Autonomous driving decision method, device and vehicle
Patent Information
- Application Number
- CN202611434874.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-16
- Publication Date
- 2026-10-20
AI Technical Summary
然而，在长尾场景和强交互场景下，该方法无法输出可靠驾驶决策
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This application relates to the field of vehicle technology, and in particular to an autonomous driving decision-making method, device, and vehicle. Background Technology
[0002] Autonomous driving technology is at the forefront of artificial intelligence and automotive engineering, and one of its core challenges lies in intelligent decision-making in complex traffic environments. Autonomous driving decision-making systems primarily rely on real-time vehicle driving status information and surrounding traffic conditions to intelligently output driving strategies and control vehicle actions, directly determining the driving safety, stability, and comfort of autonomous vehicles.
[0003] Currently, autonomous driving decision-making methods typically evaluate candidate driving behaviors based on preset rules, learning models, or trajectory prediction results to select and output driving strategies. However, in long-tail scenarios and highly interactive scenarios, this method cannot output reliable driving decisions.
[0004] There is currently no effective solution to the problem that related technologies cannot output reliable driving decisions. Summary of the Invention
[0005] This embodiment provides an autonomous driving decision-making method, device, and vehicle, which are designed to output reliable driving decisions.
[0006] Firstly, this embodiment provides an autonomous driving decision-making method, which is applied to a vehicle; the autonomous driving decision-making method includes:
[0007] Acquire driving observation data of the vehicle and determine the preset structural causal model corresponding to the driving observation data;
[0008] Based on the preset structural causal model and the driving observation data, determine the posterior distribution of the exogenous variables in the preset structural causal model;
[0009] Multiple candidate driving strategies are determined based on the driving observation data, and for each candidate driving strategy, the counterfactual result corresponding to the candidate driving strategy is determined based on the preset structural causal model, the candidate driving strategy, and the posterior distribution.
[0010] Based on the counterfactual results corresponding to each of the candidate driving strategies, the target driving strategy for the vehicle is determined, and the vehicle is controlled according to the target driving strategy.
[0011] Secondly, this embodiment provides an autonomous driving decision-making device, which is applied to a vehicle; the autonomous driving decision-making device includes:
[0012] The retrieval module is used to acquire the driving observation data of the vehicle and determine the preset structural causal model corresponding to the driving observation data;
[0013] The processing module is used to determine the posterior distribution of the exogenous variables in the preset structural causal model based on the preset structural causal model and the driving observation data.
[0014] The analysis module is used to determine multiple candidate driving strategies based on the driving observation data, and for each candidate driving strategy, determine the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy and the posterior distribution;
[0015] The decision module is used to determine the target driving strategy of the vehicle based on the counterfactual results corresponding to each of the candidate driving strategies, and to control the vehicle to perform corresponding driving actions according to the target driving strategy.
[0016] Thirdly, this embodiment provides a vehicle including a control device, the control device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the autonomous driving decision-making method described in the first aspect above.
[0017] Compared with related technologies, this embodiment provides an autonomous driving decision-making method, device, and vehicle. The autonomous driving decision-making method includes: acquiring driving observation data of the vehicle and determining a preset structural causal model corresponding to the driving observation data; determining the posterior distribution of exogenous variables in the preset structural causal model based on the preset structural causal model and the driving observation data; determining multiple candidate driving strategies based on the driving observation data, and for each candidate driving strategy, determining the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy, and the posterior distribution; determining the target driving strategy of the vehicle based on the counterfactual results corresponding to each candidate driving strategy, and controlling the vehicle according to the target driving strategy.
[0018] In this way, by introducing a pre-defined structural causal model, the posterior distribution of exogenous variables in the model is inferred using driving observation data. Counterfactual analysis is then performed on each candidate driving strategy based on this posterior distribution, yielding counterfactual results for different candidate strategies. Driving strategy selection is then completed based on these counterfactual results, thus fully considering the impact of unobservable uncertainties on the driving process and improving the accuracy of risk prediction for each candidate driving behavior. Simultaneously, the counterfactual analysis of each candidate strategy is based on the same posterior distribution, ensuring that comparisons between different strategies are not affected by differences in exogenous random factors. This ensures that the final output target driving strategy has reliable safety guarantees and interpretable causal basis, thereby enabling the output of reliable driving decisions.
[0019] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. Attached Figure Description
[0020] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:
[0021] Figure 1 This is a flowchart of an embodiment of the autonomous driving decision-making method provided in this application;
[0022] Figure 2 This is a flowchart of Embodiment 2 of the autonomous driving decision-making method provided in this application;
[0023] Figure 3 This is a flowchart of Embodiment 3 of the autonomous driving decision-making method provided in this application;
[0024] Figure 4 This is a flowchart of Embodiment 4 of the autonomous driving decision-making method provided in this application;
[0025] Figure 5 This is a flowchart of Embodiment 5 of the autonomous driving decision-making method provided in this application;
[0026] Figure 6 This is a flowchart of Embodiment Six of the Autonomous Driving Decision-Making Method provided in this application;
[0027] Figure 7 This is a flowchart of Embodiment Seven of the Autonomous Driving Decision-Making Method provided in this application;
[0028] Figure 8 This is a flowchart of Embodiment 8 of the autonomous driving decision-making method provided in this application;
[0029] Figure 9This is a structural block diagram of the autonomous driving decision-making device provided in this application. Detailed Implementation
[0030] To better understand the purpose, technical solution, and advantages of this application, the application is described and illustrated below in conjunction with the accompanying drawings and embodiments.
[0031] Unless otherwise defined, the technical or scientific terms used in this application shall have the general meaning understood by one of ordinary skill in the art to which this application pertains. Words such as “a,” “an,” “an,” “the,” “the,” and “these” used in this application do not indicate quantitative limitation and may be singular or plural. The terms “comprising,” “including,” “having,” and any variations thereof used in this application are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or device that comprises a series of steps or modules (units) is not limited to the listed steps or modules (units) but may include steps or modules (units) not listed, or may include other steps or modules (units) inherent to these processes, methods, products, or devices. Words such as “connected,” “linked,” and “coupled” used in this application are not limited to physical or mechanical connections but may include electrical connections, whether direct or indirect. “Multiple” used in this application refers to two or more. “And / or” describes the relationship between related objects, indicating that three relationships may exist; for example, “A and / or B” can represent: A alone, A and B simultaneously, and B alone. Normally, the character " / " indicates that the objects before and after it are in an "or" relationship. The terms "first," "second," "third," etc., used in this application are merely to distinguish similar objects and do not represent a specific order of objects.
[0032] The following specific embodiments are given to illustrate the technical solution of this application in detail.
