Vehicle control method, vehicle, and computer-readable storage medium

Abstract

Embodiments of the present application provide a vehicle control method, a vehicle and a computer readable storage medium. The method comprises: obtaining a game target corresponding to a target vehicle and a game scenario, wherein the game target is a traffic participant target that overlaps with a target running track of the target vehicle, and the game scenario is used to represent a current driving scene type of the target vehicle; determining a game initial state based on ego data of the target vehicle, target data of the game target and the game scenario; executing an interactive game decision with the game initial state as a root node according to a master-slave progressive strategy response mechanism to obtain a decision result; and sending a control instruction to the target vehicle based on the decision result. The present application solves the technical problem that in the related art, when an interactive game simulation of a vehicle and a weak road user is performed, the uncertainty of the behavior of the weak road user is ignored, resulting in inaccurate game results, and further causing vehicle decision errors, thereby affecting vehicle driving safety.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and more specifically, to a vehicle control method, a vehicle, and a computer-readable storage medium. Background Technology

[0002] In the field of autonomous driving, the interaction and interaction between vehicles and vulnerable road users (including pedestrians and non-motorized cyclists) in complex urban scenarios such as intersections without traffic lights, crosswalks, and roadside parking areas is one of the core challenges to achieving vehicle driving safety.

[0003] Current mainstream game theory decision-making methods rely primarily on deterministic prediction models when simulating the interaction between vehicles and vulnerable road users. However, by ignoring the high degree of uncertainty in the behavior of vulnerable road users, the game results are not accurate enough, which in turn leads to vehicle decision-making errors and affects vehicle driving safety.

[0004] There is currently no good solution to the above problems. Summary of the Invention

[0005] This application provides a vehicle control method, a vehicle, and a computer-readable storage medium to at least solve the technical problem in the related art where, in simulating the interaction game between a vehicle and a vulnerable road user, the uncertainty of the vulnerable road user's behavior is ignored, leading to inaccurate game results, which in turn causes vehicle decision-making errors and affects vehicle driving safety.

[0006] According to one aspect of the embodiments of this application, a vehicle control method is provided, comprising: acquiring a game objective and a game scenario corresponding to a target vehicle, wherein the game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle, and the game scenario is used to characterize the current driving scenario type of the target vehicle; determining an initial game state based on the vehicle data of the target vehicle, the target data of the game objective, and the game scenario; executing an interactive game decision with the initial game state as the root node according to a master-slave progressive strategy response mechanism to obtain a decision result, wherein the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game objective, and the decision result includes a target vehicle action sequence represented by a tree structure and a game objective action sequence corresponding to the target vehicle action sequence; and sending a control command to the target vehicle based on the decision result.

[0007] Furthermore, based on the master-slave progressive strategy response mechanism, an interactive game decision is executed with the initial state of the game as the root node to obtain the decision result, including: constructing the root node based on the initial state of the game; cyclically executing the interactive game loop, wherein the interactive game loop includes: starting from the root node, selecting nodes to determine the path to be expanded; for the child nodes to be expanded on the path to be expanded, with the target vehicle as the interactive leader and the game target as the interactive follower, performing node expansion according to the game constraint parameters and the predefined target vehicle expansion action set to obtain the expansion path, wherein the child node to be expanded is the endpoint of the path to be expanded; starting from the expanded child nodes on the expansion path, performing random action simulation to obtain the simulated reward, wherein the expanded child node is the endpoint of the expansion path; backtracking the simulated reward along the expansion path to update the cumulative reward value and access count corresponding to the nodes on the expansion path; and stopping the execution of the interactive game loop in response to the fulfillment of the loop termination condition to obtain the decision result.

[0008] Furthermore, starting from the root node, node selection is performed to determine the path to be expanded, including: starting from the root node, using an upper confidence interval selection strategy to select nodes and obtain a candidate node set, wherein the upper confidence interval selection strategy is used to measure the expandability of nodes; and based on the candidate node set, the path to be expanded is determined.

[0009] Furthermore, the target vehicle's extended action set includes: a first vehicle action and a second vehicle action. For the sub-nodes to be extended on the path to be extended, with the target vehicle as the interaction leader and the game objective as the interaction follower, node extension is performed according to game constraint parameters and a predefined target vehicle extended action set to obtain the extended path. This includes: constructing extended sub-nodes connected to the sub-nodes to be extended on the path to be extended according to game constraint parameters and the target vehicle extended action set, thus obtaining the extended path. The sub-nodes to be extended include: the target vehicle's first vehicle action and first vehicle state, and the game objective's first target action and first target state; the extended sub-nodes include: the target vehicle's second vehicle action and second vehicle state, and the game objective's second target action and second target state. The first vehicle state represents the state of the target vehicle after performing the first vehicle action, the second vehicle state represents the state of the target vehicle after performing the second vehicle action, the first target action represents the game objective's first response to the first vehicle action, and the second target action represents the game objective's second response to the second vehicle action. The first vehicle action, the second vehicle action, the first target action, and the second target action satisfy the constraints of the game constraint parameters.

[0010] Furthermore, starting from the extended child nodes on the extended path, random action simulation is performed to obtain the simulated reward, including: starting from the extended child nodes, executing the simulation loop, wherein the simulation loop includes: randomly selecting any vehicle action from the target vehicle's extended action set as the target vehicle's current action; randomly generating the game target's response action to the current action, wherein the target vehicle's extended action set includes multiple vehicle actions; stopping the execution of the simulation loop in response to the simulation termination condition being met, and obtaining the simulation result; and calculating the simulated reward based on the simulation result.

