Autonomous driving active decision-making methods, systems and vehicles for interactive scenarios

Abstract

This invention provides an autonomous driving active decision-making method, system, and vehicle for interactive scenarios. The method includes: identifying context anchors that play a decisive role in the evolution of the current interactive scenario; predicting the interaction relationship between the vehicle and the context anchors and the high-level semantic intent of each context anchor; generating a macro-strategy based on long-term goals, and generating candidate trajectories as vehicle actions under the constraints of the macro-strategy; recursively inferring a counterfactual interaction spectrum for each vehicle action; calculating the comprehensive risk potential energy of the counterfactual interaction spectrum using a graph neural network, and mapping the comprehensive risk potential energy into a neural risk potential field; dynamically selecting an action determination strategy based on the geometric morphological characteristics of the neural risk potential field and the prediction confidence of the high-level semantic intent, and determining the vehicle's action sequence based on the action strategy; and converting the action sequence into low-level vehicle control commands and executing them. This invention improves the traffic efficiency, safety, and decision robustness of autonomous vehicles in interactive scenarios.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving technology, specifically to an autonomous driving active decision-making method, system, and vehicle for interactive scenarios. Background Technology

[0002] Currently, the planning and decision-making systems for autonomous vehicles generally adopt a sequential architecture of "perception-prediction-planning". In this architecture, the prediction module predicts the behavior of dynamic participants (such as vehicles and pedestrians) in the environment based on historical trajectories, and the planning module generates a safe and comfortable driving trajectory for the vehicle based on these predictions. However, existing technical solutions have the following limitations: First, the decision-making mechanism is passive. Existing systems only see themselves as observers and adaptors to the traffic environment, unable to proactively guide or influence the future decisions of other participants through their own behavior. In highly interactive game scenarios such as unprotected left turns, ramp merging, and intersection passage, the vehicle often falls into a conservative, hesitant, or even deadlocked state because it cannot change the expected behavior of others, leading to low traffic efficiency or safety hazards. Second, there is a loose coupling between prediction and planning. In traditional methods, the multimodal trajectory output by the prediction module is passed to the planning module as a fixed input. The planning algorithm struggles to assess how different actions of the vehicle will inversely affect the future responses of other participants, resulting in a lack of foresight and coordination in decision-making.

[0003] Therefore, there is an urgent need to provide an autonomous driving active decision-making method, system, and vehicle for interactive scenarios, which can realize active guidance and interactive game reasoning capabilities to improve traffic efficiency, safety, and decision robustness in interactive scenarios. Summary of the Invention

[0004] In view of this, it is necessary to provide an autonomous driving active decision-making method, system and vehicle for interactive scenarios, in order to solve the technical problems of existing technologies that are unable to actively guide and lack interactive game reasoning capabilities, resulting in poor traffic efficiency, safety and decision robustness in interactive scenarios.

[0005] To address the aforementioned technical problems, in a first aspect, the present invention provides an autonomous driving active decision-making method for interactive scenarios, comprising: Identify at least one key dynamic actor from multimodal perception information that plays a decisive role in the evolution of the current interaction scenario, and use it as a context anchor. Based on graph attention network, predict the interaction relationship between the vehicle and the context anchors, as well as the high-level semantic intent of each context anchor; A macro strategy is generated based on a preset long-term goal, and a set of candidate trajectories that meet the vehicle dynamics requirements are generated under the constraints of the macro strategy as the vehicle's actions. Based on the interaction relationship and the high-level semantic intent, a corresponding objective function is set for each context anchor point; the context anchor point is used as a secondary decision-maker with the objective function, and the responses of other vehicles to the context anchor point's actions are recursively inferred, generating a counterfactual interaction spectrum that includes the actions of the vehicle, the responses of other vehicles, and a system joint future state sequence jointly determined by the actions of the vehicle and the responses of other vehicles. The comprehensive risk potential energy of the counterfactual interaction spectrum is calculated using a graph neural network, and the comprehensive risk potential energy is mapped into a neural risk potential field. Based on the geometric morphological characteristics of the neural risk potential field and the prediction confidence of the high-level semantic intent, a guidance strategy, gradient descent strategy, or information gathering strategy is dynamically selected to determine the action sequence. The sequence of actions is converted into vehicle-level control commands and executed.

[0006] In one possible implementation, the multimodal perception information includes visual data, LiDAR data, and V2X data; then, identifying at least one key dynamic participant that plays a decisive role in the evolution of the current interaction scenario from the multimodal perception information includes: The visual data, lidar data, and V2X data are input into a lightweight spatiotemporal Transformer network to obtain at least one key dynamic participant.

[0007] In one possible implementation, the interaction relationship includes competition, courtesy, or cooperation; the high-level semantic intent includes forceful passage, observing and yielding, or slowing down and yielding.

[0008] In one possible implementation, the graph attention network uses the intent statement of the context anchor as prior information.

[0009] In one possible implementation, the step of generating a macroscopic strategy based on a preset long-term goal, and generating a set of candidate trajectories that meet vehicle dynamics requirements under the constraints of the macroscopic strategy, as the vehicle's actions, includes: At least one macro strategy that aligns with the long-term objective is selected from a predefined macro strategy library; the macro strategies include: lane keeping, lane changing, gentle merging, defensive waiting, or decisive passage. The initial cost function of the iterative linear quadratic regulator is generated based on the macro strategy. The initial cost function includes multiple parameters, including target speed, lane change time, or merging gap. Perturb at least one parameter in the initial cost function to generate multiple perturbed cost functions; Using the cost function after each disturbance as the optimization objective, an iterative linear quadratic regulator is run under vehicle dynamics constraints and comfort constraints to generate multiple smooth candidate trajectories, which are then used as the vehicle's actions.