[0033] Figure 1 This is a flowchart of an embodiment of the autonomous driving decision-making method provided in this application, which is applied to a vehicle. Please refer to... Figure 1 The method provided in this embodiment may include the following steps:
[0034] Step S101: Obtain the vehicle's driving observation data and determine the preset structural causal model corresponding to the driving observation data.
[0035] The vehicle's driving observation data includes, but is not limited to, the vehicle's motion status (e.g., vehicle position, vehicle speed, vehicle acceleration, vehicle heading angle, etc.), the status of surrounding traffic participants (e.g., the position, speed, and other types of information of other vehicles, pedestrians, two-wheeled vehicles, etc.), road and environmental information (e.g., lane geometry, road boundaries, drivable areas, traffic signs, signal light status, etc.), the current task objective, and sensor quality.
[0036] After acquiring driving observation data, a pre-defined structural causal model (SCM) that matches the driving observation data is retrieved. In specific implementations, the pre-defined structural causal model can also be determined by combining factors such as the current event type (e.g., going straight, passing through an intersection, merging / exiting), vehicle type, and the road rules of the road where the vehicle is currently located (e.g., driving on the right). This ensures that the invoked structural causal model can accurately represent the causal mechanism in the current driving scenario.
[0037] The pre-defined structural causal model is a structured expression of the causal relationships between variables in the current traffic scenario, specifically defined as M=(U,V,F,P(U)). Here, U represents the set of exogenous variables, which are uncertainties in the traffic scenario that cannot be directly observed but significantly affect decision-making. These include, but are not limited to, the behavioral intentions of surrounding traffic participants, road surface disturbances, perception system biases, and random environmental disturbances. V represents the set of endogenous variables, which are system state variables that can be directly or indirectly obtained through observation. These include, but are not limited to, the vehicle's motion state, the state of surrounding traffic participants, and road and environmental information. F represents the set of structural equations, F={f k}, each endogenous variable V k Determined by the corresponding structural equation, i.e., V k =f k (Pa k U k ), Pa k For V k The causal parent node is used to characterize the causal generation mechanism and dependency relationship between variables; P(U) represents the prior distribution corresponding to the exogenous variable, which is used to characterize the initial understanding of exogenous uncertainty before the current observation data is obtained.
[0038] It should be noted that structural causal models can be pre-built using expert knowledge or causal discovery algorithms and stored in the form of configurable model versions for the decision-making process to call on demand in different scenarios.
[0039] Step S102: Based on the preset structural causal model and driving observation data, determine the posterior distribution of the exogenous variables in the preset structural causal model.
[0040] In this step, the currently acquired driving observation data is used as deterministic observational evidence of endogenous variables in the pre-defined structural causal model. Combined with the causal dependencies defined by the structural equations in the pre-defined structural causal model and the prior distributions of the exogenous variables, the posterior distribution of the exogenous variables under the given observational evidence is calculated.
[0041] In practice, the posterior distribution of the exogenous variables can be calculated using algorithms such as Bayesian filtering, particle filtering, variational inference, and Markov chain-Monte Carlo, based on the structural equation and the prior distribution of the exogenous variables.
[0042] Step S103: Based on driving observation data, determine multiple candidate driving strategies, and for each candidate driving strategy, determine the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy and the posterior distribution.
[0043] In this embodiment, multiple candidate behaviors for the vehicle (e.g., lane keeping, changing lanes left / right, following other vehicles, slowing down to yield, etc.) can be generated based on driving observation data. Subsequently, the vehicle's current driving environment data is acquired, including but not limited to road topology information, lane line information, drivable area range, traffic signals, and traffic signs. Environmental constraint information is determined based on the current driving environment data, including but not limited to the set of drivable lanes and hard rules (e.g., stopping at red lights, slowing down to yield at yield signs, etc.).
[0044] For each candidate behavior, the candidate behavior is aggregated with environmental constraint information to form a behavior envelope corresponding to the candidate behavior. The trajectory planner or behavior planner samples or optimizes within each behavior envelope to generate at least one specific behavior instance that satisfies all the constraints in the behavior envelope. This specific behavior instance is the candidate driving strategy.
[0045] Furthermore, counterfactual reasoning is performed for each candidate driving strategy to determine its corresponding counterfactual outcome. The counterfactual outcome characterizes the future evolution of a vehicle under given external uncertainties, assuming a particular candidate driving strategy. Specifically, the posterior distribution is sampled to obtain multiple exogenous variable samples. The vehicle's future trajectory is determined based on the candidate driving strategy. For each exogenous variable sample, the predicted response behavior of surrounding traffic participants is determined based on the exogenous variable sample, the future trajectory, and the structural equations corresponding to surrounding traffic participants in the pre-defined structural causal model. Subsequently, a correlation analysis is performed on the future trajectory and the predicted response behavior. Based on the correlation analysis results, multiple evaluation indicators corresponding to the candidate driving strategy are determined. These indicators include, but are not limited to, relative distance and time-to-collision (TTC). The multiple evaluation indicators are summarized to obtain the sub-counterfactual outcomes corresponding to the exogenous variable samples. Finally, the counterfactual outcome is determined based on each sub-counterfactual outcome. In this way, all candidate driving strategies reuse the same set of exogenous samples from the same source, thus ensuring that the differences in counterfactual results between different strategies come only from the candidate behavior itself, rather than from the interference of exogenous random factors, making the counterfactual assessment fair and interpretable.
[0046] Step S104: Determine the target driving strategy for the vehicle based on the counterfactual results corresponding to each candidate driving strategy, and control the vehicle according to the target driving strategy.
[0047] Specifically, a preset baseline driving strategy for the vehicle can be obtained, and the corresponding risk assessment value can be determined. For each candidate driving strategy, a risk assessment is performed based on the counterfactual results corresponding to the candidate driving strategy, resulting in a risk assessment value. The difference between the risk assessment value corresponding to the preset baseline driving strategy and the risk assessment value corresponding to the candidate driving strategy is then determined as the risk improvement amount corresponding to the candidate driving strategy. Subsequently, based on the risk improvement amounts corresponding to each candidate driving strategy, a target driving strategy is selected from all candidate driving strategies.
[0048] It should be noted that if the risk improvement corresponding to a candidate driving strategy is greater than zero, it indicates that the candidate driving strategy can reduce risk compared to the preset benchmark driving strategy, i.e., it has better driving performance. When selecting a target driving strategy, the candidate driving strategy with the largest risk improvement can be selected from all candidate driving strategies with a risk improvement greater than zero. Furthermore, if no candidate driving strategy has a risk improvement greater than zero, it indicates that all candidate driving strategies are not superior to the preset benchmark driving strategy, and the preset benchmark driving strategy can be used as the target driving strategy. After determining the target driving strategy, it is converted into control commands for the vehicle's underlying actuators (e.g., chassis control commands such as throttle opening, brake pressure, steering angle, and yaw control parameters) to control the vehicle to perform corresponding driving actions.