[0011] Furthermore, based on the simulation results, the simulation reward is calculated, including: determining a first simulation reward and a second simulation reward based on the simulation results, wherein the first simulation reward is used to characterize the reward of the target vehicle for its own game behavior, and the second simulation reward is used to characterize the reward of the target vehicle's own game behavior on the game target; and the first simulation reward and the second simulation reward are weighted and fused to obtain the simulation reward.

[0012] Furthermore, based on the decision results, control commands are sent to the target vehicle, including: determining the target game path based on the decision results, wherein the cumulative reward value of the path corresponding to the target game path is greater than the cumulative reward value of the path corresponding to the remaining game paths in the game decision tree; interpolating and completing the target game path to obtain the target game trajectory; and sending control commands to the target vehicle based on the target game trajectory.

[0013] Furthermore, the game objectives include: non-motorized vehicle drivers and pedestrians; the game scenarios include: left turn scenario, right turn scenario, straight-ahead scenario, and U-turn scenario; and the game constraint parameters include: game duration, the target vehicle's speed and acceleration limit, and the game objective's speed and acceleration limit.

[0014] According to another aspect of the embodiments of this application, a vehicle control system is also provided, comprising: an acquisition module, configured to acquire a game objective and a game scenario corresponding to a target vehicle, wherein the game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle, and the game scenario is used to characterize the current driving scenario type of the target vehicle; a determination module, configured to determine an initial state of the game based on the vehicle data of the target vehicle, the target data of the game objective, and the game scenario; an execution module, configured to execute an interactive game decision with the initial state of the game as the root node according to a master-slave progressive strategy response mechanism, and obtain a decision result, wherein the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game objective, and the decision result includes a target vehicle action sequence represented by a tree structure and a game objective action sequence corresponding to the target vehicle action sequence; and a sending module, configured to send control commands to the target vehicle based on the decision result.

[0015] According to another aspect of the embodiments of this application, a vehicle is also provided, including: a memory storing an executable program; and a processor for running the executable program, wherein the executable program executes the methods in various embodiments of this application when it runs.

[0016] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0017] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0018] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0019] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the methods in various embodiments of this application.

[0020] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods of the various embodiments of this application.

[0021] In this embodiment, the game objective and game scenario corresponding to the target vehicle are obtained. The game objective is a traffic participation target that overlaps with the target vehicle's target trajectory, and the game scenario is used to characterize the current driving scenario type of the target vehicle. Based on the target vehicle's vehicle data, the game objective's target data, and the game scenario, the initial game state is determined. According to the master-slave progressive strategy response mechanism, an interactive game decision is executed with the initial game state as the root node to obtain the decision result. The interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game objective. The decision result includes the target vehicle's action sequence represented by a tree structure and the game objective's action sequence corresponding to the target vehicle's action sequence. Based on the decision result, a control command is sent to the target vehicle. This application first identifies traffic participants whose trajectory overlaps with that of the target vehicle as game targets, i.e., traffic participants that currently affect the safe driving of the target vehicle are taken as game objects, thus clarifying the target of subsequent games. By acquiring the game scenario, the game environment is clarified. The acquisition of game targets and game scenarios is a prerequisite for subsequent game execution. Based on the vehicle data, the target data of the game targets, and the game scenario, the initial state of the game is constructed together, ensuring the physical realism and environmental adaptability of the game model. Furthermore, a master-slave progressive strategy response mechanism is adopted, with the target vehicle as the interaction leader and the game target as the interaction follower. Through dynamic iteration of a tree structure, the action sequences of both parties are explored, so that the response behavior of the game target can evolve in real time with the action strategy of the vehicle. This explicitly models the dynamic reaction capability of the game target under uncertainty, thereby significantly improving the accuracy of the representation of the decision results to the real interaction scenario, thus obtaining more accurate decision results for the target vehicle to make correct vehicle decisions. Finally, based on the decision results, control commands are sent to the target vehicle, enabling the vehicle to execute effective decision actions to ensure the driving safety of the target vehicle. In summary, the technical solution disclosed in this application effectively solves the technical problem in related technologies that, when simulating the interactive game between vehicles and vulnerable road users, the uncertainty of the behavior of vulnerable road users is ignored, leading to inaccurate game results, which in turn causes vehicle decision-making errors and affects vehicle driving safety. The solution achieves the technical effect of improving the accuracy of the game results between vehicles and vulnerable road users, thereby ensuring the correctness of vehicle decisions and guaranteeing vehicle driving safety. Attached Figure Description

[0022] 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:

[0023] Figure 1 This is a flowchart of a vehicle control method according to an embodiment of this application;

[0024] Figure 2 This is a flowchart of another vehicle control method according to an embodiment of this application;

[0025] Figure 3 This is an example diagram of the game subjects according to an embodiment of this application;

[0026] Figure 4 This is a structural block diagram of a vehicle control system according to an embodiment of this application;

[0027] Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort should fall within the scope of protection of the present application.

[0029] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0030] According to an embodiment of this application, an embodiment of a vehicle control method is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, 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.

[0031] This embodiment provides a vehicle control method. Figure 1 This is a flowchart of a vehicle control method according to an embodiment of this application, such as... Figure 1 As shown, the method includes the following steps:

[0032] Step S10: Obtain the game objective and game scenario corresponding to the target vehicle. The game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle. The game scenario is used to characterize the current driving scenario type of the target vehicle.

[0033] Step S11: Determine the initial state of the game based on the target vehicle's own vehicle data, the target data of the game objective, and the game scenario;

[0034] Step S12: According to the master-slave progressive strategy response mechanism, execute the interactive game decision with the initial state of the game as the root node and obtain the decision result. In this case, the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game target. The decision result includes the target vehicle action sequence represented by a tree structure and the game target action sequence corresponding to the target vehicle action sequence.