[0010] In one possible implementation, the recursive inference of the responses of each context anchor point to other vehicles in response to the actions of this vehicle includes: The physical and signal changes caused by the vehicle's actions to the surrounding environment are assessed to obtain an updated environmental state; the physical changes include changes in occupied space and speed, and the signal changes include changes in the status of the vehicle's turn signals. The situational anchor is modeled as a secondary decision-maker with its own objective function. Based on the updated environmental state, the optimal response strategy that maximizes the situational anchor's own benefit is determined, and the optimal response strategy is used as the response of the other vehicle.

[0011] In one possible implementation, the training objective of the graph neural network is to minimize the difference between the output and the target value, where the target value is a weighted sum of collision probability, efficiency loss, comfort, traffic law compliance, and social cooperability indicators. The social cooperability indicators are used to evaluate the degree to which the system's joint future state sequence evolves into an ordered state and conforms to the norms of interaction behavior.

[0012] In one possible implementation, the guidance strategy is as follows: when there is a low-risk basin in the neural risk potential field that can be reached through a smooth trajectory sequence, select the action that can fastest move the current motion state of the vehicle along the descent direction of the low-risk basin; The gradient descent strategy is as follows: when the neural risk potential field structure is flat or there are no low-risk basins, calculate the gradient of the neural risk potential field with respect to the current motion state of the vehicle, determine the opposite direction of the gradient as the direction of the fastest local decrease of risk potential energy, and select the action from the actions of the vehicle that makes the angle between the direction of change of the vehicle's motion state and the direction of the fastest decrease the smallest. The information collection strategy is as follows: when the prediction confidence of the high-level semantic intent is lower than the confidence threshold, a tentative action is selected to stimulate the response of the context anchor point in order to improve the prediction confidence of the next prediction cycle.

[0013] Secondly, the present invention also provides an autonomous driving active decision-making system for interactive scenarios, comprising: The context anchor point determination unit is used to identify at least one key dynamic participant that plays a decisive role in the evolution of the current interaction scenario from multimodal perception information, and to serve as the context anchor point. The semantic understanding unit is used to predict the interaction relationship between the vehicle and the context anchors, as well as the high-level semantic intent of each context anchor, based on a graph attention network. The vehicle motion generation unit is used to generate a macro strategy based on a preset long-term goal, and generate a set of candidate trajectories that meet the vehicle dynamics requirements under the constraints of the macro strategy, as the vehicle motion. The counterfactual interaction spectrum generation unit is used to set a corresponding self-objective function for each context anchor point according to the interaction relationship and the high-level semantic intent; and to recursively infer the responses of other vehicles to the context anchor point's actions in response to the actions of the current vehicle, and generate a counterfactual interaction spectrum that includes the actions of the current vehicle, the responses of other vehicles, and a system joint future state sequence jointly determined by the actions of the current vehicle and the responses of other vehicles. The neural risk potential field construction unit is used to calculate the comprehensive risk potential energy of the counterfactual interaction spectrum using a graph neural network, and to map the comprehensive risk potential energy into a neural risk potential field. An action sequence generation unit is used to dynamically select an action determination strategy based on the geometric morphological features of the neural risk potential field and the prediction confidence of the high-level semantic intent, and to determine the action sequence of the vehicle based on the action strategy. The instruction execution unit is used to convert the action sequence into vehicle-level control instructions and execute them.

[0014] Thirdly, the present invention also provides a vehicle, including a processor and a memory, wherein, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the autonomous driving active decision-making method for interactive scenarios described in any of the above possible implementations.

[0015] The beneficial effects of this invention are as follows: The autonomous driving active decision-making method for interactive scenarios provided by this invention identifies at least one contextual anchor point that plays a decisive role in the evolution of the scenario from multimodal perception information, and predicts the interaction relationship between the vehicle and the contextual anchor point and the high-level semantic intent of the contextual anchor point based on a graph attention network, enabling subsequent decisions to focus on key interactive objects. Furthermore, by generating a counterfactual interaction spectrum and constructing a neural risk potential field, it is possible to evaluate the responses of other vehicles and the overall scenario evolution caused by different actions of the vehicle. Compared to the traditional method where prediction and planning are decoupled and the vehicle only passively adapts to the environment, this invention enables the vehicle to proactively evaluate the responses of other vehicles and actively select action sequences, effectively solving the problems of conservatism, hesitation, or deadlock in highly interactive scenarios such as unprotected left turns and ramp merging.

[0016] Furthermore, this invention generates a counterfactual interaction spectrum by recursively inferring the responses of other vehicles to each vehicle's action at the context anchor point, thus jointly evaluating the vehicle's actions and responses within the same framework. This counterfactual interaction spectrum quantitatively characterizes the causal chain of the vehicle's actions-other vehicle responses-system joint future, enabling the decision-making process to compare the differences in the system's joint future state sequences caused by different vehicle actions. Compared to the unidirectional mode in traditional methods where prediction results are passed to the planning module as fixed inputs, this invention significantly improves the foresight and synergy of decision-making.

[0017] In other words, this invention significantly improves the traffic efficiency, safety, and decision-making robustness of autonomous vehicles in interactive scenarios. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a schematic flowchart of an embodiment of the autonomous driving active decision-making method for interactive scenarios provided by the present invention; Figure 2 For the present invention Figure 1 A schematic flowchart of an embodiment of step S103; Figure 3 This is a schematic flowchart of an embodiment of the present invention for recursively inferring the responses of other vehicles to the actions of the vehicle itself at each context anchor point. Figure 4 A schematic diagram of an embodiment of the autonomous driving active decision-making system for interactive scenarios provided by the present invention; Figure 5 A schematic diagram of an embodiment of the vehicle provided by the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0021] It should be understood that the illustrative drawings are not drawn to scale. The flowcharts used in this invention illustrate operations implemented according to some embodiments of the invention. It should be understood that the operations in the flowcharts may be implemented out of order, and steps without logical contextual relationships may be reversed or performed simultaneously. Furthermore, those skilled in the art, guided by the content of this invention, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities may be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor systems and / or microcontroller systems.