[0049] The autonomous driving decision-making method provided in this embodiment specifically includes: acquiring driving observation data of the vehicle and determining a preset structural causal model corresponding to the driving observation data; determining the posterior distribution corresponding to the exogenous variables in the preset structural causal model based on the preset structural causal model and the driving observation data; determining multiple candidate driving strategies based on the driving observation data, and for each candidate driving strategy, determining the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy, and the posterior distribution; determining the target driving strategy of the vehicle based on the counterfactual results corresponding to each candidate driving strategy, and controlling the vehicle according to the target driving strategy.
[0050] In this way, by introducing a pre-defined structural causal model, the posterior distribution of exogenous variables in the model is inferred using driving observation data. Counterfactual analysis is then performed on each candidate driving strategy based on this posterior distribution, yielding counterfactual results for different candidate strategies. Driving strategy selection is then completed based on these counterfactual results, thus fully considering the impact of unobservable uncertainties on the driving process and improving the accuracy of risk prediction for each candidate driving behavior. Simultaneously, the counterfactual analysis of each candidate strategy is based on the same posterior distribution, ensuring that comparisons between different strategies are not affected by differences in exogenous random factors. This ensures that the final output target driving strategy has reliable safety guarantees and interpretable causal basis, thereby enabling the output of reliable driving decisions.
[0051] Figure 2 This is a flowchart of Embodiment 2 of the autonomous driving decision-making method provided in this application. Please refer to... Figure 2 Based on the above embodiments, according to the preset structural causal model and driving observation data, the posterior distribution corresponding to the exogenous variables in the preset structural causal model is determined, including:
[0052] Step S201: Obtain each structural equation in the preset structural causal model, and infer the driving state based on each structural equation to obtain driving prediction data.
[0053] Obtain the structural equations f in the pre-defined structural causal model k Structural equation modeling is used to characterize endogenous variable V. k Its causal parent node Pa k and exogenous variable U k The functional mapping relationship between them, i.e., V k =f k (Pa k U k Based on the structural equations, the driving state is inferred, and the predicted values of each endogenous variable under the current assumptions of the exogenous variables are calculated to form driving prediction data.
[0054] Step S202: Determine the likelihood value corresponding to the driving observation data based on the driving observation data and the driving prediction data.
[0055] driving observation data (O 0:t The likelihood value of the driving observation data is determined by comparing it with the driving prediction data, under the given assumptions of exogenous variables. For example, based on a pre-defined observation noise model (such as a Gaussian noise model, a Laplace noise model, etc.), the conditional probability density of the driving observation data relative to the driving prediction data is calculated, and this conditional probability density is the likelihood value.
[0056] Step S203: Obtain the prior distribution corresponding to the exogenous variables in the preset structural causal model, and determine the posterior distribution based on the likelihood value and the prior distribution.
[0057] In this step, the prior distribution P(U) corresponding to the exogenous variables in the pre-defined causal model is obtained, and then the posterior distribution q(U) is calculated by combining the likelihood value and the prior distribution. t |O 0:t ).
[0058] The method provided in this embodiment extrapolates the driving state by obtaining the various structural equations in the structural causal model, then determines the likelihood value by comparing the driving observation data with the predicted data, and finally obtains the posterior distribution of the exogenous variable under given observational evidence by Bayesian fusion of the likelihood value and the prior distribution, thereby realizing the quantitative perception of external unobservable uncertainties.
[0059] Figure 3 This is a flowchart of Embodiment 3 of the autonomous driving decision-making method provided in this application. Please refer to... Figure 3 Based on the above embodiments, multiple candidate driving strategies are determined according to driving observation data, including:
[0060] Step S301: Based on driving observation data, generate multiple candidate behaviors for the vehicle.
[0061] In practice, driving observation data can be input into a pre-trained driving behavior prediction model, which will then output multiple candidate behaviors for the vehicle. It should be noted that these candidate behaviors represent the vehicle's driving intentions, and include, but are not limited to, lane keeping, changing lanes to the left, changing lanes to the right, following other vehicles, slowing down to yield, and pulling over to the side of the road.
[0062] Step S302: Obtain the vehicle's current driving environment data and determine environmental constraint information based on the current driving environment data.
[0063] Current driving environment data includes, but is not limited to, road topology information (e.g., number of lanes, lane connectivity, lane width, etc.), lane line information (e.g., solid / dashed line type, etc.), drivable area range (e.g., road boundaries, curbs, construction areas, areas occupied by obstacles, etc.), traffic signals (e.g., traffic light colors and remaining time, etc.), and traffic signs (e.g., speed limit signs, yield signs, no-entry signs, etc.).
[0064] Analyzing the aforementioned current driving environment data yields environmental constraint information. This information characterizes the permitted and prohibited boundaries of driving behavior in the external environment, specifically including the set of passable lanes and mandatory rules. The set of passable lanes represents the range of lanes a vehicle can occupy in the current scenario; mandatory rules include traffic light rules, yield rules, and speed limits.
[0065] Step S303: Obtain the preset physical constraint information of the vehicle, and for each candidate behavior, determine the motion feasible domain corresponding to the candidate behavior based on the preset physical constraint information and the candidate behavior.
[0066] For each candidate behavior, the feasible motion domain corresponding to that candidate behavior is determined based on preset physical constraint information and the behavior type. The preset physical constraint information consists of inherent constraints of the vehicle, including but not limited to vehicle dynamic limits (e.g., maximum braking deceleration, maximum longitudinal acceleration, maximum lateral acceleration, minimum turning radius, etc.), actuator capability boundaries (e.g., maximum steering wheel angle, maximum braking pressure, maximum drive torque, actuator response delay, etc.), and ride comfort limits (e.g., maximum impact, maximum rate of change of lateral acceleration, etc.).
[0067] In practice, based on the different requirements of candidate behaviors on vehicle motion states, the specific constraint interval corresponding to the candidate behavior is extracted from the aforementioned preset physical constraint information to form the motion feasible region corresponding to the candidate behavior. The motion feasible region includes parameter constraint ranges such as the longitudinal acceleration feasible region, the lateral acceleration feasible region, and the trajectory curvature feasible region.
[0068] In this way, the vehicle's physical capabilities and driving comfort constraints are transformed into continuous boundary constraints of driving trajectory and motion parameters, ensuring that the subsequently generated candidate driving strategies are executable and avoiding the generation of invalid strategies that exceed the vehicle's physical limits.
[0069] Step S304: Generate candidate driving strategies based on candidate behaviors, environmental constraint information, and the motion feasible domain corresponding to the candidate behaviors.