[0035] Step S13: Based on the decision result, send control commands to the target vehicle.

[0036] The aforementioned game objective refers to vulnerable road users (VRUs) such as non-motorized vehicles or pedestrians who have potential spatial intersections or temporal conflicts within the predicted driving path range of autonomous vehicles.

[0037] VRUs include, but are not limited to, pedestrians, electric bicycle riders, and bicycle riders, who are traffic participants who do not possess the attributes of motor vehicles but have independent movement capabilities.

[0038] In one optional embodiment, the game objective and the target vehicle's target trajectory overlap. This overlap can be defined as follows: within a preset time window (e.g., the next 3 seconds), the planned driving path of the target vehicle (the vehicle center trajectory generated based on lane lines, navigation trajectory and dynamic obstacle avoidance requirements) and the predicted motion trajectory of the VRU share at least one geometric region in space.

[0039] The aforementioned game scenario refers to the current traffic semantic context in which the target vehicle is located, such as going straight, turning left, turning right, making a U-turn, or having no traffic lights at the intersection to yield.

[0040] In one optional embodiment, the game scenario is determined by jointly analyzing the semantic tags of the high-precision map (such as intersection type and lane function) and the vehicle's navigation instructions (such as "turn left in 200 meters ahead"). The game scenario is used to constrain the spatial conditions and priority logic of the game behavior.

[0041] In one optional embodiment, VRU targets around the target vehicle are first identified in real time using a multi-sensor fusion algorithm (such as LiDAR point cloud clustering, camera semantic segmentation, and radar tracking), and their position, speed, orientation, and motion intention are extracted. Then, the future trajectory prediction of each VRU (e.g., using LSTM) is compared with the vehicle's target trajectory for spatiotemporal collision detection, filtering out potential game targets with a collision probability higher than a threshold (e.g., 0.3). Simultaneously, by combining the topology of the high-precision map with lane change instructions output by the vehicle navigation module, the current driving stage is determined to belong to a specific scenario type such as "turning left at an intersection" or "going straight through a pedestrian crossing," serving as the contextual basis for subsequent game strategy space design.

[0042] In another optional embodiment, the target status of surrounding VRUs is first detected in real time by an onboard perception system (such as a camera and millimeter-wave radar fusion perception module), including their two-dimensional position coordinates, velocity vector, heading angle, and body length and width in the global coordinate system. Then, based on a motion prediction model, a sequence of predicted trajectory points for the VRUs over the next five time steps is output. Simultaneously, the vehicle outputs the target's trajectory generated based on the navigation path and road semantic information at the current moment. This trajectory contains several discrete points, and its spatial range extends outward from the vehicle's width (typically 1.8~2.0m) to form a "safety interaction zone." When the predicted trajectory point of any VRU overlaps spatially with this interaction zone (i.e., the shortest distance from the VRU's predicted point to the target's trajectory is less than or equal to half the vehicle's width plus the VRU's radius) and is time-synchronized (the time difference between the prediction time and the vehicle's arrival at the spatial point is less than ±0.5s), the VRU is marked as a game target. This selection process prioritizes lightweight computation, retaining only candidate objects that meet both spatial and temporal constraints, avoiding subsequent high-computational-cost game modeling for irrelevant targets.

[0043] Optionally, in complex urban intersection scenarios, some VRUs may experience multimodal uncertainty in trajectory prediction due to occlusion or sensing noise (e.g., pedestrians may walk straight, stop, or cross). In this case, a multi-hypothesis trajectory sampling strategy is adopted to generate 3-5 possible trajectory branches for each VRU, and calculate the overlap probability between each branch and the vehicle's path. Only when at least one trajectory branch satisfies the above overlap criterion is the VRU considered as the game objective.

[0044] The aforementioned initial state of the game refers to the complete set of environment and behavioral representations used to construct the root node of the search tree before the Monte Carlo Tree Search (MCTS) algorithm is started. It contains multi-dimensional state parameters of the player and the game objective, which are used to support subsequent strategy deduction.

[0045] In one optional embodiment, the vehicle data includes the vehicle's current position coordinates, speed, acceleration, heading angle, vehicle dimensions (length and width), maximum longitudinal acceleration and deceleration, and maximum lateral acceleration. The game objective data includes the game objective's real-time position, speed, acceleration, orientation, motion type (pedestrian / non-motorized vehicle), and historical motion inertial parameters (such as the statistical mean of the rate of change of acceleration). The game scenario describes high-level information about the vehicle's current driving intention and environmental context, such as "waiting to turn left," "going straight," "merging right," or "U-turn." This information originates from navigation instructions and lane-level semantic annotations on high-precision maps and is used to constrain the spatial applicability and behavioral rationality of the game strategy.

[0046] In one optional implementation, the CAN bus data of the vehicle is first obtained from the vehicle control domain, and high-precision pose information output by the positioning module is fused to construct the vehicle's state vector. For each selected game target, the motion state estimate output by the vehicle's perception system is extracted, and its kinematic model parameters are initialized according to its motion type. Next, the potential collision point position between the vehicle and the VRU is calculated based on the game scenario (e.g., "left turn"), serving as the key reference coordinates in the state space. Finally, all the above parameters are integrated into a high-dimensional state vector as the initial state of the game. The initial state of the game includes the vehicle state, the game target state, the scene semantic identifier, and the collision time estimate, providing a unified decision input basis for subsequent strategy progression deduction.