[0022] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0023] This invention provides an autonomous driving active decision-making method, system, and vehicle for interactive scenarios, which will be described below.

[0024] Before demonstrating the embodiments, let's illustrate some of the problems with existing technologies with an example. Specifically: In an unprotected left-turn scenario, i.e., at an intersection without a dedicated left-turn green light (i.e., a left-turn arrow light), the vehicle needs to engage in a game with oncoming straight-ahead vehicles: who can pass through the intersection first? Existing methods typically treat themselves as observers of the environment, and their behavioral strategies are essentially reactive. They passively determine whether it is safe and whether there is sufficient clearance based on the current position, speed, and other states of oncoming vehicles.

[0025] If oncoming traffic is heavy, existing methods cause the vehicle to wait until a sufficiently large gap is detected before making a left turn. However, in real-world traffic, human drivers sometimes signal their intention to turn left to oncoming vehicles by slowly moving forward, slightly accelerating, or occupying part of the intersection space, thus guiding them to slow down and give way. Existing systems lack this proactive guidance capability and can only passively wait, resulting in: prolonged delays in completing left turns, causing congestion behind the vehicle; and, when traffic gaps are small, the inability to assess whether proactive acceleration will prompt the oncoming traffic to yield, leading to a deadlock.

[0026] To address the aforementioned technical problems, embodiments of the present invention provide an autonomous driving active decision-making method for interactive scenarios, such as... Figure 1As shown, autonomous driving active decision-making methods for interactive scenarios include: S101. Identify at least one key dynamic participant from multimodal perception information that plays a decisive role in the evolution of the current interaction scenario, and use it as a context anchor.

[0027] Specifically, the context anchor point can be determined based on the spatial distance and relative speed between the vehicle and each dynamic participant, as well as the degree of conflict with the vehicle's planned path.

[0028] For example, in an unprotected left turn scenario, the contextual anchor is an oncoming vehicle going straight and a pedestrian preparing to cross the road.

[0029] By identifying contextual anchor points rather than all dynamic participants, this invention reduces computational load while ensuring traffic efficiency and safety, thereby improving the decision-making efficiency of autonomous driving's proactive decision-making.

[0030] In a specific embodiment of the present invention, the multimodal perception information includes visual data, lidar data, and V2X data; therefore, step S101 specifically comprises: Visual data, LiDAR data, and V2X data are input into a lightweight spatiotemporal Transformer network to obtain at least one key dynamic participant.

[0031] S102. Based on graph attention network, predict the interaction relationship between the vehicle and context anchors, as well as the high-level semantic intent of each context anchor.

[0032] In specific embodiments of the present invention, the interaction relationship includes competition, courtesy, or cooperation; the semantic intent includes forceful passage, observing and yielding, or slowing down and yielding.

[0033] It should be noted that graph attention networks can also output the probability of high-level semantic intents. Specifically, the result can be intent: force passage, probability 0.7 or intent: observe and prepare to give way, probability 0.3.

[0034] This invention, through a graph attention network, outputs high-level semantic intent of contextual anchors. Compared to traditional methods that only output interaction relationships, it directly reveals the interaction tendencies and decision-making motivations of other participants, thereby improving the interpretability of the vehicle's decisions. For example, it can explicitly output that because the other party's intent is to force passage, the vehicle chooses to slow down and wait. As another example, when it detects that the other party's intent is to observe and yield, the vehicle can proactively transform the other party's intent into slowing down and yielding through actions such as tentatively moving forward or flashing its lights, thus guiding the interaction process.

[0035] S103. Generate a macro strategy based on the preset long-term goal, and generate a set of candidate trajectories that meet the vehicle dynamics requirements under the constraints of the macro strategy, as the vehicle's actions.

[0036] In this context, the long-term objective refers to the driving intention that the vehicle hopes to achieve within a macro-timescale (typically a driving planning cycle of 5-10 seconds or longer), including but not limited to: lane keeping, lane changing, gentle merging, defensive waiting, or decisive passage. The long-term objective guides the generation of macro-strategies, ensuring that the vehicle's actions meet current environmental constraints while not deviating from the ultimate expectations at the mission level.

[0037] Vehicle dynamics requirements refer to the requirement that the candidate trajectory generated by the vehicle must meet its physical kinematics and dynamic constraints to ensure that the trajectory is executable in actual control. These requirements include: the continuity and boundary limits of position, velocity, and acceleration (such as maximum steering angle, maximum braking deceleration, and maximum lateral acceleration); the kinematic relationships of the vehicle model; and the physical limits of actuators (such as the response bandwidth and saturation limit of the throttle, brake, and steering motors). By meeting these requirements, it is ensured that the generated candidate trajectory does not exceed the vehicle's physical capabilities while providing a smooth and comfortable ride.

[0038] S104. Based on the interaction relationship and high-level semantic intent, set a corresponding objective function for each context anchor point; treat the context anchor point as a secondary decision-maker with its own objective function, recursively infer the responses of other vehicles to the context anchor point's actions, and generate a counterfactual interaction spectrum that includes the actions of the current vehicle, the responses of other vehicles, and the system's joint future state sequence determined by the actions of the current vehicle and the responses of other vehicles.

[0039] For example, when step S102 predicts that the interaction between the oncoming straight-going vehicle and this vehicle is competition, and its high-level semantic intent is to force passage, the objective function of the context anchor will be set to efficiency first, prioritizing its own straight-going.

[0040] This invention sets its own objective function based on interaction relationships and high-level semantic intent as context anchors, making the recursively deduced responses of other vehicles highly consistent with real driving behavior, thereby significantly improving the credibility of the counterfactual interaction spectrum; at the same time, the vehicle can formulate differentiated guidance strategies to achieve more human-like, efficient and robust interactive decision-making.

