future trajectory predictions. For and In each state, this invention uses the local trajectory or high-level intent output by the coach as initial guidance or soft constraints. A mechanistic planner then re-optimizes the trajectory under stricter safety and comfort constraints to obtain a high-precision reference trajectory. Alternatively, it adjusts and trims collision-prone, overly aggressive, or overly conservative segments of the coach trajectory to obtain corrected planning output labels. The corrected data form new data subsets: The sample set in the upper left quadrant, modified by mechanistic planning. The sample set in the upper right quadrant is modified by mechanism planning.

[0070] The multi-expert adaptive end-to-end decision programming model proposed in this invention includes three expert models: Expert Small, Medium, and Large. Based on the dataset constructed in the aforementioned steps, imitation learning is used to obtain expert models of different capacities.

[0071] Expert Small models are used in the lower left and upper left quadrants. It adopts a lightweight structure of "image encoder + measurement encoder + planning head".

[0072] Image encoder: Input is a front-view RGB camera image. ( (Resolution). The front-end uses a ResNet 18 with pruned channels and layers as the backbone network, removing the classification layers and retaining only convolutional and residual blocks to encode the image into a one-dimensional visual feature vector. .

[0073] Measurement encoder: Input measurement vector This includes low-dimensional quantities such as the vehicle's current speed, steering angle, throttle, brakes, gear position, and a two-dimensional direction vector pointing to the next navigation waypoint. The measurement encoder consists of two fully connected network layers. -ReLU- Mapping the input to a measurement feature vector .

[0074] Feature fusion and bottleneck layer: and By concatenating the features along the feature dimension, we obtain the fused features. Through two layers of fully connected networks -ReLU-FC - ReLU yields bottleneck characteristics As a unified representation of driving semantics, FC stands for Fully Connected Layer, and ReLU stands for Rectified Linear Unit.

[0075] Multi-branch planning head: Selects the corresponding branch based on high-level navigation instructions {go straight, turn left, turn right, keep lane, change lane}. Each branch contains a trajectory head and a speed head. Trajectory head: Two-layer fully connected network. Output from the vehicle coordinate system A future reference trajectory point Speed ​​Header: Two-Layer Fully Connected Network Output the current or short-term target speed. This allows the lower-level PID or MPC controller to track the signal.

[0076] Small models in the lower left quadrant dataset and the top left quadrant corrected dataset Joint training is performed on the same network. For Medium to large batches of well-structured and low-noise "everyday scenarios" are used to reinforce the local reference trajectories output by the learning coach. and measured speed As the primary supervisory signal, it enables the model to fit the coach's decision-making patterns in common scenarios. For In scenarios with relatively simple environments but rare structural morphologies or scarce training samples, known as "structural blind spots," a high-precision mechanistic motion planner generates corrected trajectories based on global privileged information. As the primary supervisor label, while retaining the coach's trajectory. As a soft constraint, it improves the safety and rationality of such scenarios without changing the computational cost of small models.

[0077] The training steps for a small model can be summarized as follows: 1) Based on the uncertainty assessment results, the coach trajectory samples are divided into... and ,in The trajectory has been corrected offline using a mechanistic motion planner.

[0078] 2) From The system randomly samples mini-batches at a preset ratio and uses Huber loss to minimize the predicted trajectory. Deviation from the target trajectory: Sample, with For the goal; to Sample, with The primary objective is to add a soft loss constraint with a small weight. Do not deviate Too many. The velocity sensor uses mean square error to regress the target velocity.

[0079] 3) Employ the Adam (Adam Adaptive Moment Estimation) optimizer (initial learning rate) (Batch size 64), combined with learning rate decay and gradient clipping, monitors the trajectory error and task completion rate of the validation set in daily scenarios and structural blind spot scenarios, and stops training and solidifies the model parameters after reaching the preset indicators.

[0080] Through the above structure and training process, the small model can cover most low-risk daily scenarios while maintaining low parameter count and inference latency, and can also safely handle the top left quadrant scenarios with rare structures and scarce data with the assistance of the mechanism planner.

[0081] Medium-sized models are used in the lower right quadrant. A fixed structure of "multi-frame image encoder + measurement encoder + feature fusion + multi-branch planning head" is used to model the lower right quadrant dataset. Highly random interaction scenarios with medium to high aleatoric and low epistemic characteristics.