[0047] Optionally, to improve the robustness of the initial state, a state uncertainty modeling mechanism is introduced. For example, for VRUs with high perception noise, their position and velocity information are represented by a Gaussian distribution, with the mean being the observed value and the variance dynamically calculated by the sensor confidence level. This allows for random sampling to simulate perception errors in the real world during subsequent interactive game decisions rooted at the initial state (such as the node expansion phase). Furthermore, the semantics of the game scenario not only affect the legitimacy of the action space (e.g., prohibiting rapid acceleration of vehicles in a "left turn" scenario) but also adjust the reward weight of the game objective during subsequent interactive game decisions. For example, in a "right turn" scenario, the system tends to assign a higher safety priority to pedestrians.

[0048] The aforementioned master-follower progressive strategy response mechanism is a sequential decision-making framework based on Stackelberg game theory. In this framework, the target vehicle, as the "leader" (i.e., the dominant player in the interaction), chooses its action strategy first. The objective vehicle, as the "follower" (i.e., the player in the interaction), observes the actions of the target vehicle and makes its optimal response based on the principle of maximizing its own utility, forming a hierarchical response loop of "decision-feedback-re-decision." This mechanism differs from the synchronous game of Nash equilibrium and more closely reflects the real-world scenario in traffic where the target vehicle has greater control and information advantage.

[0049] The above interactive game decision-making is based on Monte Carlo Tree Search (MCTS), with the initial state of the game as the root node, and the decision tree is expanded layer by layer.

[0050] Within the MCTS framework, this master-slave progressive policy response mechanism is implemented through hierarchical expansion of tree nodes. Each node represents a choice of action by the autonomous vehicle and the optimal response state of the VRU under that action.

[0051] The decision results are multi-step action sequence pairs organized in a tree structure, including the acceleration and steering angle command sequence of the vehicle for several future time steps (e.g., 3 to 5 steps, each step Δt=0.2s), and the corresponding optimal motion state response sequence of the VRU at each step (e.g., speed, acceleration, and orientation change).

[0052] In one optional implementation, MCTS executes four core phases: node selection, node expansion, simulation, and backtracking. In the node selection phase, child nodes with high cumulative rewards but low access frequency are prioritized. In the node expansion phase, the vehicle selects actions from a preset action space, generates new child nodes, and calculates the optimal response of the VRU for each vehicle action based on the Stackelberg mechanism. Under its own kinematic constraints, the VRU solves for its optimal acceleration and angular velocity by minimizing a comprehensive cost function (including collision risk, target deviation, and comfort penalty). In the simulation phase, a lightweight stochastic strategy (such as uniformly sampled actions) is used to perform forward inference from leaf nodes until the game termination condition is met (such as time exhaustion, both speeds approaching zero, or safe passage). In the backtracking phase, the cumulative reward values ​​of all nodes on the path are updated backward according to the designed hierarchical reward function (including vehicle efficiency, safety, comfort rewards, and VRU behavior impact penalties). Finally, after satisfying the maximum search depth or time limit, the final decision tree, i.e., the decision result, is obtained.

[0053] The aforementioned control command refers to the action command corresponding to the target action sequence of the target vehicle determined based on the decision results. This command will be transmitted to the vehicle's underlying motion controller to drive the vehicle to perform safe, efficient, and comfortable game behavior.

[0054] In one alternative embodiment, after each MCTS decision is completed, the first action in the target action sequence (e.g., acceleration = 0.3 m / s², steering angle rate = 0.05 rad / s) is extracted, converted into a control signal recognizable by the vehicle CAN bus (e.g., throttle opening percentage, steering wheel angle command), and sent to the chassis execution unit through the vehicle domain controller.

[0055] In another alternative embodiment, the target motion sequence is linearly or cubically interpolated at time steps (e.g., 0.1 s) to generate smooth velocity and acceleration curves. Based on this result, control commands are generated and sent to the chassis execution unit via the vehicle domain controller.

[0056] In addition, after executing the instruction, the VRU trajectory changes are continuously monitored, and the MCTS process is re-triggered in the next control cycle (e.g., after 100ms) to achieve closed-loop dynamic game.

[0057] In this embodiment, the game objective and game scenario corresponding to the target vehicle are obtained. The game objective is a traffic participation target that overlaps with the target vehicle's target trajectory, and the game scenario is used to characterize the current driving scenario type of the target vehicle. Based on the target vehicle's vehicle data, the game objective's target data, and the game scenario, the initial game state is determined. According to the master-slave progressive strategy response mechanism, an interactive game decision is executed with the initial game state as the root node to obtain the decision result. The interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game objective. The decision result includes the target vehicle's action sequence represented by a tree structure and the game objective's action sequence corresponding to the target vehicle's action sequence. Based on the decision result, a control command is sent to the target vehicle. This application first identifies traffic participants whose trajectory overlaps with that of the target vehicle as game targets, i.e., traffic participants that currently affect the safe driving of the target vehicle are taken as game objects, thus clarifying the target of subsequent games. By acquiring the game scenario, the game environment is clarified. The acquisition of game targets and game scenarios is a prerequisite for subsequent game execution. Based on the vehicle data, the target data of the game targets, and the game scenario, the initial state of the game is constructed together, ensuring the physical realism and environmental adaptability of the game model. Furthermore, a master-slave progressive strategy response mechanism is adopted, with the target vehicle as the interaction leader and the game target as the interaction follower. Through dynamic iteration of a tree structure, the action sequences of both parties are explored, so that the response behavior of the game target can evolve in real time with the action strategy of the vehicle. This explicitly models the dynamic reaction capability of the game target under uncertainty, thereby significantly improving the accuracy of the representation of the decision results to the real interaction scenario, thus obtaining more accurate decision results for the target vehicle to make correct vehicle decisions. Finally, based on the decision results, control commands are sent to the target vehicle, enabling the vehicle to execute effective decision actions to ensure the driving safety of the target vehicle. In summary, the technical solution disclosed in this application effectively solves the technical problem in related technologies that, when simulating the interactive game between vehicles and vulnerable road users, the uncertainty of the behavior of vulnerable road users is ignored, leading to inaccurate game results, which in turn causes vehicle decision-making errors and affects vehicle driving safety. The solution achieves the technical effect of improving the accuracy of the game results between vehicles and vulnerable road users, thereby ensuring the correctness of vehicle decisions and guaranteeing vehicle driving safety.