[0041] A secondary decision-maker refers to modeling each situational anchor point as an intelligent agent with independent decision-making capabilities. This agent possesses its own state space, set of possible actions, and the aforementioned objective function. In hierarchical game counterfactual inference, the secondary decision-maker can proactively evaluate and select the response strategy that maximizes its own gains based on the updated environmental state after its own actions, thereby generating an interactive closed loop of its own actions and the responses of other vehicles. This allows the counterfactual interaction spectrum to truly reflect the causal logic of the game between the two parties.

[0042] The system's joint future state sequence refers to a joint spatiotemporal state matrix formed by aligning the vehicle's trajectory (generated after the vehicle performs a certain action), the response trajectories of each context anchor point according to its own objective function, and the evolutionary trajectories of other participants in the scene, on a unified time axis with the same discrete step size. This sequence encompasses complete state information of all relevant participants within the next few seconds, including their position, speed, orientation, and intentions. It is directly input into a graph neural network to calculate the comprehensive risk potential energy, thereby quantifying the overall scene evolution risk caused by the vehicle's action.

[0043] S105. Calculate the comprehensive risk potential energy of the counterfactual interaction spectrum using a graph neural network, and map the comprehensive risk potential energy into a neural risk potential field.

[0044] Specifically, the mapping process of the neural risk potential field is as follows: taking the vehicle state (such as position and speed) in the system joint future state sequence corresponding to each vehicle action as the independent variable and the comprehensive risk potential energy output by the graph neural network as the dependent variable, a discrete point cloud is constructed. Then, the discrete point cloud is fitted by radial basis function (RBF) interpolation or by training a small fully connected neural network to generate a continuously differentiable neural risk duration.

[0045] It should be noted that in practical applications, the number of counterfactual interaction spectra is enormous. Therefore, in some embodiments of the present invention, this step can be executed in parallel on a dedicated AI accelerator (such as a GPU / TPU) to ensure the real-time performance of the entire complex process.

[0046] S106. Dynamically select an action determination strategy based on the geometric morphological characteristics of the neural risk potential field and the prediction confidence of the high-level semantic intent, and determine the action sequence of the vehicle based on the action strategy.

[0047] S107. Convert the action sequence into vehicle-level control commands and execute them.

[0048] Among them, the vehicle's underlying control commands can be the control of the accelerator, brakes, and steering.

[0049] It should be understood that the autonomous driving active decision-making method for interactive scenarios in this embodiment of the invention can be implemented in any autonomous driving computing platform based on this active decision-making method, such as an autonomous driving domain controller, a central computing platform, or a dedicated decision-making and planning unit. Specifically, the autonomous driving active decision-making method for interactive scenarios is stored in the aforementioned device in the form of a pre-programmed program or a hardware acceleration core. When the vehicle is in a driving state and the interactive scenario is activated, the program is called or the hardware module is triggered, and the method is executed in real time to generate the optimal action sequence for the vehicle and send it to the underlying execution mechanism.

[0050] Compared with existing technologies, the autonomous driving active decision-making method for interactive scenarios provided by this invention identifies at least one contextual anchor point that plays a decisive role in scenario evolution from multimodal perception information. Based on a graph attention network, it predicts the interaction relationship between the vehicle and the contextual anchor point, as well as the high-level semantic intent of the contextual anchor point, enabling subsequent decisions to focus on key interactive objects. Furthermore, by generating a counterfactual interaction spectrum and constructing a neural risk potential field, it can evaluate the responses of other vehicles and the overall scenario evolution caused by different actions of the vehicle. Compared to the traditional method where prediction and planning are decoupled and the vehicle passively adapts to the environment, this invention enables the vehicle to proactively evaluate the responses of other vehicles and actively select action sequences, effectively solving the problems of conservatism, hesitation, or deadlock in highly interactive scenarios such as unprotected left turns and ramp merging.

[0051] Furthermore, this embodiment of the invention recursively infers the responses of other vehicles to each vehicle's action at the context anchor point, generating a counterfactual interaction spectrum that jointly evaluates the vehicle's actions and responses within the same framework. This counterfactual interaction spectrum quantitatively characterizes the causal chain of the vehicle's actions-other vehicle responses-system joint future, enabling the decision-making process to compare the differences in the system's joint future state sequences caused by different vehicle actions. Compared to the unidirectional mode in traditional methods where prediction results are passed to the planning module as fixed inputs, this embodiment of the invention significantly improves the foresight and synergy of decision-making.

[0052] In other words, the embodiments of the present invention significantly improve the traffic efficiency, safety, and decision-making robustness of autonomous vehicles in interactive scenarios.

[0053] In real-world vehicle-to-infrastructure (V2X) environments, intent declarations actively broadcast by surrounding vehicles can be directly obtained via V2X communication. Using these declarations as prior information input to a graph attention network significantly reduces the uncertainty in predicting future behavior at contextual anchor points, avoiding fuzzy inferences based solely on historical trajectories. Furthermore, when perceptual data is occluded or delayed, intent declarations can serve as strong constraints to correct the allocation of attention weights, making the network's output interaction relationships and semantic intent more accurate and reliable.

[0054] Therefore, in a preferred embodiment of the present invention, the graph attention network uses the intent declaration of the context anchor as prior information.

[0055] In some embodiments of the present invention, such as Figure 2 As shown, step S103 includes: S201. Select at least one macro strategy that meets the long-term goal from the predefined macro strategy library; S202. Generate an initial cost function for an iterative linear quadratic regulator (iLQR) based on a macro strategy. The initial cost function contains multiple parameters, including target speed, lane change time, or merging gap. S203. Perturb at least one parameter in the initial cost function to generate multiple perturbed cost functions; S204. Using the cost function after each disturbance as the optimization objective, run an iterative linear quadratic regulator under vehicle dynamics constraints and comfort constraints to generate multiple smooth candidate trajectories, and use the candidate trajectories as the vehicle's actions.