[0082] Multi-frame image encoder: Input is the most recent Frame-forward RGB camera image sequence In implementation, Frame images are stacked in the channel dimension to form a size of tensor ( The resolution is The front-end uses a full ResNet-34 as the backbone network, removing the classification head and retaining only the convolutional and residual blocks to encode multiple frames of images into one-dimensional visual feature vectors. .

[0083] Measurement encoder: Input measurement vector This includes low-dimensional state variables such as vehicle speed, steering angle, throttle, braking, gear position, and a two-dimensional direction vector pointing to the next navigation waypoint. The measurement encoder consists of two fully connected network layers, such as... -ReLU- Mapping the input to a measurement feature vector .

[0084] Feature fusion and bottleneck layer: concatenating visual features and measurement features to obtain... Input two-layer fully connected network FC -ReLU, to obtain bottleneck characteristics , as a unified driving semantic representation describing the current highly random interaction state.

[0085] Multi-branch planning head: based on high-level navigation instructions Go straight, turn left, turn right, stay in your lane, change lanes Select the corresponding branch. Each branch contains a trajectory header and a velocity header. The trajectory header consists of a two-layer fully connected network. Output from the vehicle coordinate system Prediction of future trajectory points Speed ​​Header: Two-Layer Fully Connected Network Output the current or short-term target vehicle speed This allows the lower-level PID (Proportional-Integral-Derivative) or MPC (Model Predictive Control) controllers to track the signal.

[0086] In terms of training strategy, the medium-sized student model is only used on the lower right quadrant dataset. Training on this topic. Aleatoric uncertainty in this type of scenario. The relatively high level of uncertainty, coupled with the low level of epistemic uncertainty, indicates that the coach has considerable experience in these states, but future performance is still significantly affected by environmental randomness. This invention employs a behavioral cloning training method that combines coach-led supervision with uncertainty reweighting.

[0087] Medium model The training steps and details can be summarized as follows: 1) Data preparation and uncertainty labeling: Construct training sample pairs from the dataset in the lower right quadrant ( ),in To enhance the learning coach's role at all times Output local reference trajectory, For the corresponding target or actual vehicle speed, The Aleatoric uncertainty label is given to the SQR uncertainty assessment network.

[0088] 2) Weighted Behavior Cloning Training: For each The sample weights in the sample are defined according to the Aleatoric uncertainty:

[0089] in These are hyperparameters used to moderately reduce the impact of high-noise samples on the overall loss. For each sample, the trajectory regression loss and velocity regression loss are calculated:

[0090] The overall loss of a medium-sized student model can be written as:

[0091] in for size, This is the loss weighting coefficient.

[0092] 3) Optimization and convergence determination: The Adam optimizer was used (initial learning rate set to 1). (batch size 64), combined with learning rate decay and gradient pruning to ensure training stability. On the validation set, metrics such as trajectory error, task completion rate, and collision rate in high-traffic, complex interaction scenarios are monitored. When these metrics stabilize and reach preset thresholds, training is stopped and the parameters of the medium-sized student model are fixed.

[0093] A large student model is used in the upper right quadrant. This model employs the largest network structure in terms of parameters, introducing higher-resolution feature maps, deeper layers, or temporal Transformers into the visual encoder to enhance its ability to model details and long-term dependencies in high-risk, rare scenes. This model primarily refines the dataset in the upper right quadrant. During training, the supervision signal mainly consists of high-precision trajectories generated by a mechanistic motion planner, supplemented by auxiliary supervision such as coach value estimation and uncertainty labeling. This enables the model to learn conservative, safe, and feasible decision-making strategies even in scenarios that are rare but have the greatest potential harm.

[0094] Large student model It adopts a high-capacity structure of "multi-frame high-resolution image encoder + temporal aggregation module + measurement encoder + feature fusion + multi-branch planning head": Multi-frame high-resolution image encoder: Input is the most recent Frame-forward RGB image sequence The resolution of each frame is Each frame of image is fed into a ResNet 50 backbone network with shared weights. The classification layer is removed, and only the convolutional and residual blocks are retained to obtain frame-level visual feature vectors. .Will The frame-level features are arranged in chronological order to form a sequence. Inputting two layers of lightweight temporal Transformer encoders (4 attention heads) yields temporally aggregated visual features. It is used to characterize long-term interactive information in high-risk scenarios.