[0070] In this step, candidate behaviors, environmental constraints, and the feasible motion domain are jointly constrained to generate a candidate driving strategy for each candidate behavior. Specifically, for each candidate behavior, the candidate behavior, environmental constraints, and the feasible motion domain are aggregated to form a behavior envelope corresponding to that candidate behavior. Subsequently, a trajectory planner or behavior planner samples or optimizes within each behavior envelope to generate at least one concrete behavior instance that satisfies all the constraints of that behavior envelope, serving as the candidate driving strategy for that candidate behavior.
[0071] The method provided in this embodiment generates multiple candidate behaviors based on driving observation data, then determines environmental constraint information based on driving environment data, incorporates external hard constraints such as road boundaries, traffic signals, and traffic rules into the strategy generation conditions, and determines the motion feasible domain based on preset physical restriction information. The vehicle dynamic limits, actuator capability boundaries, and comfort limitations are transformed into formal constraints of continuous parameter intervals. Finally, the candidate behaviors, environmental constraint information, and motion feasible domain are jointly constrained to form a behavior envelope, and specific behavior instances are generated within the envelope as candidate driving strategies. This ensures that the generated candidate strategies simultaneously meet the triple conditions of intention rationality, environmental compliance, and physical feasibility, thereby improving the reliability and safety of decision-making.
[0072] Figure 4 The flowchart for Embodiment 4 of the autonomous driving decision-making method provided in this application is shown below. Please refer to... Figure 4 Based on the above embodiments, according to the preset structural causal model, candidate driving strategies, and posterior distribution, the counterfactual results corresponding to the candidate driving strategies are determined, including:
[0073] Step S401: Sample the posterior distribution to obtain multiple exogenous variable samples.
[0074] Specifically, sampling is performed on the posterior distribution corresponding to the exogenous variable to obtain a set of homologous exogenous samples. This set includes multiple exogenous variable samples, each of which is... (s=1,2,……,K).
[0075] Step S402: Determine the future trajectory of the vehicle based on the candidate driving strategy, and for each exogenous variable sample, determine the predicted response behavior of the surrounding traffic participants based on the exogenous variable sample, the future trajectory, and the structural equations corresponding to the surrounding traffic participants in the preset structural causal model.
[0076] In this step, a causal intervention is performed. That is, in the pre-defined causal model, the vehicle's decision variable A is set as the current candidate driving strategy. The remaining structural equations in the model remain unchanged. Here, H represents the prediction time domain. To predict time variables within the prediction time domain, under this intervention condition, a vehicle kinematics or dynamics model is used to generate the vehicle's future trajectory within the prediction time domain based on the candidate driving strategy. Subsequently, for each exogenous variable sample, the structural equation model corresponding to the surrounding traffic participants is used to extrapolate the exogenous variable sample and the future trajectory, obtaining the predicted response behavior of the surrounding traffic participants, i.e., extrapolating the response trajectory of the surrounding traffic participants to the vehicle's behavior within the prediction time domain.
[0077] Step S403: Perform correlation analysis on the future motion trajectory and predicted response behavior, and determine multiple evaluation indicators corresponding to the candidate driving strategies based on the correlation analysis results.
[0078] A correlation analysis is performed on the future trajectory and predicted response behavior to analyze the spatiotemporal correlation characteristics between the vehicle and surrounding traffic participants at each time step within the prediction time domain, obtaining the correlation analysis results. Subsequently, based on the correlation analysis results, multiple evaluation indicators corresponding to candidate driving strategies are determined, including relative distance and collision time. Simultaneously, evaluation indicators for candidate driving strategies in dimensions such as road boundary preservation, traffic rule compliance, physical feasibility of the actuator, and ride comfort can also be obtained based on the analysis of the vehicle's future trajectory.
[0079] Step S404: Summarize multiple evaluation indicators to obtain the sub-counterfactual results corresponding to the exogenous variable samples.
[0080] Specifically, by summarizing the above evaluation indicators, the sub-counterfactual results corresponding to the exogenous variable sample are obtained. (s=1,2,……,K). Sub-counterfactual results are used to characterize the performance of the candidate driving strategy under specific exogenous assumptions.
[0081] Step S405: Determine the counterfactual result based on each sub-counterfactual result.
[0082] Based on the sub-counterfactual results corresponding to all exogenous variable samples, i.e. { }, determine the counterfactual outcome corresponding to the candidate driving strategy.
[0083] It should be noted that all candidate driving strategies are performed using the same set of homologous exogenous samples, which ensures the comparability of counterfactual results between different strategies.
[0084] The method provided in this embodiment introduces causal intervention and uses vehicle kinematics or dynamics models to generate the vehicle's future trajectory. Then, it uses structural equations corresponding to surrounding traffic participants to extrapolate exogenous variable samples and future trajectories, making the generation of surrounding response behavior dependent on both candidate strategy differences and external uncertainty assumptions. Based on this, multidimensional evaluation indicators are obtained through correlation analysis and summarized into sub-counterfactual results. These are then aggregated across samples to obtain counterfactual results, enabling the counterfactual results to simultaneously reflect the comprehensive performance of candidate driving strategies under multidimensional objectives. Furthermore, all candidate strategies are based on the same homogeneous exogenous sample set to execute the above extrapolation process, ensuring the comparability of counterfactual results between different strategies and effectively improving the safety and reliability of autonomous driving decision-making in complex interaction scenarios.
[0085] Figure 5 This is a flowchart of Embodiment 5 of the autonomous driving decision-making method provided in this application. Please refer to... Figure 5 Based on the above embodiments, the target driving strategy for the vehicle is determined according to the counterfactual results corresponding to each candidate driving strategy, including:
[0086] Step S501: Based on the counterfactual results corresponding to each candidate driving strategy, select alternative driving strategies from among the candidate driving strategies.
[0087] In this step, based on the counterfactual results corresponding to each candidate driving strategy, alternative driving strategies that meet the hard safety conditions are selected from the candidate driving strategies.
[0088] In practice, multiple preset hard constraint functions are obtained, each used to evaluate the margin of the candidate strategy in a certain safety dimension. The preset hard constraint functions include, but are not limited to, collision safety constraints (e.g., collision time not less than the safety threshold, minimum relative distance not less than the safety clearance, etc.), road boundary constraints (e.g., vehicles not exceeding the drivable area or lane boundary, etc.), traffic regulation constraints (e.g., not violating traffic light rules and lane marking rules, etc.), dynamic stability constraints (e.g., the required adhesion coefficient does not exceed the available adhesion coefficient of the road surface, etc.), and actuator capability constraints (e.g., the expected steering wheel angle, braking pressure, and driving torque do not exceed the available capability boundary of the actuator, etc.).