[0056] This invention first selects a finite number of high-level behavioral patterns that align with long-term goals from a predefined strategy library, ensuring that the action space broadly covers different driving intentions rather than blindly sampling within a continuous action space. Then, for each selected macro-strategy, multiple perturbed cost functions are generated by perturbing key parameters in the iLQR cost function, and optimized separately under dynamics and comfort constraints. This yields a smooth and feasible cluster of candidate trajectories under the same macro-behavioral pattern, achieving deep coverage of the local action space for that strategy. This hierarchical sampling structure allows the vehicle's actions to encompass diverse driving styles while avoiding the redundancy and inefficiency of random sampling in a high-dimensional continuous space, thus achieving a balance between breadth and depth with minimal sampling cost.

[0057] In some embodiments of the present invention, such as Figure 3 As shown, the recursive inference of the responses of each situational anchor point to the actions of other vehicles in response to the actions of this vehicle in step S104 includes: S301. Evaluate the physical and signal changes caused by the vehicle's actions to the surrounding environment to obtain an updated environmental state; physical changes include changes in occupied space and speed, and signal changes include changes in the status of the vehicle's turn signals. S302. Model the situation anchor as a secondary decision-maker with its own objective function, determine the optimal response strategy of the situation anchor to maximize its own benefit based on the updated environmental state, and use the optimal response strategy as the response of other vehicles.

[0058] For example, if the vehicle's action is to accelerate decisively to merge, it will infer whether the oncoming vehicle's objective function is more beneficial to slow down and yield or to maintain speed. If the benefit of slowing down and yielding is higher than the benefit of maintaining speed, the oncoming vehicle will choose to slow down and yield.

[0059] It should be noted that, based on steps S301 and S302, the recursive reasoning is not a single calculation of the other vehicle's response, but rather iterates the above process repeatedly on the time axis: First, the vehicle takes action and updates the environmental state, and each context anchor point outputs the optimal response strategy according to its own objective function; then, these responses are regarded as new environmental changes, the vehicle's secondary response is evaluated again, and the secondary response of the anchor point is triggered again, and so on until the interaction converges or the preset recursion depth is reached.

[0060] The embodiments of the present invention can perform deep modeling of the cyclic causal chain of "this vehicle - other vehicles - this vehicle" in complex interactive scenarios through a recursive mechanism, thereby generating a more robust optimal action sequence.

[0061] Existing risk assessment methods typically rely solely on discrete rules designed based on physical indicators such as collision time or safe distance, making it difficult to comprehensively measure traffic efficiency, passenger comfort, and implicit interaction norms generally followed by human drivers. Furthermore, manually designed weighting coefficients often depend on expert experience and cannot adapt to complex and ever-changing interaction scenarios.

[0062] To address this technical problem, in some embodiments of the present invention, the training objective of the graph neural network is to minimize the difference between the output and the target value. The target value is a weighted sum of collision probability, efficiency loss, comfort, traffic law compliance, and social cooperability indicators. The social cooperability indicator is used to evaluate the degree to which the system's joint future state sequence evolves into an ordered state and conforms to the interaction behavior norms.

[0063] Among these, the interaction behavior norms are not the hard rules stipulated by laws and regulations, but rather the implicit interaction norms that human drivers generally follow. For example, when an intersection narrows and vehicles need to merge, vehicles in the left and right lanes usually take turns passing.

[0064] This invention, by setting the comprehensive risk potential energy output by a graph neural network to simultaneously reflect physical safety, traffic efficiency, ride comfort, regulatory compliance, and social cooperation, allows autonomous driving decisions to move beyond a purely collision-avoidance-based approach. Instead, it enables autonomous driving to comprehensively weigh multiple objectives, much like a human driver, and to act safely, efficiently, and in accordance with social expectations in complex interactions. Secondly, by introducing social cooperation indicators, the vehicle's behavior patterns are more easily understood and accepted by other road users, significantly improving the integration and cooperation of autonomous vehicles in mixed traffic flows. Furthermore, because the graph neural network is trained with a weighted sum as its objective, its internal nonlinear mapping implicitly captures the complex coupling between various indicators, avoiding the abrupt changes and discontinuities at boundaries inherent in manual rules. This results in a smoother risk potential field and more reliable gradient information, ultimately enhancing the robustness and generalization ability of the entire decision-making system.

[0065] In a specific embodiment of the present invention, the action determination strategy includes a guidance strategy, a gradient descent strategy, and an information collection strategy; The guidance strategy is as follows: when there is a low-risk basin in the neural risk potential field that can be reached through a smooth trajectory sequence, select the action that can fastest move the current motion state of the vehicle along the downward direction of the low-risk basin. The gradient descent strategy is as follows: when the neural risk potential field structure is flat or there are no low-risk basins, calculate the gradient of the neural risk potential field with respect to the current motion state of the vehicle, determine the opposite direction of the gradient as the direction of the fastest local decrease of risk potential energy, and select the action from the vehicle's actions that makes the angle between the direction of change of the vehicle's motion state and the direction of the fastest decrease the smallest. The information gathering strategy is as follows: when the prediction confidence of the high-level semantic intent is lower than the confidence threshold, a tentative action is selected to stimulate the response of the context anchor, so as to improve the prediction confidence of the next prediction cycle.

[0066] In this context, a low-risk basin refers to a connected local region within the neural risk potential field. Within this region, the risk potential energy values ​​at all points are below a preset safety threshold, and the potential energy gradient at the region boundary points inward, meaning the potential energy monotonically decreases from the boundary to the center. Simultaneously, the second derivative (Hessian matrix) of this region is positive definite, ensuring that the potential energy surface exhibits a "bowl-shaped" concavity within the region. Specifically, basins can be identified by calculating the local minima and their attraction regions within the risk potential field. The attraction region is the set of all state points that can eventually converge to the same low-risk basin via a gradient descent path. Furthermore, the existence of a smooth trajectory sequence (e.g., the potential energy values ​​at all points on the path are below the current potential energy, and the gradient direction consistency is above a preset angle threshold) is evaluated.