[0095] Measurement encoder: Input measurement vector This includes the vehicle's current speed, steering angle, throttle, braking, gear position, a two-dimensional direction vector pointing to the next navigation waypoint, and optional historical acceleration. The measurement encoder consists of two fully connected networks: Output measurement feature vector .

[0096] Feature fusion and bottleneck layer: concatenating temporal visual features with measurement features to obtain... Through two layers of fully connected networks Bottleneck characteristics were obtained. , serving as a unified semantic representation for driving in high-risk and rare scenarios.

[0097] Multi-branch planning head: based on high-level navigation instructions Select the corresponding branch. Each branch contains a trajectory header, a velocity header, and a value header. The trajectory header consists of a two-layer fully connected network. Output from the vehicle coordinate system A future reference trajectory point Speed ​​Header: Two-Layer Fully Connected Network Output target vehicle speed Value Head: Two-Layer Fully Connected Network Output of coach value estimation regression value Explicitly encode information about "future risks and returns".

[0098] Large models only correct the dataset in the upper right quadrant. In the training, the supervision signal primarily consists of high-precision trajectories generated by a mechanistic motion planner, supplemented by coach value estimation and uncertainty labels. The training process can be simplified to the following steps: 1) Data Construction: For each sample time in the upper right quadrant Construct training tuples .in This refers to the corrected trajectory output by the mechanistic motion planner under global privileged information. To enhance the learning coaching trajectory, For coach valuation, and These are the Aleatoric and Epistemic uncertainty labels, respectively.

[0099] 2) Loss function design: For each sample, define the mechanistic trajectory regression loss:

[0100] Coach trajectory soft constraint loss (small weight):

[0101] Value regression loss:

[0102] Trajectory smoothing regularization:

[0103] This is used to suppress overly aggressive changes in trajectory curvature. Weights are assigned to samples based on uncertainty labels:

[0104] in is a hyperparameter used to emphasize rare scenarios with high epistemicity, i.e., "model cognitive inadequacy". The overall loss of a large model is then:

[0105] in This ensures that the output of the mechanism planner is the primary supervisory factor.

[0106] 3) Optimize the process: The Adam optimizer was used (initial learning rate set to 1). (Batch size 32), combined with learning rate decay and gradient clipping to control training stability. In constructing... Time to high Or, samples near high-risk termination events such as collisions or serious violations may be moderately oversampled to ensure that large models focus on learning scenarios that are rare but pose the greatest risk. The trajectory safety margin (minimum) in high-risk scenarios is monitored on the validation set. The training stops and the parameters of the large student model are fixed once the above indicators (minimum distance), task completion rate, and collision rate are stably satisfied with the preset thresholds.

[0107] Through the above structure and training design, the large student model achieves time-series performance. With the combined effect of strong supervision from a mechanistic motion planner, it can achieve high performance in the upper right quadrant. And high The scenario learns more conservative, safe, and feasible end-to-end decision-making and planning strategies.

[0108] In this invention, to address the limited computing power and power consumption of in-vehicle domain controllers, as well as the stringent real-time requirements of end-to-end decision-making and planning, a multi-expert adaptive deployment mechanism based on uncertainty assessment results is proposed. This mechanism, implemented on an in-vehicle edge computing platform, achieves an integrated online inference process from sensor input to uncertainty estimation, expert selection, and trajectory output, enabling the aforementioned uncertainty-based four-quadrant scene partitioning and multi-capacity expert model to be implemented in real-world vehicles.

[0109] In one embodiment, the in-vehicle computing platform consists of a domain controller integrating a CPU and a GPU or dedicated AI acceleration unit (NPU, DSP, etc.), equipped with limited on-chip cache and video memory, running a real-time operating system or automotive-grade Linux, and communicating with actuators such as steering and braking via in-vehicle Ethernet or CAN bus. The software stack includes middleware (ROS2, self-developed middleware), sensor drivers, and the multi-expert end-to-end decision planning module of this invention.

[0110] To adapt to hardware from different manufacturers, this invention uniformly converts the offline-trained uncertainty evaluation network and multi-capacity expert models (Expert S / Expert M / Expert L / R rule expert) into the intermediate representation format ONNX (Open Neural Network Exchange), and generates the corresponding inference graph or engine file according to the compilation toolchain of the target chip, ensuring that the same network structure can achieve low-latency and high-reliability online inference on multiple automotive edge platforms.