[0089] For each candidate driving strategy, the robustness margin of the candidate driving strategy is calculated based on the sub-counterfactual results under all corresponding exogenous variable samples. For example, for each exogenous variable sample, based on the sub-counterfactual results corresponding to the exogenous variable sample and each preset hard constraint function, the constraint value corresponding to each preset hard constraint function is determined, and the minimum value among all constraint values under the exogenous variable sample is taken as the sample margin corresponding to the exogenous variable sample. Subsequently, the sample margins corresponding to all exogenous variable samples are sorted, and the sample margin of the specified quantile in the sorting result is taken as the robustness margin of the candidate driving strategy.
[0090] Furthermore, candidate strategies with robustness margin less than a preset margin threshold (e.g., the preset margin threshold can be 0) are identified as unusable strategies and eliminated, while candidate strategies with robustness margin not less than the preset margin threshold are retained as alternative driving strategies.
[0091] Step S502: Obtain the vehicle's preset baseline driving strategy and determine the first risk assessment value corresponding to the preset baseline driving strategy.
[0092] Specifically, in the same homologous exogenous sample set The same counterfactual reasoning process as the candidate driving strategy is performed on the preset baseline driving strategy to obtain the sub-counterfactual results of the preset baseline driving strategy under each exogenous sample. Subsequently, multidimensional evaluation indicators are extracted from each sub-counterfactual result and fused according to preset weights to obtain the single-sample risk value under each exogenous variable sample. Then, the single-sample risk values under all exogenous variable samples are aggregated across samples using preset aggregation operators (e.g., expected value, quantile, maximum value, conditional risk value, etc.) to obtain the first risk assessment value, denoted as [missing value]. .
[0093] Step S503: For each alternative driving strategy, a risk assessment is performed on the alternative driving strategy based on the counterfactual results corresponding to the alternative driving strategy to obtain a second risk assessment value, and the difference between the first risk assessment value and the second risk assessment value is determined as the risk improvement amount corresponding to the alternative driving strategy.
[0094] For example, in one embodiment, for each sub-counterfactual result corresponding to the alternative driving strategy, a basic risk value corresponding to each evaluation indicator is determined based on multiple evaluation indicators in the sub-counterfactual result, and a single-sample risk value corresponding to the sub-counterfactual result is determined based on each basic risk value; and a second risk assessment value is determined based on the single-sample risk value corresponding to each sub-counterfactual result.
[0095] It should be noted that the counterfactual result of each alternative driving strategy includes its evaluation index under each exogenous variable sample. Multiple evaluation indices corresponding to each exogenous variable sample are fused according to preset weights to obtain a single-sample risk value. Then, using preset aggregation operators (e.g., expected value, quantile, maximum value, conditional risk value, etc.), cross-sample aggregation is performed on the single-sample risk values corresponding to each sub-counterfactual result to obtain the second risk assessment value corresponding to the alternative driving strategy, denoted as [missing value]. .
[0096] Subsequently, the difference between the first risk assessment value and the second risk assessment value is determined as the risk improvement amount corresponding to the alternative driving strategy, i.e., the risk improvement amount. Among them, the risk improvement corresponding to the alternative driving strategies is greater than zero (i.e., This indicates that the alternative driving strategy can reduce risk compared to the preset baseline driving strategy.
[0097] Step S504: Select the target driving strategy from the alternative driving strategies based on the risk improvement amount corresponding to each alternative driving strategy.
[0098] In practice, the target driving strategy can be selected from all alternative driving strategies with a risk improvement greater than zero. If no alternative driving strategy with a risk improvement greater than zero exists, the preset baseline driving strategy will be used as the target driving strategy.
[0099] The method provided in this embodiment selects alternative driving strategies that meet safety conditions through hard constraint robustness margin, and then calculates the risk improvement of each alternative driving strategy with reference to a preset benchmark driving strategy. Based on this, the driving strategy with the greatest risk reduction is selected as the target driving strategy, thereby ensuring that the final output target driving strategy obtains the maximum safety gain relative to the benchmark strategy while meeting hard safety constraints, effectively improving the safety and reliability of autonomous driving decisions in complex scenarios.
[0100] Figure 6 A flowchart of Embodiment Six of the Autonomous Driving Decision-Making Method Provided in this Application. Please refer to... Figure 6 Based on the above embodiments, alternative driving strategies are selected from each candidate driving strategy according to the counterfactual results corresponding to each candidate driving strategy, including:
[0101] Step S601: For each sub-counterfactual result corresponding to the candidate driving strategy, determine the constraint value corresponding to each preset hard constraint function based on the sub-counterfactual result and each preset hard constraint function, and take the minimum value among the constraint values as the reference margin corresponding to the sub-counterfactual result.
[0102] Specifically, for each sub-counterfactual result corresponding to the candidate driving strategy (s=1,2,……,K), based on the sub-counterfactual result and each preset hard constraint function (j=1,2,……,J), calculate the constraint value corresponding to each preset hard constraint function. The minimum value among all constraint values is used as the reference margin corresponding to the sub-counterfactual result.
[0103] Each preset hard constraint function is used to evaluate the margin of the candidate strategy in a certain safety dimension. The preset hard constraint functions include, but are not limited to, collision safety constraints (e.g., collision time not less than the safety threshold, minimum relative distance not less than the safety clearance, etc.), road boundary constraints (e.g., vehicles not exceeding the drivable area or lane boundary, etc.), traffic regulation constraints (e.g., not violating traffic light rules and lane marking rules, etc.), dynamic stability constraints (e.g., the required adhesion coefficient does not exceed the available adhesion coefficient of the road surface, etc.), and actuator capability constraints (e.g., the expected steering wheel angle, braking pressure, and driving torque do not exceed the available capability boundary of the actuator, etc.).
[0104] Step S602: Take the minimum value among the reference margins corresponding to each sub-counterfactual result as the robust margin corresponding to the candidate driving strategy.
[0105] In this step, the minimum value among the reference margins corresponding to each sub-counterfactual result is taken as the robustness margin corresponding to the candidate driving strategy, and the specific expression is as follows:
[0106] (1)
[0107] In equation (1), This represents the robustness margin corresponding to the i-th candidate driving strategy; This indicates the reference margin corresponding to the exogenous variable sample; This indicates a pre-defined hard constraint function; This represents the sub-counterfactual result corresponding to the exogenous variable sample.
[0108] Step S603: Candidate driving strategies with robustness margin less than a preset margin threshold are identified as unavailable strategies, and candidate driving strategies other than unavailable strategies are identified as alternative driving strategies.
[0109] Set the robustness margin to less than a preset margin threshold (for example, the preset margin threshold can be 0, i.e., ...). Candidate strategies that are not available are identified and eliminated, while candidate strategies with a robustness margin not less than a preset margin threshold are retained as alternative driving strategies.