[0058] The vehicle control method in the embodiments of this application will be further described below.

[0059] Optionally, based on the master-slave progressive strategy response mechanism, an interactive game decision is executed with the initial state of the game as the root node to obtain the decision result, including the following steps:

[0060] Step S121: Based on the initial state of the game, construct the root node;

[0061] Step S122: Execute the interactive game loop repeatedly, which includes: starting from the root node, selecting nodes to determine the path to be expanded; for the child nodes to be expanded on the path to be expanded, with the target vehicle as the interaction leader and the game target as the interaction follower, perform node expansion according to the game constraint parameters and the predefined target vehicle expansion action set to obtain the expansion path, where the child node to be expanded is the endpoint of the path to be expanded; starting from the expanded child node on the expansion path, perform random action simulation to obtain simulated rewards, where the expanded child node is the endpoint of the expansion path; backtrack the simulated rewards along the expansion path and update the cumulative reward value and access count corresponding to the nodes on the expansion path.

[0062] Step S123: In response to the fulfillment of the loop termination condition, stop executing the interactive game loop and obtain the decision result.

[0063] The root node is the starting state node of the MCTS algorithm. It carries all the information included in the initial state of the game and serves as the sole entry point for subsequent node expansion and strategy search.

[0064] The node selection mentioned above refers to the strategy in the MCTS framework of selecting child nodes layer by layer from the root node of the current tree according to the Upper Confidence Bound for Trees (UCB) algorithm, in order to balance the utilization of known high-reward paths and the exploration of insufficiently explored paths.

[0065] The aforementioned node expansion refers to using Stackelberg game as the decision-making mechanism. Based on the child nodes to be expanded, the leader (auto-vehicle) first chooses an action, and the VRUs (followers) respond in real time according to the leader's action, thereby generating new game state nodes.

[0066] In one alternative embodiment, the target vehicle's extended action set (i.e., the vehicle's action space) is defined as a discretized sequence of acceleration and steering angle changes, forming a finite set of actions. For each vehicle action, its future state is deduced based on a dynamic model. The VRU, as a follower, calculates its optimal response action based on its kinematic model and behavioral response mechanism, given a vehicle action, through a cost function minimization strategy. The VRU's action space includes jerk and angular acceleration. Each vehicle-VRU action pair generates a child node whose state includes the predicted trajectories of both parties, time to collision (TTC), relative position, and speed difference, constituting the endpoint of the extended path.

[0067] The aforementioned random action simulation refers to a forward simulation of the game process using a strategy-free random action sampling method based on the leaf nodes (i.e., extended child nodes) generated by node expansion, until the game termination conditions are met (such as reaching the maximum number of simulation steps, both parties' speeds approaching zero, or a collision / safe passage occurring). During the simulation, both the autonomous vehicle and the VRU randomly sample actions according to their respective preset action spaces, without using an optimization model, only simulating the state evolution after the actions are executed. This process is used to evaluate the long-term game outcome of the current path, compensating for the short-sightedness of MCTS at limited depth. In the simulation, collision events are detected in real time (e.g., when the body boxes of both vehicles overlap in space and the time difference is less than 1.0 second, a collision event is considered to have occurred), and it is recorded whether "safe passage" (i.e., both parties did not collide but completed the interaction) or "game failure" (both parties stop) is achieved.

[0068] The simulated rewards mentioned above are calculated based on a preset reward function, which comprehensively considers safety, efficiency, comfort, and behavioral consistency. For example, safety rewards are exponentially negatively correlated with TTC, efficiency rewards are inversely proportional to time deviation, and comfort rewards are negatively correlated with acceleration mutations.

[0069] The aforementioned backtracking refers to updating the cumulative reward value (Q value) and the number of visits (N value) of each node layer by layer along the complete path (i.e., the extended path) from the leaf node back to the root node, based on the simulated reward obtained from the simulation.

[0070] The aforementioned loop termination conditions include, but are not limited to: reaching a preset total number of iterations (e.g., 1500) to ensure search sufficiency; finding a path whose cumulative reward value is significantly higher than other paths and whose visit count accounts for more than a threshold (e.g., 70%), indicating convergence; and simulation time exceeding a preset maximum game duration (e.g., 3.5 seconds) to prevent decision delays from affecting real-time performance. When any termination condition is met, the MCTS iteration stops, and the currently generated decision tree is taken as the decision result.

[0071] In another optional embodiment, based on the finally generated decision tree, all leaf nodes and intermediate nodes are traversed, and the action sequence with the maximum cumulative reward and the highest number of visits is selected as the final decision result. This sequence contains the optimal acceleration and steering commands of the vehicle in the next few time steps, as well as the predicted response trajectory of the corresponding VRU. Optionally, the final decision result can be used to correct the future predicted trajectory of the VRU and generate a smooth, safe, and efficient vehicle motion path.