[0111] Based on the aforementioned design, this invention introduces a lightweight uncertainty assessment and expert routing module at the vehicle end, enabling adaptive selection of a suitable expert model based on real-time uncertainty assessment results within each control cycle. For example... Figure 2 As shown, the overall process includes the following steps: 1) Shared coding and lightweight uncertainty assessment During vehicle operation, each control cycle acquires the latest image frame sequence from the forward-facing camera and reads vehicle speed, steering angle, throttle, braking, gear position, and navigation commands from the vehicle bus. These multimodal inputs first pass through a shared, lightweight coding network to extract compact state feature vectors.

[0112] To avoid the computational overhead of directly running "multi-SQR deep integration" on the vehicle side, this invention distills the deep integration uncertainty evaluation network into a single uncertainty prediction head in the offline stage, and couples it with the aforementioned shared coding network, directly outputting the result on the vehicle side. Uncertainty estimation and Uncertainty estimation Output total uncertainty .

[0113] To further conserve onboard computing power, this invention does not construct a separate, independent uncertainty student network. Instead, it is applied as an additional branch (head) to the real-time decision-making and planning model running on the vehicle. The specific structural design is as follows: Shared Feature Extraction Layer: Reuses the backbone network of the decision-making and planning model, taking real-time sensor data (image sequences, measurement vectors) as input and outputting a high-dimensional environmental feature vector. This process does not add additional computational overhead. Independent Uncertainty Distillation Head: After the shared feature vector, a lightweight multilayer perceptron (MLP) is connected. This MLP contains two fully connected layers (ReLU), and the output layer contains two independent neurons, corresponding to the predicted values ​​of aleatoric uncertainty and epistemic uncertainty, respectively. Activation Function Design: Since the uncertainty value must be non-negative, the output layer uses the Softplus activation function to ensure that the predicted risk value conforms to the physical meaning.

[0114] Multi-objective regression distillation loss function: To enable the lightweight student network to simultaneously learn environmental noise and model cognitive bias, this invention designs a combined loss function that includes relative and absolute errors. During training, the shared backbone network parameters are frozen (or fine-tuned), and only the distillation head parameters are updated. The total distillation loss is defined as:

[0115] Where: Aleatoric regression loss The aim is to fit environmental noise. This is considering the large range of uncertainty values ​​(from 0.1 for low risk to [missing value] for high risk). The log-space smoothed L1 loss is used to balance the sensitivity to large and small values.

[0116] Epistemic regression loss The aim is to fit the cognitive blind spots of the model. Since epistemic uncertainty is a key signal triggering intervention by high-level experts, and only significantly increases in rare scenarios, a weighted mean squared error is introduced for high-uncertainty samples (i.e.,...). The threshold samples are given higher weights. This forces networks to focus on their ability to identify the "unknown":

[0117] 2) Scene risk assessment based on four-quadrant boundaries This invention reuses the partitioning threshold obtained from the uncertainty distribution statistics during the training phase during the deployment phase. Median of uncertainty and Median of uncertainty ), will be estimated in real time and Project onto the same two-dimensional plane and determine the risk quadrant to which the current state belongs: when and At that time, it corresponds to the easy scenario in the lower left quadrant, characterized by "low noise and low cognitive uncertainty"; when and When, it corresponds to the "structural blind spot" scenario in the upper left quadrant; when and When, it corresponds to a highly random interaction scenario in the lower right quadrant; when and At that time, it corresponds to the high-risk, rare scenario in the upper right quadrant. The quadrant determination logic mentioned above is consistent with the data division and expert training strategy described above, thereby achieving "online routing based on the risk allocation during training" during the deployment phase.

[0118] 3) Multi-capacity expert selection and reasoning execution Based on the quadrant to which the current state belongs, the lightweight expert routing module of this invention adopts the following expert selection strategy: When the state falls in the lower left or upper left quadrant, the small expert model Expert S is activated, outputting the future reference trajectory and target velocity to cover most low-risk daily scenarios and scenarios with rare structures but simple environments; when the state falls in the lower right quadrant, the medium expert model Expert M is activated, utilizing its stronger temporal modeling and representation capabilities to handle highly stochastic interaction scenarios; when the state falls in the upper right quadrant, the large expert model Expert L is activated, and the conservative strategy learned under the strong supervision of the mechanistic motion planner is used to deal with high-risk rare scenarios with high aleatoric and high epistemic characteristics.