[0110] The method provided in this embodiment calculates the constraint values of each preset hard constraint function for each sub-counterfactual result, selects the minimum value among all constraint values as the reference margin, and takes the minimum value among the reference margins corresponding to all exogenous variable samples as the robustness margin of the candidate strategy. This quantifies the safety boundary of the candidate driving strategy under various external uncertainties in the form of a conservative lower bound. Finally, strategies with robustness margins less than a preset threshold are eliminated, ensuring that all candidate strategies entering subsequent risk assessment and selection have passed the verification of all hard constraints under all considered exogenous uncertainties. This improves the reliability and credibility of the final output target strategy, thereby effectively enhancing the robustness of autonomous driving decisions in complex scenarios.
[0111] Figure 7 This is a flowchart of Embodiment Seven of the Autonomous Driving Decision-Making Method provided in this application. Please refer to... Figure 7 Based on the above embodiments, before controlling the vehicle according to the target driving strategy, the method further includes:
[0112] Step S701: Obtain the latest value of the specified features of the vehicle; the specified features include a combination of one or more of the following: driving event identifiers corresponding to driving observation data, posterior distributions corresponding to exogenous variables, driving observation data, robustness margins corresponding to the target driving strategy, and execution capability boundaries of the vehicle actuators.
[0113] In this embodiment, before controlling the vehicle according to the target driving strategy, the latest values of specified vehicle features are obtained in advance for pre-execution safety verification. These specified vehicle features include, but are not limited to, driving event identifiers corresponding to driving observation data (e.g., driving scenario identifiers and their associated version numbers), posterior distributions corresponding to exogenous variables, driving observation data, robustness margins corresponding to the target driving strategy, and the execution capability boundaries of the vehicle's actuators (e.g., maximum steering wheel angle, maximum braking pressure, maximum drive torque).
[0114] It should be noted that when a substantial change occurs in the driving scenario (e.g., a change in road topology, a change in mission objective, etc.), a new driving scenario identifier is assigned to the new scenario. Within the lifecycle of the scenario corresponding to the same driving scenario identifier, the version number monotonically increments with each update of valid observation data (e.g., a refresh of the vehicle's position, etc.) to accurately record the evolution of the data state under the same scenario.
[0115] Step S702: Verify the latest value of the specified feature, and control the vehicle according to the target driving strategy when the verification passes.
[0116] The latest value of the specified feature obtained at the current moment is compared one by one with the corresponding snapshot value frozen when the target driving strategy was authorized. Specifically, this includes: (1) verifying whether the current driving event identifier is consistent with the driving event identifier at the time of authorization. If they are consistent, the verification is deemed to pass; otherwise, the verification fails. (2) verifying whether the difference between the posterior distribution corresponding to the current exogenous variable and the posterior distribution at the time of authorization (which can be measured by KL divergence, JS distance or sample statistical difference index) is within the preset range. If the difference is within the preset range, the verification is deemed to pass; otherwise, the verification fails. (3) verifying whether the current driving observation data is still within the tolerance range of the counterfactual inference prediction at the time of authorization. If it is still within the tolerance range of the counterfactual inference prediction at the time of authorization, the verification is deemed to pass; otherwise, the verification fails. (4) recalculating the robustness margin of the target driving strategy based on the latest data and verifying whether it is still greater than or equal to the preset margin threshold. If it is still greater than or equal to the preset margin threshold, the verification is deemed to pass; otherwise, the verification fails. (5) Verify whether the control command corresponding to the target driving strategy falls completely within the execution capability boundary of the actuator. If it falls completely within the execution capability boundary of the actuator, the verification is considered successful; otherwise, the verification fails.
[0117] In addition, the expiration timestamps corresponding to the driving observation data (i.e., the effective expiration time of the driving observation data) can be obtained, and the expiration timestamps can be compared with the current system time to determine their timeliness, so as to verify whether the target driving strategy is effective at the current moment.
[0118] When all the above checks pass, the target driving strategy is confirmed to be effective and safe at the current moment. The target driving strategy is then translated into control commands for the vehicle's underlying actuators, controlling the vehicle to perform the corresponding driving actions. If any of the above checks fail, the current target driving strategy can be marked as invalid and revoked, triggering an emergency response mechanism. For example, a preset minimum-risk behavior (such as safe parking or emergency braking) can be activated to bring the vehicle into a safe state; or the entire decision-making process from observation acquisition, posterior inference to strategy generation can be re-executed to cope with new situations arising from changes in the scenario or updated cognition.
[0119] The method provided in this embodiment, by pre-obtaining the latest values of specified vehicle features and performing multi-dimensional comprehensive verification before issuing and executing the target driving strategy, can effectively identify safety risks introduced by factors such as scene switching and changes in external uncertainties within the time window from strategy authorization to actual execution. Simultaneously, by using a fallback mechanism of triggering minimum-risk behavior or re-executing the entire decision-making process when verification fails, it ensures that the vehicle has executable safety commands under any circumstances. Thus, it effectively avoids safety hazards introduced by the time delay between decision-making and execution, improving the safety of the autonomous driving system in dynamic environments.
[0120] The present embodiment will now be described and illustrated through preferred embodiments.
[0121] Figure 8 The flowchart of Embodiment 8 of the autonomous driving decision-making method provided in this application is as follows: Figure 8 As shown, the autonomous driving decision-making method includes the following steps:
[0122] Step S801: Obtain the vehicle's driving observation data and determine the preset structural causal model corresponding to the driving observation data;
[0123] Step S802: Based on the preset structural causal model and driving observation data, determine the posterior distribution of the exogenous variables in the preset structural causal model;
[0124] Step S803: Determine multiple candidate driving strategies based on driving observation data, and sample the posterior distribution to obtain multiple exogenous variable samples;
[0125] Step S804: For each candidate driving strategy, determine the future trajectory of the vehicle based on the candidate driving strategy, and for each exogenous variable sample, determine the predicted response behavior of the surrounding traffic participants based on the exogenous variable sample, the future trajectory, and the structural equations corresponding to the surrounding traffic participants in the preset structural causal model.
[0126] Step S805: Perform correlation analysis on future motion trajectory and predicted response behavior, determine multiple evaluation indicators corresponding to candidate driving strategies based on the correlation analysis results, and summarize the multiple evaluation indicators to obtain the sub-counterfactual results corresponding to the exogenous variable samples.
[0127] Step S806: Determine the counterfactual result based on each sub-counterfactual result, and determine the target driving strategy for the vehicle based on the counterfactual result corresponding to each candidate driving strategy, and control the vehicle to perform the corresponding driving actions according to the target driving strategy.
[0128] It should be noted that the steps shown in the above process or in the flowchart of the accompanying figures can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.