[0067] Structural flatness refers to a situation where, within the local neighborhood of the vehicle's current state, the magnitude of the first-order gradient of the potential function in the neural risk potential field is less than a preset flatness threshold, and the eigenvalues ​​of the second derivative are all close to zero. This indicates that the potential energy changes extremely slowly within this region, and there are no obvious local minima or maxima. In this case, the terrain of the risk potential field approximates a horizontal plane or has slight undulations, making it impossible to identify attractive low-risk basins or determine a single steepest descent direction.

[0068] A probing action refers to a small, short-duration, reversible maneuver that the vehicle actively chooses when the prediction confidence of the high-level semantic intent is lower than the confidence threshold. Its purpose is not to directly minimize the current risk potential energy, but to stimulate a clear response from the situational anchor point by changing its own state (such as slight deceleration, slight lateral drift, short honking or flashing lights), so as to obtain higher confidence information about the response of other vehicles in the next decision cycle and reduce the uncertainty of the risk potential field.

[0069] The ultimate goal of these three action determination strategies is to enable the autonomous driving system to adaptively select the most appropriate action determination logic in any interaction scenario, thereby achieving a dynamic optimal balance between safety and efficiency. The guidance strategy utilizes the basin structure in the potential field for long-term planning, avoiding getting trapped in local optima; the gradient descent strategy quickly reduces current risks and ensures basic safety when there is no clear objective; and the information gathering strategy proactively addresses cognitive uncertainty, acquiring key information through trial and error to optimize subsequent decisions. This approach balances safety, traffic efficiency, and decision robustness across all scenarios.

[0070] The following are examples of the three strategies in specific scenarios: Example of a guidance strategy: In a scenario where a vehicle merges from a ramp into a main road, the vehicle is in the acceleration lane, and there is a vehicle traveling at a constant speed behind it on the main road. At this point, the neural risk potential field presents a clear low-risk basin: a stable state after the vehicle successfully merges into the main road and maintains a safe following distance. There exists a trajectory sequence that can reach this basin through smooth acceleration and timely merging. However, if the vehicle continues at its current constant speed, it will miss the merging window and enter a high-risk area. The recommended guidance strategy is for the vehicle to decisively and moderately accelerate to preemptively occupy the merging point. While this action may momentarily shorten the collision time with the following vehicle (the instantaneous risk is not the lowest), its purpose is to quickly push the state towards the downward direction of the basin, actively guiding the following vehicle to perceive the vehicle's merging intention and slow down to give way, ultimately allowing both vehicles to slide into a low-risk, orderly merging state.

[0071] Example of gradient descent strategy: In a scenario of driving on a straight, open road, there are no vehicles or obstacles in front of the vehicle, lane lines are clear, and traffic conditions are simple. At this time, the neural risk potential field is generally flat: the potential energy gradient amplitude is extremely small, there are no obvious low-risk basins or high-risk ridges, and the risk change from any slight deviation from the lane or speed fluctuation is almost zero. In this situation, no attractive long-term objective can be identified, so the gradient descent strategy is chosen: The gradient of the current potential field with respect to the vehicle's motion state is calculated, and the negative gradient direction is found to point towards maintaining the current lane centerline and driving at a constant speed. Actions that minimize the angle between the vehicle's direction of motion and this negative gradient direction are selected from the vehicle's actions (i.e., almost no steering wheel adjustment, throttle opening remains unchanged), thus maintaining the current safe state at the lowest cost. This strategy avoids meaningless probing or jittery movements when there is no clear optimization objective.

[0072] Example of information gathering strategy: In an unprotected left turn scenario, the vehicle reaches the stop line at an intersection, where there is an oncoming vehicle going straight. Since the oncoming vehicle neither uses its turn signal nor broadcasts any V2X intent, and its historical trajectory fluctuates between stable straight-going and slight deceleration, the probability of both aggressive clearance and yielding intent in the counterfactual interaction spectrum is 0.5, resulting in a prediction confidence level below the set threshold (e.g., 0.7). At this point, the risk potential field exhibits high uncertainty in the areas corresponding to the vehicle's forward movement and waiting actions, meaning it's impossible to be certain whether the oncoming vehicle will slow down. Instead of blindly choosing to turn left or wait, a probing action is initiated: the vehicle slowly coasts forward 0.5 meters and slightly illuminates its brake lights while flashing its left turn signal. This action itself carries extremely low risk (not entering the conflict zone), but it can elicit a clear response from the oncoming vehicle: if the oncoming vehicle begins to decelerate, the confidence level shifts towards yielding, and a decisive left turn can be made in the next cycle; if the oncoming vehicle accelerates or honks, the confidence level shifts towards aggressive clearance, and a stop and wait option is chosen in the next cycle. Through information gathering strategies, the system transforms uncertainty into observable information, thereby enabling better decisions in subsequent cycles.

[0073] In summary, the autonomous driving proactive decision-making method for interactive scenarios proposed in this invention represents a paradigm shift from passive adaptation to proactive guidance. It resolves traffic impasses through explicit strategic actions, improving traffic efficiency and safety. The introduced hierarchical game theory model endows the system with high-level collaborative capabilities, making interactive behavior more human-like and easier to understand. Simultaneously, the neural risk potential field provides an intuitive decision-making "topographical map," greatly enhancing the system's interpretability and verifiability. At the product and commercial level, this solution effectively reduces travel time and improves driving comfort. Its modular design and hardware acceleration optimization ensure the real-time performance of complex algorithms, enhancing the system's robustness and market competitiveness in long-tail scenarios. In terms of social benefits, intelligent vehicles with proactive collaborative capabilities can act as "traffic lubricants," optimizing macro-level traffic flow and providing a key technological foundation for the future construction of a vehicle-road cooperative ecosystem.