[0119] In one embodiment, to further reduce the average computational overhead, the present invention may also introduce computing power status variables (such as GPU utilization, NPU temperature and remaining time budget) as auxiliary inputs to dynamically fine-tune the above routing thresholds. For example, when computing power is tight, the upper right quadrant boundary may be tightened appropriately, and Expert M may be selected to replace Expert L, thereby ensuring both safety and real-time performance.

[0120] To ensure system security under extremely high uncertainty or abnormal computing power conditions, this invention introduces a rule-based expert model, Expert, outside of the multi-expert framework. As a fallback control module.

[0121] The rule expert model, Expert R, does not rely on deep neural network inference. Instead, it employs a classic driving rule model that has been theoretically verified and calibrated in engineering. Under extreme high-risk or computationally unsound conditions, it provides deterministic and verifiable safety control behavior for vehicles. The rule expert model consists of three parts: a longitudinal safety control submodule, a lateral stability control submodule, and multi-level safety degradation decision logic. Its specific contents are as follows.

[0122] (1) Vertical safety control submodule The longitudinal safety control submodule achieves safe following distance maintenance and emergency braking control based on the Intelligent Driver Model (IDM) and Time-to-Collision (TTC) metrics. The IDM model calculates the expected longitudinal acceleration based on the vehicle's current speed v, the net distance Δd to the vehicle in front, and the speed difference Δv between the two vehicles. The formula for the expected acceleration of the IDM model is:

[0123] in For maximum comfort acceleration, To determine the desired speed, the current speed limit or the target speed for the reduced speed zone is taken. For acceleration index, The desired following distance is defined as:

[0124] in Minimum parking distance, For safe headway, For comfortable braking deceleration, the parameters of the IDM model are pre-calibrated based on vehicle type, road speed limit, and current risk level. Simultaneously, the longitudinal safety control submodule calculates the collision time in real time.

[0125] in To prevent division by zero of extremely small positive numbers. When TTC is below the emergency braking threshold. At this time, the system triggers maximum braking force output. When the TTC is at the emergency braking threshold With warning threshold During this period, the system calculates the braking force using linear interpolation:

[0126] The final acceleration command for longitudinal safety control is the smaller value between the IDM output and the TTC braking command, i.e. This ensures that the hard constraints for collision avoidance are prioritized under all circumstances.

[0127] (2) Lateral stability control submodule The lateral stability control submodule employs a lane-keeping strategy based on proportional-derivative (PD) control. This strategy uses the lateral deviation of the vehicle's centerline relative to the current lane centerline as the metric. and the rate of change of deviation As input, calculate the desired steering angle:

[0128] in and These are the proportional gain and derivative gain, pre-calibrated based on the vehicle's lateral dynamics and steering system response characteristics. To ensure safety, the steering angle is expected to be limited to the physically feasible maximum steering angle range, i.e. Meanwhile, the rate of change of steering angle is constrained to Within a certain range, smooth steering is ensured. When lane line detection is unavailable or the confidence level is below a preset threshold, the lateral control submodule switches to steering angle hold mode, which maintains the current steering angle or slowly returns it to center, avoiding sudden deflection due to sensing failure.

[0129] (3) Multi-level security degradation decision logic The rule expert model implements a three-level safety degradation strategy based on the uncertainty level and current driving status: Level 1 is low-speed cruise control. When the total uncertainty first exceeds the upper limit threshold, the rule expert reduces the target speed to a preset low-speed cruise speed while maintaining the current lane to reduce kinetic energy and reaction requirements. Longitudinal control uses the IDM model to track the low-speed target, while lateral control maintains lane keeping. Level 2 is deceleration and obstacle avoidance. When uncertainty continues to increase or TTC drops to the warning level, the rule expert further reduces the target speed to an extremely low speed and increases the safety margin with obstacles in front and to the sides. The safe headway in the IDM model... The distance is dynamically increased to proactively create a safe distance. The third level is safe parking, which occurs when the total uncertainty continuously exceeds the upper limit threshold and exceeds a preset time window. Or TTC is below the emergency braking threshold The rule expert outputs the target deceleration. Based on the braking command, guide the vehicle to decelerate smoothly to a complete stop within the current lane; during the stopping process, continuously monitor the situation of vehicles behind, and if there is a risk of rear-end collision, appropriately reduce the deceleration to avoid secondary accidents.