[0129] This embodiment also provides an autonomous driving decision-making device for implementing the above embodiments and preferred embodiments; details already described will not be repeated. The terms "module," "unit," "subunit," etc., used below refer to combinations of software and / or hardware that implement a predetermined function. Although the device described in the following embodiments is preferably implemented in software, hardware implementation, or a combination of software and hardware, is also possible and contemplated.
[0130] Figure 9 This is a structural block diagram of the autonomous driving decision-making device in this embodiment, as shown below. Figure 9 As shown, the device includes:
[0131] The retrieval module 10 is used to acquire the vehicle's driving observation data and determine the preset structural causal model corresponding to the driving observation data.
[0132] Processing module 20 is used to determine the posterior distribution of exogenous variables in the preset structural causal model based on the preset structural causal model and driving observation data.
[0133] The analysis module 30 is used to determine multiple candidate driving strategies based on driving observation data, and for each candidate driving strategy, determine the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy and the posterior distribution.
[0134] The decision module 40 is used to determine the target driving strategy of the vehicle based on the counterfactual results corresponding to each candidate driving strategy, and to control the vehicle to perform corresponding driving actions according to the target driving strategy.
[0135] Optionally, in one possible implementation, processing module 20 is specifically used for:
[0136] Obtain the structural equations in the preset structural causal model, and infer the driving state based on the structural equations to obtain driving prediction data;
[0137] Based on driving observation data and driving prediction data, determine the likelihood value corresponding to the driving observation data;
[0138] Obtain the prior distributions of exogenous variables in the predefined causal model, and determine the posterior distributions based on the likelihood values and prior distributions.
[0139] Optionally, in one possible implementation, the analysis module 30 is specifically used for:
[0140] Based on driving observation data, multiple candidate behaviors for the vehicle are generated;
[0141] Acquire the vehicle's current driving environment data and determine environmental constraint information based on the current driving environment data;
[0142] Obtain the vehicle's preset physical constraint information, and for each candidate behavior, determine the motion feasible domain corresponding to the candidate behavior based on the preset physical constraint information and the candidate behavior;
[0143] Based on candidate behaviors, environmental constraint information, and the feasible motion domain corresponding to the candidate behaviors, candidate driving strategies are generated.
[0144] Optionally, in one possible implementation, the analysis module 30 is specifically used for:
[0145] By sampling the posterior distribution, multiple samples of exogenous variables are obtained;
[0146] The vehicle's future trajectory is determined based on the candidate driving strategy. For each exogenous variable sample, the predicted response behavior of the surrounding traffic participants is determined based on the exogenous variable sample, the future trajectory, and the structural equations corresponding to the surrounding traffic participants in the pre-defined structural causal model.
[0147] A correlation analysis is performed on the future motion trajectory and the predicted response behavior, and multiple evaluation indicators corresponding to the candidate driving strategies are determined based on the correlation analysis results.
[0148] By summarizing multiple evaluation indicators, sub-counterfactual results corresponding to the exogenous variable samples are obtained;
[0149] The counterfactual result is determined based on each sub-counterfactual result.
[0150] Optionally, in one possible implementation, the decision module 40 is specifically used for:
[0151] Based on the counterfactual results corresponding to each candidate driving strategy, alternative driving strategies are selected from each candidate driving strategy.
[0152] Obtain the vehicle's preset baseline driving strategy and determine the first risk assessment value corresponding to the preset baseline driving strategy;
[0153] For each alternative driving strategy, a risk assessment is performed on the alternative driving strategy based on the counterfactual results corresponding to the alternative driving strategy to obtain a second risk assessment value, and the difference between the first risk assessment value and the second risk assessment value is determined as the risk improvement amount corresponding to the alternative driving strategy.
[0154] Based on the risk improvement corresponding to each alternative driving strategy, the target driving strategy is selected from the alternative driving strategies.
[0155] Optionally, in one possible implementation, the decision module 40 is specifically used for:
[0156] For each sub-counterfactual result corresponding to the candidate driving strategy, the constraint value corresponding to each preset hard constraint function is determined based on the sub-counterfactual result and each preset hard constraint function, and the minimum value among the constraint values is used as the reference margin corresponding to the sub-counterfactual result.
[0157] The minimum value among the reference margins corresponding to each sub-counterfactual result is taken as the robust margin corresponding to the candidate driving strategy.
[0158] Candidate driving strategies with a robustness margin less than a preset margin threshold are identified as unavailable strategies, and candidate driving strategies other than unavailable strategies are identified as alternative driving strategies.
[0159] Optionally, in one possible implementation, the decision module 40 is specifically used for:
[0160] For each sub-counterfactual outcome corresponding to the alternative driving strategy, the basic risk value corresponding to each evaluation indicator is determined based on multiple evaluation indicators in the sub-counterfactual outcome, and the single sample risk value corresponding to the sub-counterfactual outcome is determined based on each basic risk value.
[0161] The second risk assessment value is determined based on the single-sample risk value corresponding to each sub-counterfactual result.
[0162] Optionally, in one possible implementation, the decision module 40 is specifically used for:
[0163] Obtain the latest value of specified features of the vehicle; the specified features include a combination of one or more of the following: driving event identifiers corresponding to driving observation data, posterior distributions corresponding to exogenous variables, driving observation data, robustness margins corresponding to the target driving strategy, and the execution capability boundaries of the vehicle actuators;
[0164] The latest value of the specified feature is validated, and the vehicle is controlled according to the target driving strategy when the validation passes.
[0165] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.
[0166] This embodiment also provides a vehicle including a control device, which includes a memory and a processor. The memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above method embodiments.
[0167] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementations, and will not be repeated in this embodiment.
[0168] Furthermore, in conjunction with the autonomous driving decision-making methods provided in the above embodiments, this embodiment can also provide a storage medium for implementation. The storage medium stores a computer program; when executed by a processor, the computer program implements any of the autonomous driving decision-making methods described in the above embodiments.
[0169] It should be understood that the specific embodiments described herein are merely illustrative of the application and not intended to limit it. All other embodiments derived by those skilled in the art based on the embodiments provided in this application without inventive effort are within the scope of protection of this application.
[0170] Obviously, the accompanying drawings are merely some examples or embodiments of this application. Those skilled in the art can apply this application to other similar situations based on these drawings without any creative effort. Furthermore, it is understood that although the work done in this development process may be complex and lengthy, for those skilled in the art, certain design, manufacturing, or production modifications made based on the technical content disclosed in this application are merely conventional technical means and should not be considered as insufficient disclosure of this application.
[0171] The term "embodiment" in this application refers to a specific feature, structure, or characteristic described in connection with an embodiment that may be included in at least one embodiment of this application. The appearance of this phrase in various places in the specification does not necessarily imply the same embodiment, nor does it imply that it is mutually exclusive with or independent of other embodiments. It will be clearly or implicitly understood by those skilled in the art that the embodiments described in this application may be combined with other embodiments without conflict.