[0074] On the other hand, embodiments of the present invention also provide an autonomous driving active decision-making system for interactive scenarios, such as... Figure 4 As shown, the autonomous driving active decision-making system 400 for interactive scenarios includes: The context anchor point determination unit 401 is used to identify at least one key dynamic participant that plays a decisive role in the evolution of the current interaction scenario from multimodal perception information, and use it as a context anchor point. The semantic understanding unit 402 is used to predict the interaction relationship between the vehicle and the context anchors and the high-level semantic intent of each context anchor based on the graph attention network. The vehicle motion generation unit 403 is used to generate a macro strategy based on a preset long-term goal, and generate a set of candidate trajectories that meet the requirements of vehicle dynamics under the constraints of the macro strategy, as the vehicle motion. The counterfactual interaction spectrum generation unit 404 is used to set a corresponding objective function for each context anchor point according to the interaction relationship and high-level semantic intent; the context anchor point is used as a secondary decision-maker with its own objective function, and the response of the context anchor point to the action of the vehicle is recursively inferred to the response of the vehicle to the action of the vehicle, and a counterfactual interaction spectrum containing the action of the vehicle, the response of the vehicle, and the system joint future state sequence determined by the action of the vehicle and the response of the vehicle. The neural risk potential field construction unit 405 is used to calculate the comprehensive risk potential energy of the counterfactual interaction spectrum using a graph neural network and map the comprehensive risk potential energy into a neural risk potential field. Action sequence generation unit 406 is used to dynamically select an action determination strategy based on the geometric morphological features of the neural risk potential field and the prediction confidence of the high-level semantic intent, and determine the action sequence of the vehicle based on the action strategy. The instruction execution unit 407 is used to convert the action sequence into vehicle low-level control instructions and execute them.

[0075] The autonomous driving active decision-making system 400 for interactive scenarios provided in the above embodiments can implement the technical solutions described in the above embodiments of the autonomous driving active decision-making method for interactive scenarios. The specific implementation principles of each module or unit can be found in the corresponding content in the above embodiments of the autonomous driving active decision-making method for interactive scenarios, and will not be repeated here.

[0076] like Figure 5 As shown, the present invention also provides a vehicle 500. The vehicle 500 includes a processor 501, a memory 502, and a display 503. Figure 5 Only some components of vehicle 500 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0077] In some embodiments, processor 501 may be a microcontroller in a vehicle, used to run program code stored in memory 502 or process data, such as the autonomous driving active decision-making method for interactive scenarios in this invention.

[0078] In some embodiments, memory 502 may be an internal storage unit of vehicle 500, such as a hard disk or memory of vehicle 500. In other embodiments, memory 502 may also be an external storage device of vehicle 500, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on vehicle 500.

[0079] Furthermore, the memory 502 may include both internal storage units of the vehicle 500 and external storage devices. The memory 502 is used to store application software and various types of data installed on the vehicle 500.

[0080] In some embodiments, display 503 may be an LED display, a liquid crystal display, a touch-screen liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 503 is used to display information about vehicle 500 and to display a visual user interface. Components 501-503 of vehicle 500 communicate with each other via a device bus.

[0081] In some embodiments of the present invention, when the processor 501 executes the autonomous driving active decision-making program for interactive scenarios stored in the memory 502, the following steps can be implemented: Identify at least one key dynamic actor from multimodal perception information that plays a decisive role in the evolution of the current interaction scenario, and use it as a context anchor. Based on graph attention network, predict the interaction relationship between the vehicle and context anchors, as well as the high-level semantic intent of each context anchor; A macro strategy is generated based on a preset long-term goal, and a set of candidate trajectories that meet the requirements of vehicle dynamics are generated under the constraints of the macro strategy as the vehicle's actions. Based on the interaction relationship and high-level semantic intent, a corresponding objective function is set for each context anchor point; the context anchor point is treated as a secondary decision-maker with its own objective function, and the responses of other vehicles to the context anchor point's actions are recursively inferred to generate a counterfactual interaction spectrum that includes the actions of the current vehicle, the responses of other vehicles, and the system's joint future state sequence determined by the actions of the current vehicle and the responses of other vehicles. The comprehensive risk potential energy of the counterfactual interaction spectrum is calculated using a graph neural network, and then mapped to a neural risk potential field. Based on the geometric morphological characteristics of the neural risk potential field and the prediction confidence of the high-level semantic intent, an action determination strategy is dynamically selected, and the action sequence of the vehicle is determined based on the action strategy. The action sequence is converted into vehicle-level control commands and executed.

[0082] It should be understood that when the processor 501 executes the autonomous driving active decision-making program based on the interactive scenario in the memory 502, in addition to the functions mentioned above, it can also implement other functions, as can be found in the description of the corresponding method embodiments above.

[0083] Those skilled in the art will understand that all or part of the processes of the methods described in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.), and the computer program can be stored in a computer-readable storage medium. The computer-readable storage medium may be a disk, optical disk, read-only memory, or random access memory, etc.