[0130] The switching between the three strategies mentioned above adopts a hysteresis mechanism, that is, the trigger threshold for downgrading from a low level to a high level is strictly higher than the recovery threshold for restoring from a high level to a low level, so as to avoid frequent switching near the uncertainty threshold, which would lead to unstable vehicle behavior.

[0131] The triggering and workflow of the rule expert model specifically include: 1) Rule-based fallback triggered by high uncertainty When total uncertainty Continuously exceeding the preset upper limit threshold A certain time window, or When the cognitive uncertainty threshold is exceeded on its own, the system determines that the current scene has significantly exceeded the training distribution of the multi-expert model. At this point, no learning expert is activated, and the rule expert takes over. Takeover generates safety control commands such as safe stopping, deceleration and avoidance, or low-speed maintenance, enabling a safe degradation from a learning-based strategy to a verifiable rule-based strategy.

[0132] 2) Timeout and resource anomaly monitoring This invention sets a maximum allowable inference latency threshold in the end-to-end decision planning module. The system monitors the total computation time from input sensor data to output trajectory / control commands in each control cycle. When the single-frame inference time exceeds the threshold, or when the inference engine returns an error code or there are resource anomalies such as insufficient video memory, the system immediately interrupts the expert inference of the current cycle and switches to a rule-based expert model or a traditional longitudinal / lateral control strategy to ensure that the vehicle can maintain a controllable and safe motion state even when computing power degrades.

[0133] 3) Multi-level collaboration ensures efficient vehicle deployment. Through the aforementioned lightweight uncertainty assessment, four-quadrant risk determination, multi-capacity expert routing, and rule-based fallback, this invention achieves dynamic matching of model capacity based on scenarios at the algorithm level. Large experts with high computational consumption are only activated in high-risk, rare scenarios, while decision-making and planning are mostly completed by small or medium-sized experts. This significantly reduces the average inference latency and power consumption on the vehicle-mounted edge computing platform while maintaining multi-scenario risk perception and high-precision trajectory planning capabilities.

[0134] To implement the methods of the above embodiments, the present invention provides an adaptive end-to-end decision planning system for vehicle-mounted edge computing platforms, comprising: The data acquisition module is used to acquire the vehicle's driving parameters, navigation instructions, and the latest image frame sequence captured by the forward-facing camera, and to extract state feature vectors. The uncertainty estimation module is used to calculate environmental uncertainty based on the state feature vector using a pre-trained SQR quantile regression network model, and to calculate model uncertainty by integrating multiple SQR quantile regression networks, and then sum them to obtain the total uncertainty. The environmental uncertainty, model uncertainty, and total uncertainty are compared with the classification threshold obtained from uncertainty distribution statistics to determine the risk quadrant to which the current vehicle state belongs. The lightweight expert routing module is used to select and activate offline-trained small expert models, medium expert models, or large expert models to output future reference trajectories and target velocities based on the risk quadrant to which the current state belongs. Among them, small expert models cover most low-risk daily scenarios and scenarios with rare structures but simple environments; medium expert models handle highly random interaction scenarios; and large expert models are used to deal with high-risk and rare scenarios with high environmental uncertainty and high model uncertainty.

[0135] Furthermore, the lightweight expert routing module is also used to take over when the total uncertainty continues to exceed the preset upper limit threshold for a certain time window, or when the environmental uncertainty alone exceeds the highest uncertainty threshold. In this case, the rule expert model takes over and generates safety control instructions such as safe stopping, deceleration and avoidance, or low-speed maintenance, thus safely downgrading from a learning strategy to a verifiable rule strategy.

[0136] Each module is mainly used to implement the various steps of the method implementation, which will not be elaborated here.