[0072] Through the above steps, a root node is constructed to fully represent the real-time states (position, speed, acceleration, etc.) of the autonomous vehicle and the VRU, providing computable initial conditions for dynamic game theory. The UCB-guided node selection mechanism achieves an adaptive balance between exploring unknown VRU behavior and utilizing existing experience, avoiding misjudgments caused by perceived noise or prediction bias. The Stackelberg game framework drives node expansion, enabling the autonomous vehicle, as the "leader," to proactively predict the optimal response of the VRU (as the "follower") under different actions, transforming the originally uncontrollable VRU behavior into a modelable conditional response, thereby bypassing the limitations of traditional methods. Traditional methods rely on a "yielding probability threshold." In the simulation phase, a stochastic strategy is used to explore behavioral diversity, combined with a custom multi-dimensional reward function (safety, efficiency, comfort, behavioral inertia, etc.), so that decision-making no longer depends on a single hard rule, but comprehensively evaluates long-term game gains. A backtracking mechanism dynamically updates the cumulative rewards and access counts of nodes, enabling the decision-making process to have online learning and adaptive optimization capabilities, effectively mitigating the problem of false braking caused by unpredictable behaviors such as sudden lane changes and emergency stops by VRUs. Finally, the decision result is output when the termination condition is met, achieving robust decision-making for complex, dynamic, and uncertain interactive scenarios.

[0073] Optionally, starting from the root node, node selection is performed to determine the path to be expanded, including the following steps:

[0074] Step S1221: Starting from the root node, select nodes using the upper confidence interval selection strategy to obtain a candidate node set. The upper confidence interval selection strategy is used to measure the expandability of nodes.

[0075] Step S1222: Determine the path to be extended based on the candidate node set.

[0076] The Upper Confidence Interval (UCB) selection strategy is a decision-making mechanism that balances exploration and exploitation in uncertain environments, widely used in Monte Carlo tree search. The core idea of ​​this strategy is to prioritize the node with the highest overall evaluation value among multiple alternative child nodes. This evaluation value consists of two parts: the node's historical average return (exploitation term) and the uncertainty of the node's access frequency (exploration term). The formula for calculating UCB is:

[0077] (1)

[0078] Where v′ is the current child node and v is the parent node; Q(v') represents the total reward accumulated by the child node since it was visited, reflecting its past performance; N(v') is the number of times the child node has been visited, and N(v) is the total number of times the parent node has been visited; c is the exploration weight coefficient, used to adjust the balance between exploration and exploitation. This formula ensures that nodes with fewer visits have a chance to be explored, while high-reward nodes are still prioritized, thus efficiently approximating the optimal path with limited computing resources.

[0079] For example, the current game tree has 4 nodes, including a, b, c, d, and e, where a is the root node, b and c are child nodes of a, and d and e are child nodes of c. The node selection process is as follows: start selecting nodes from the root node. Since there is only one node at the root node level, the root node a is selected as the first candidate node. Then, traverse the child nodes of a, since c... Value greater than b Therefore, c is selected as the second candidate node. Further, the child nodes of c are traversed, since d... Values ​​greater than e Since the value is d, d is selected as the third candidate node. At this point, the candidate node set includes a, c and d, and the path to be expanded is the path composed of a, c and d.

[0080] It should be noted that the UCB selection process is repeated in each round of MCTS iteration until a node that is not fully developed or a game termination node is reached. In addition, to enhance the sensitivity to the diversity of low-speed VRU behavior, dynamic weighting of the action space can be introduced into the UCB calculation, such as assigning a slightly higher exploration weight to lateral lane change actions to deal with high-risk situations such as non-motorized vehicles suddenly crossing the road.

[0081] Optionally, the target vehicle's extended action set includes: a first vehicle action and a second vehicle action. For the child nodes to be extended on the path to be extended, with the target vehicle as the interaction leader and the game objective as the interaction follower, node extension is performed according to the game constraint parameters and the predefined target vehicle extended action set to obtain the extended path. This includes the following steps: Based on the game constraint parameters and the target vehicle extended action set, extended child nodes connected to the child nodes to be extended on the path to be extended are constructed to obtain the extended path. The child nodes to be extended include: the target vehicle's first vehicle action and first vehicle state, and the game objective's first target action and first target state; the extended child nodes include: the target vehicle's second vehicle action and second vehicle state, and the game objective's second target action and second target state. The first vehicle state is used to represent the state of the target vehicle after performing the first vehicle action, the second vehicle state is used to represent the state of the target vehicle after performing the second vehicle action, the first target action is used to represent the game objective's first response action to the first vehicle action, and the second target action is used to represent the game objective's second response action to the second vehicle action. The first vehicle action, the second vehicle action, the first target action, and the second target action satisfy the constraints of the game constraint parameters.

[0082] The aforementioned target vehicle extended action set is a set of discrete control commands that the vehicle can choose during the game decision-making process, used to drive its state transitions, and forms the basis of the MCTS tree search branch. The first vehicle action and the second vehicle action represent control strategies that the vehicle can execute in different dimensions, such as longitudinal acceleration changes, lateral offset rates, and angular velocity adjustments. Both are independent candidate actions in the action space, forming multidimensional discrete action combinations through parameterized sampling. The first vehicle action can be considered the baseline action for the current game stage, such as maintaining the current speed or slightly decelerating; the second vehicle action is a more exploratory action, such as significant deceleration, lateral avoidance, or acceleration through. Together, they cover the typical behavior patterns of the vehicle in complex interaction scenarios. The construction of the target vehicle extended action set must satisfy vehicle dynamics constraints, such as maximum acceleration, upper limit of jerk, and minimum turning radius, to ensure that all candidate actions are physically executable.