[0172] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of patent protection. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the appended claims.
Claims
1. An autonomous driving decision-making method, characterized in that, The autonomous driving decision-making method is applied to vehicles; The autonomous driving decision-making method includes: Acquire driving observation data of the vehicle and determine the preset structural causal model corresponding to the driving observation data; Based on the preset structural causal model and the driving observation data, determine the posterior distribution of the exogenous variables in the preset structural causal model; Multiple candidate driving strategies are determined based on the driving observation data, and for each candidate driving strategy, the counterfactual result corresponding to the candidate driving strategy is determined based on the preset structural causal model, the candidate driving strategy, and the posterior distribution. Based on the counterfactual results corresponding to each of the candidate driving strategies, the target driving strategy for the vehicle is determined, and the vehicle is controlled according to the target driving strategy.
2. The autonomous driving decision-making method according to claim 1, characterized in that, The step of determining the posterior distribution of exogenous variables in the preset structural causal model based on the preset structural causal model and the driving observation data includes: Obtain each structural equation in the preset structural causal model, and infer the driving state based on each structural equation to obtain driving prediction data; Based on the driving observation data and the driving prediction data, determine the likelihood value corresponding to the driving observation data; Obtain the prior distribution corresponding to the exogenous variable in the preset causal model, and determine the posterior distribution based on the likelihood value and the prior distribution.
3. The autonomous driving decision-making method according to claim 1, characterized in that, The step of determining multiple candidate driving strategies based on the driving observation data includes: Based on the driving observation data, multiple candidate behaviors of the vehicle are generated; Obtain the current driving environment data of the vehicle, and determine environmental constraint information based on the current driving environment data; Obtain the preset physical constraint information of the vehicle, and for each candidate behavior, determine the motion feasible domain corresponding to the candidate behavior based on the preset physical constraint information and the candidate behavior; The candidate driving strategy is generated based on the candidate behavior, the environmental constraint information, and the motion feasible domain corresponding to the candidate behavior.
4. The autonomous driving decision-making method according to claim 1, characterized in that, The step of determining the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy, and the posterior distribution includes: The posterior distribution is sampled to obtain multiple exogenous variable samples; The future trajectory of the vehicle is determined based on the candidate driving strategy, and for each exogenous variable sample, the predicted response behavior of the surrounding traffic participants is determined based on the exogenous variable sample, the future trajectory, and the structural equation corresponding to the surrounding traffic participants in the preset structural causal model. A correlation analysis is performed on the future motion trajectory and the predicted response behavior, and multiple evaluation indicators corresponding to the candidate driving strategy are determined based on the correlation analysis results. By summarizing the multiple evaluation indicators, the sub-counterfactual results corresponding to the exogenous variable samples are obtained; The counterfactual result is determined based on each of the aforementioned sub-counterfactual results.
5. The autonomous driving decision-making method according to claim 4, characterized in that, The step of determining the target driving strategy for the vehicle based on the counterfactual results corresponding to each of the candidate driving strategies includes: Based on the counterfactual results corresponding to each of the candidate driving strategies, alternative driving strategies are selected from each of the candidate driving strategies; Obtain the preset baseline driving strategy of the vehicle, and determine the first risk assessment value corresponding to the preset baseline driving strategy; For each alternative driving strategy, a risk assessment is performed on the alternative driving strategy based on the counterfactual result corresponding to the alternative driving strategy to obtain a second risk assessment value, and the difference between the first risk assessment value and the second risk assessment value is determined as the risk improvement amount corresponding to the alternative driving strategy. Based on the risk improvement amount corresponding to each of the alternative driving strategies, the target driving strategy is selected from the alternative driving strategies.
6. The autonomous driving decision-making method according to claim 5, characterized in that, The step of selecting alternative driving strategies from the candidate driving strategies based on the counterfactual results corresponding to each candidate driving strategy includes: For each sub-counterfactual result corresponding to the candidate driving strategy, the constraint value corresponding to each preset hard constraint function is determined based on the sub-counterfactual result and each preset hard constraint function, and the minimum value among the various constraint values is used as the reference margin corresponding to the sub-counterfactual result; The minimum value among the reference margins corresponding to each of the sub-counterfactual results is taken as the robust margin corresponding to the candidate driving strategy. Candidate driving strategies with a robustness margin less than a preset margin threshold are identified as unavailable strategies, and candidate driving strategies other than the unavailable strategies are identified as alternative driving strategies.
7. The autonomous driving decision-making method according to claim 5, characterized in that, The step of assessing the risk of the alternative driving strategies based on the counterfactual results corresponding to the alternative driving strategies, and obtaining a second risk assessment value, includes: For each sub-counterfactual result corresponding to the alternative driving strategy, a basic risk value corresponding to each evaluation indicator is determined based on multiple evaluation indicators in the sub-counterfactual result, and a single-sample risk value corresponding to the sub-counterfactual result is determined based on each basic risk value. The second risk assessment value is determined based on the single-sample risk value corresponding to each of the aforementioned counterfactual results.
8. The autonomous driving decision-making method according to claim 6, characterized in that, Before controlling the vehicle according to the target driving strategy, the method further includes: Obtain the latest value of a specified feature of the vehicle; the specified feature includes a combination of one or more of the following: driving event identifiers corresponding to the driving observation data, posterior distributions corresponding to the exogenous variables, driving observation data, robustness margins corresponding to the target driving strategy, and execution capability boundaries of the vehicle actuators; The latest value of the specified feature is verified, and if the verification passes, the vehicle is controlled according to the target driving strategy.
9. An autonomous driving decision-making device, characterized in that, The autonomous driving decision-making device is applied to the vehicle; The autonomous driving decision-making device includes: The retrieval module is used to acquire the driving observation data of the vehicle and determine the preset structural causal model corresponding to the driving observation data; The processing module is used to determine the posterior distribution of the exogenous variables in the preset structural causal model based on the preset structural causal model and the driving observation data. The analysis module is used to determine multiple candidate driving strategies based on the driving observation data, and for each candidate driving strategy, determine the counterfactual result corresponding to the candidate driving strategy based on the preset structural causal model, the candidate driving strategy and the posterior distribution; The decision module is used to determine the target driving strategy of the vehicle based on the counterfactual results corresponding to each of the candidate driving strategies, and to control the vehicle to perform corresponding driving actions according to the target driving strategy.
10. A vehicle, characterized in that, include: A control device, the control device including a memory, a processor, and a computer program; The computer program is stored in the memory and configured to be executed by the processor to implement the autonomous driving decision-making method as described in any one of claims 1 to 8.