[0084] The present invention provides a detailed description of an autonomous driving active decision-making method, system, and vehicle for interactive scenarios. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, those skilled in the art will recognize that there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An autonomous driving active decision-making method for interactive scenarios, characterized in that, include: Identify at least one key dynamic actor from multimodal perception information that plays a decisive role in the evolution of the current interaction scenario, and use it as a context anchor. Based on graph attention network, predict the interaction relationship between the vehicle and the context anchors, as well as the high-level semantic intent of each context anchor; A macro strategy is generated based on a preset long-term goal, and a set of candidate trajectories that meet the vehicle dynamics requirements are generated under the constraints of the macro strategy as the vehicle's actions. Based on the interaction relationship and the high-level semantic intent, a corresponding objective function is set for each context anchor point; the context anchor point is used as a secondary decision-maker with the objective function, and the responses of other vehicles to the context anchor point's actions are recursively inferred, generating a counterfactual interaction spectrum that includes the actions of the vehicle, the responses of other vehicles, and a system joint future state sequence jointly determined by the actions of the vehicle and the responses of other vehicles. The comprehensive risk potential energy of the counterfactual interaction spectrum is calculated using a graph neural network, and the comprehensive risk potential energy is mapped into a neural risk potential field. Based on the geometric morphological features of the neural risk potential field and the prediction confidence of the high-level semantic intent, a dynamic action determination strategy is selected, and the action sequence of the vehicle is determined based on the action strategy. The sequence of actions is converted into vehicle-level control commands and executed.

2. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The multimodal perception information includes visual data, LiDAR data, and V2X data; therefore, identifying at least one key dynamic participant that plays a decisive role in the evolution of the current interaction scenario from the multimodal perception information includes: The visual data, lidar data, and V2X data are input into a lightweight spatiotemporal Transformer network to obtain at least one key dynamic participant.

3. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The interaction relationships include competition, courtesy, or cooperation; the high-level semantic intents include forceful passage, observing and yielding, or slowing down and yielding.

4. The autonomous driving active decision-making method for interactive scenarios according to claim 1 or 3, characterized in that, The graph attention network uses the intent declaration of the context anchor as prior information.

5. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The process of generating a macro-strategy based on a preset long-term goal, and generating a set of candidate trajectories that meet vehicle dynamics requirements under the constraints of the macro-strategy as the vehicle's actions, includes: At least one macro strategy that aligns with the long-term objective is selected from a predefined macro strategy library; the macro strategies include: lane keeping, lane changing, gentle merging, defensive waiting, or decisive passage. The initial cost function of the iterative linear quadratic regulator is generated based on the macro strategy. The initial cost function includes multiple parameters, including target speed, lane change time, or merging gap. Perturb at least one parameter in the initial cost function to generate multiple perturbed cost functions; Using the cost function after each disturbance as the optimization objective, an iterative linear quadratic regulator is run under vehicle dynamics constraints and comfort constraints to generate multiple smooth candidate trajectories, which are then used as the vehicle's actions.

6. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The recursive inference of the responses of each situational anchor point to the actions of other vehicles in response to the actions of this vehicle includes: The physical and signal changes caused by the vehicle's actions to the surrounding environment are assessed to obtain an updated environmental state; the physical changes include changes in occupied space and speed, and the signal changes include changes in the status of the vehicle's turn signals. The situational anchor is modeled as a secondary decision-maker with its own objective function. Based on the updated environmental state, the optimal response strategy that maximizes the situational anchor's own benefit is determined, and the optimal response strategy is used as the response of the other vehicle.

7. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The training objective of the graph neural network is to minimize the difference between the output and the target value. The target value is a weighted sum of collision probability, efficiency loss, comfort, traffic law compliance, and social coordination indicators. The social coordination indicator is used to evaluate the degree to which the system's joint future state sequence evolves into an ordered state and conforms to the interaction behavior norms.

8. The autonomous driving active decision-making method for interactive scenarios according to claim 1, characterized in that, The action determination strategy includes a guidance strategy, a gradient descent strategy, and an information gathering strategy; The guidance strategy is as follows: when there is a low-risk basin in the neural risk potential field that can be reached through a smooth trajectory sequence, select the action that can fastest move the current motion state of the vehicle along the descent direction of the low-risk basin; The gradient descent strategy is as follows: when the neural risk potential field structure is flat or there are no low-risk basins, calculate the gradient of the neural risk potential field with respect to the current motion state of the vehicle, determine the opposite direction of the gradient as the direction of the fastest local decrease of risk potential energy, and select the action from the actions of the vehicle that makes the angle between the direction of change of the vehicle's motion state and the direction of the fastest decrease the smallest. The information collection strategy is as follows: when the prediction confidence of the high-level semantic intent is lower than the confidence threshold, a tentative action is selected to stimulate the response of the context anchor point in order to improve the prediction confidence of the next prediction cycle.

9. An autonomous driving active decision-making system for interactive scenarios, characterized in that, include: The context anchor point determination unit is used to identify at least one key dynamic participant that plays a decisive role in the evolution of the current interaction scenario from multimodal perception information, and to serve as the context anchor point. The semantic understanding unit is used to predict the interaction relationship between the vehicle and the context anchors, as well as the high-level semantic intent of each context anchor, based on a graph attention network. The vehicle motion generation unit is used to generate a macro strategy based on a preset long-term goal, and generate a set of candidate trajectories that meet the vehicle dynamics requirements under the constraints of the macro strategy, as the vehicle motion. The counterfactual interaction spectrum generation unit is used to set a corresponding self-objective function for each context anchor point according to the interaction relationship and the high-level semantic intent; and to recursively infer the responses of other vehicles to the context anchor point's actions in response to the actions of the current vehicle, and generate a counterfactual interaction spectrum that includes the actions of the current vehicle, the responses of other vehicles, and a system joint future state sequence jointly determined by the actions of the current vehicle and the responses of other vehicles. The neural risk potential field construction unit is used to calculate the comprehensive risk potential energy of the counterfactual interaction spectrum using a graph neural network, and to map the comprehensive risk potential energy into a neural risk potential field. An action sequence generation unit is used to dynamically select an action determination strategy based on the geometric morphological features of the neural risk potential field and the prediction confidence of the high-level semantic intent, and to determine the action sequence of the vehicle based on the action strategy. The instruction execution unit is used to convert the action sequence into vehicle-level control instructions and execute them.

10. A vehicle, characterized in that, Including processor and memory, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the autonomous driving active decision-making method for interactive scenarios as described in any one of claims 1 to 8.

Images