[0083] Optionally, the autonomous vehicle is assigned the role of "leader," while VRUs (such as pedestrians and non-motorized vehicles) are modeled as followers. This role assignment is based on the engineering reality that autonomous vehicles possess greater control and path planning authority in real-world traffic. Autonomous vehicles can proactively adjust their trajectories to coordinate interactions, while the behavior of VRUs is limited by their physiological or motor capabilities, allowing them only limited responses to the autonomous vehicle's decisions. Within the Stackelberg game framework, the leader (autonomous vehicle) first chooses an action (e.g., the first vehicle's action), and then the followers (VRUs), based on that action and their own behavioral models, choose the response action that minimizes their own cost (e.g., the first target action). This sequential decision-making structure effectively simulates the "autonomous vehicle moves first, VRUs respond later" interaction logic in real-world traffic, avoiding the computational explosion and uncertainty amplification problems caused by simultaneous decisions from both sides in Nash equilibrium.

[0084] Optionally, each child node to be expanded represents the game state at a certain moment in the MCTS tree. When expanding a node, the child node is taken as the parent node, and the second vehicle action in the target vehicle's expansion action set is selected as the expansion action. The vehicle's state after time step Δt is calculated using the vehicle kinematics model, i.e., the second vehicle state. Subsequently, according to the Stackelberg game rules, the VRU response model is driven to calculate its optimal response action under this vehicle action (i.e., the second vehicle action), i.e., the second target action. This response action is determined by minimizing the total cost function of the VRU, which includes collision risk, deviation from the original trajectory, and drastic speed change. During the calculation, the VRU's response action is constrained by game-related parameters, including maximum acceleration (e.g., pedestrians not exceeding 1.5 m / s², non-motorized vehicles not exceeding 2.0 m / s²), maximum jerk (e.g., pedestrians not exceeding 3.0 m / s³), and maximum steering angular velocity (e.g., pedestrians not exceeding 0.5 rad / s), ensuring that the response behavior is within human physiological and physical limits. Ultimately, the extended child node contains the complete updated state: the second vehicle state of the autonomous vehicle (driven by the actions of the second vehicle) and the second target state of the VRU (driven by the actions of the second target). These two evolve synchronously through time step Δt, forming a new game path. This path is recorded as a new branch in the MCTS tree for subsequent simulation and backtracking.

[0085] It should be noted that the first objective action is the VRU's response to the action of the first vehicle, and the second objective action is the VRU's response to the action of the second vehicle; both are optimal responses of the VRU under the Stackelberg structure. The generation of the response action does not rely on enumeration but is achieved through continuous optimization. Specifically, for all possible combinations of jerk and angular acceleration under kinematic constraints, the optimal solution is found by minimizing the combined cost function using gradient descent or discrete sampling + greedy search strategies. The cost function consists of five parts: collision safety cost (determined based on the predicted minimum collision time τ_min, using an exponential decay model), target position cost (i.e., the squared Euclidean distance from the desired crossing point or the original path), efficiency cost (i.e., the squared deviation from the reference velocity), state cost (i.e., the squared penalty of acceleration and angular velocity), and input cost (i.e., the squared penalty of jerk and angular acceleration). In the calculation, if the VRU is a pedestrian, its speed limit is set to 1.5 m / s and its acceleration limit is 1.5 m / s²; if it is a non-motorized vehicle, its speed limit is set to 6 m / s and its acceleration limit is 2.5 m / s². Different cost weights are used for both to achieve differentiated modeling. For example, for pedestrians, the safety cost weight is set to 10 and the efficiency cost weight to 1; for non-motorized vehicles, the safety cost weight is set to 8 and the efficiency cost weight to 3, to reflect their higher need to maintain speed. The solution for the response action is executed in real time in each extension step, ensuring that the VRU state of each child node reflects its optimal adaptability to the vehicle's actions, rather than a fixed trajectory prediction.

[0086] In one alternative embodiment, the target vehicle extended action set { is first defined} }, based on the current status of the vehicle The new state after dt is obtained according to the state transition function. The number of extended actions yields an equivalent number of vehicle state extensions. According to Stackelberg game theory, the objective, as a follower, extends its own state based on the extended vehicle state after the vehicle's extension. The main components of this extension include action sampling and state prediction.

[0087] Regarding action sampling, candidate jerks and angular accelerations are generated within the obstacle kinematic constraints, and the sampling range is compared with the current velocity of the game target. Dynamic binding introduces a speed-dependent scaling factor.

[0088] The scope of jerk is as follows: jerk is a scalar, with the direction consistent with the orientation. The sampling range for angular acceleration is: If the obstacle is determined to be an obstacle that only involves vertical game progression, this sample size is 0.

[0089] The acceleration change constraint function is specifically expressed as follows:

[0090] (2)

[0091] in, The velocity decomposition point (e.g., 0.5) 0.8 ), Set the minimum scaling ratio for high speeds (e.g., 0.3) to avoid drastic acceleration and deceleration. and For the target maximum and minimum speeds, the maximum and minimum speed thresholds for pedestrians and non-motorized vehicles are different.

[0092] The constraint function for angular velocity variation is specifically expressed as:

[0093] (3)

[0094] in, The target maximum speed is defined as the maximum speed. The higher the speed, the smaller the allowable steering angular velocity.

[0095] Regarding state prediction, measure child nodes (time step). The state of the game objective (based on multiple factors) The final child node state is obtained by accumulating the data. The acceleration calculation formula is as follows:

[0096] (4)

[0097] The formula for calculating speed is as follows:

[0098] (5)

[0099] The formula for calculating the location is as follows:

[0100] (6)

[0101] (7)

[0102] (8)

[0103] (9)

[0104] The formula for calculating orientation is as follows:

[0105] (10)

[0106] Furthermore, the calculation of the objective state cost in the game mainly includes collision detection and cost calculation. Collision detection primarily involves calculating the collision time and distance between the objective and the player vehicle based on trajectory estimation results. Cost calculation mainly includes safety cost, objective position cost, efficiency cost, acceleration cost, jerk cost, and angular velocity cost.