[0137] This application also provides a computer-readable storage medium, such as flash memory, hard disk, multimedia card, card-type memory (e.g., SD or DX memory), random access memory (RAM), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, disk, optical disk, server, app store, etc., which stores a computer program. When the program is executed by a processor, it implements the corresponding function. The computer-readable storage medium of this embodiment, when executed by a processor, implements the adaptive end-to-end decision planning method for an in-vehicle edge computing platform, as described in this embodiment.

[0138] This invention systematically solves the core problems of existing end-to-end autonomous driving methods, such as insufficient unified modeling in multiple scenarios, inadequate generalization ability, and limited on-board computing power, through a technical framework of scenario uncertainty assessment, multi-expert collaborative planning, and vehicle-mounted lightweight deployment. The technical effects are reflected in the following three aspects.

[0139] (1) Achieve accurate identification and classification of driving risks in multiple scenarios, and improve the generalization ability and decision stability of the strategy. The "multi-SQR deep integrated uncertainty assessment model" constructed in this invention can output both aleatoric and epistemic uncertainties, and realize the quantitative assessment of environmental randomness and model cognitive blind spots. Based on the four-quadrant scenario division of this uncertainty label, the training data can be automatically classified according to risk characteristics, and the high uncertainty samples can be corrected by combining the mechanistic motion planner, which fundamentally reduces the misjudgment and unstable behavior of traditional end-to-end methods in long-tail, rare and high-risk scenarios, thereby significantly enhancing the generalization ability of the strategy in cross-road conditions and cross-regional scenarios.

[0140] (2) Constructing a multi-capacity expert system that can be invoked on demand, balancing performance and real-time performance under the condition of limited on-board computing power. Based on scenario uncertainty, this invention realizes a multi-expert collaborative mechanism in which "small expert models handle daily low-risk scenarios, medium-sized models handle highly random interactive scenarios, and large models handle high-risk rare scenarios". Since most states during vehicle operation are in the low-risk quadrant, this invention can activate only small or medium-sized models for most of the time, thereby significantly reducing the average inference latency and computing power consumption. In truly complex or dangerous scenarios, large experts can be activated in a timely manner to ensure the safety and accuracy of decision-making. This design of allocating model capacity on demand enables the end-to-end planning system to still balance high performance and high real-time performance on resource-constrained edge platforms.

[0141] (3) Introducing uncertainty distillation and rule-based fallback to achieve a verifiable safety assurance mechanism. This invention distills the uncertainty assessment capability of multi-model integration into a lightweight prediction head embedded in the online inference process, enabling the vehicle-mounted terminal to have continuous risk perception capabilities while maintaining extremely low computational overhead. When real-time uncertainty exceeds the threshold or the inference cycle is abnormal, the system will automatically switch to rule experts or traditional stable control strategies, achieving a safe degradation from "learning-based decision-making" to "verifiable rule-based strategies." This mechanism effectively avoids uncontrollable decision-making behavior that deep models may produce in off-distribution scenarios, significantly improving the overall system's safety robustness.

[0142] In summary, this invention has achieved significant technological improvements in key dimensions such as accurate risk identification, cross-scenario strategy stability, in-vehicle real-time performance and computing resource utilization, and safety assurance under extreme conditions. It provides a complete solution with high safety, high reliability, and high efficiency for the implementation of end-to-end autonomous driving decision planning on actual in-vehicle edge platforms.

[0143] It should be noted that, depending on the implementation needs, the various steps / components described in this application can be broken down into more steps / components, or two or more steps / components or parts of the operation of steps / components can be combined into new steps / components to achieve the purpose of this invention.

[0144] The order of the steps in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0145] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. An adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms, characterized in that, Includes the following steps: The system acquires the vehicle's driving parameters, navigation commands, and the latest image frame sequence captured by the forward-facing camera, and extracts state feature vectors. Based on the state feature vector, the environmental uncertainty is calculated using a pre-trained SQR quantile regression network model, and the model uncertainty is calculated by integrating multiple SQR quantile regression networks. The total uncertainty is obtained by summing the results. By comparing environmental uncertainty, model uncertainty, total uncertainty, and classification thresholds obtained from uncertainty distribution statistics, the risk quadrant to which the current vehicle status belongs is determined. Based on the risk quadrant of the current state, select to activate the offline-trained small expert model, medium expert model, or large expert model to output the future reference trajectory and target velocity; among them, the small expert model is used for low-risk scenarios; Medium-level expert models handle highly stochastic interaction scenarios; large-scale expert models are used for high-risk scenarios.

2. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, When both environmental uncertainty and model uncertainty are less than the corresponding threshold, the scenario corresponds to the low-risk scenario in the lower left quadrant; when environmental uncertainty is less than the corresponding threshold and model uncertainty is greater than or equal to the corresponding threshold, the scenario corresponds to the structural blind spot scenario in the upper left quadrant; when environmental uncertainty is greater than or equal to the corresponding threshold and model uncertainty is less than the corresponding threshold, the scenario corresponds to the high-randomness interaction scenario in the lower right quadrant; when both environmental uncertainty and model uncertainty are greater than the corresponding threshold, the scenario corresponds to the high-risk scenario in the upper right quadrant.

3. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, When the total uncertainty continues to exceed the preset upper limit threshold for a certain time window, or when the environmental uncertainty alone exceeds the highest uncertainty threshold, the rule expert model takes over and generates a safety control command to stop safely, decelerate to avoid obstacles, or maintain a low speed.

4. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, The training process of the SQR quantile regression network model is as follows: a coach model based on PPO proximal policy optimization is constructed; vehicle trajectories are sampled in the simulation environment using the coach model under multiple map, weather, and traffic density configurations; all collected vehicle trajectories are replayed and the expected labels are calculated to obtain a dataset with expected labels; the pre-constructed SQR quantile regression network model is trained based on the dataset with expected labels.

5. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, The multi-SQR quantile regression network ensemble is specifically as follows: multiple independent SQR networks are trained using data resampling and random initialization to form a deep ensemble; when evaluating any state, multiple preset quantiles are calculated for each SQR network, and the total uncertainty is decomposed into environmental uncertainty and model uncertainty through cross-model and intra-model variance decomposition.

6. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 5, characterized in that, By using a multi-SQR quantile regression network ensemble, each state in the vehicle trajectory collected by the coach model is labeled with environmental uncertainty and model uncertainty. The state space is divided into four quadrants with environmental uncertainty as the x-axis and model uncertainty as the y-axis. For the data in different quadrants, three different sizes of student models and a high-precision mechanistic motion planner are selected for data correction and training. Under the constraint of limited on-board computing power, adaptive modeling of scenarios with different difficulty and risk levels is achieved, resulting in small expert models, medium expert models and large expert models.

7. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, When trajectory prediction is performed using a small, medium, or large expert model, if the prediction time exceeds the inference delay threshold or an abnormal signal is detected, the current expert model prediction is interrupted and switched to other control strategies to ensure that the vehicle maintains a controllable and safe motion state when computing power degrades.

8. The adaptive end-to-end decision-making and planning method for vehicle-mounted edge computing platforms according to claim 1, characterized in that, At the algorithm implementation level, the capacity of the three expert models is dynamically matched according to the scenario.

9. An adaptive end-to-end decision-making and planning system for vehicle-mounted edge computing platforms, characterized in that, include: The data acquisition module is used to acquire the vehicle's driving parameters, navigation instructions, and the latest image frame sequence captured by the forward-facing camera, and to extract state feature vectors. The uncertainty estimation module is used to calculate environmental uncertainty based on the state feature vector using a pre-trained SQR quantile regression network model, and to calculate model uncertainty by integrating multiple SQR quantile regression networks, and then sum them to obtain the total uncertainty. The environmental uncertainty, model uncertainty, and total uncertainty are compared with the classification threshold obtained from uncertainty distribution statistics to determine the risk quadrant to which the current vehicle state belongs. The lightweight expert routing module is used to select and activate a pre-trained offline small expert model, medium expert model, or large expert model to output the future reference trajectory and target velocity based on the risk quadrant to which the current state belongs; among which, the small expert model is used for low-risk scenarios; Medium-level expert models handle highly stochastic interaction scenarios; large-scale expert models are used for high-risk scenarios.

10. The adaptive end-to-end decision planning system for vehicle-mounted edge computing platforms according to claim 9, characterized in that, The lightweight expert routing module is also used to take over when the total uncertainty continues to exceed the preset upper limit threshold for a certain time window, or when the environmental uncertainty alone exceeds the highest uncertainty threshold. In this case, the rule expert model takes over and generates safety control commands such as safe stopping, deceleration and avoidance, or low-speed maintenance.

Images