[0107] Specifically, the formula for calculating the security cost is as follows:

[0108] (11)

[0109] in, For the collision time, It is a constant.

[0110] The formula for calculating the cost of the target location is:

[0111] (12)

[0112] in, Current position For the target location, It is a constant.

[0113] The formula for calculating efficiency cost is:

[0114] (13)

[0115] in, At the current speed, For the target speed, It is a constant.

[0116] The formula for calculating acceleration cost is:

[0117] (14)

[0118] in, For the current acceleration, It is a constant.

[0119] The formula for calculating jerk costs is:

[0120] (15)

[0121] in, To accelerate the current speed, It is a constant.

[0122] The formula for calculating the cost of angular velocity is:

[0123] (16)

[0124] in, The current angular velocity, It is a constant.

[0125] The formula for calculating the total cost is:

[0126] (17)

[0127] Optionally, based on the cost of the game outcome for each action, the action of the target vehicle with the lowest cost and the corresponding game objective state are obtained, specifically represented as follows:

[0128] (18)

[0129] in, The jerk and angular velocity corresponding to the optimal response action of the game objective.

[0130] Optionally, starting from the extended child nodes on the extended path, random action simulations are performed to obtain simulated rewards, including the following steps:

[0131] Step S1223: Starting from the extended child node, execute the simulation loop, wherein the simulation loop includes: randomly selecting any vehicle action from the target vehicle extended action set as the target vehicle's current action; randomly generating the game target's response action to the current action, wherein the target vehicle extended action set includes multiple vehicle actions;

[0132] Step S1224: In response to the simulation termination condition being met, the execution of the simulation loop body is stopped, and the simulation result is obtained;

[0133] Step S1225: Calculate the simulated reward based on the simulation results.

[0134] In one alternative embodiment, each iteration of the simulation loop proceeds in time increments (e.g., Δt = 0.2 s). After the vehicle performs a sampling action, the VRU immediately generates a random response action (not necessarily the optimal response under the Stackelberg structure) based on its perception state. The state is then updated, and the simulation enters the next round. Because the response action generated by the VRU is random, it can simulate the inherent uncertainty of human behavior, avoiding an overly deterministic simulation path that would lose its exploratory significance.

[0135] The simulation termination conditions refer to several preset logical judgment criteria used in the simulation phase of MCTS to determine whether a complete game process should end. These criteria include three categories: time termination, state termination, and game termination. Time termination occurs when the cumulative simulation time reaches a preset upper limit (e.g., 3.0 seconds). Exceeding this time is considered a meaningless extension of the game, and the simulation is forcibly terminated. State termination occurs when the speed of the vehicle or VRU drops to an extremely low threshold (e.g., both less than 0.5 m / s) for more than two time steps. At this point, neither party has any obvious intention to interact, and the game is considered deadlocked. Game termination occurs when the vehicle collides with the VRU (collision time TTC ≤ 0) or the safe distance is completely breached (e.g., the closest distance is less than 0.2 meters). In this case, the simulation immediately terminates and is marked as a failure event. These conditions together constitute robust termination logic, ensuring that each simulation path converges within a reasonable range and avoiding infinite loops or invalid expansions.

[0136] In one optional embodiment, the priority of simulation termination conditions is ordered as collision > stillness > time. That is, once a collision is detected, the simulation ends immediately and is recorded as a high-penalty event; if there is no collision but the speed remains close to zero, it is considered a "game failure" and recorded as a medium penalty; only when there is no collision and both parties are still moving is the simulation allowed to continue to the time limit. This mechanism effectively distinguishes between three scenarios: "dangerous interaction," "ineffective stalemate," and "normal game," providing a clear semantic basis for subsequent reward calculation. In addition, to improve simulation efficiency, state legality checks are added during the simulation process, such as determining whether the VRU exceeds the traversable area of ​​the map or enters a prohibited area (such as a green belt or the outer edge of a sidewalk). If this occurs, the simulation is terminated early and a high penalty is imposed, ensuring that the simulation results have practical significance under both physical and legal constraints.

[0137] The aforementioned simulated reward is a scalar feedback value used to measure the overall quality of a complete simulated path. Its design integrates the three objectives of safety, efficiency, and comfort, and serves as the core basis for the guide tree search direction in MCTS. This reward consists of two parts: a reward for the vehicle itself and a negative reward for its impact on the VRU. Ultimately, these are weighted and fused using behavioral habit factors to form a comprehensive reward.

[0138] Optionally, based on the simulation results, the simulated reward is calculated, including the following steps:

[0139] Step S1226: Based on the simulation results, determine the first simulation reward and the second simulation reward respectively, wherein the first simulation reward is used to characterize the reward of the target vehicle for its own game behavior, and the second simulation reward is used to characterize the reward of the target vehicle's own game behavior on the game target.

[0140] Step S1227: The first simulated reward and the second simulated reward are weighted and fused to obtain the simulated reward.

[0141] The aforementioned first simulated reward refers to the quantified value, from the perspective of the self-vehicle (i.e., the target vehicle), evaluating the positive or negative effects of its current decision-making behavior on the efficiency, safety, and comfort of achieving its driving goals during the Monte Carlo tree search simulation phase. This reward does not directly consider the state changes of the game objective (such as pedestrians or non-motorized vehicles), but focuses on the self-vehicle's own operational performance.

[0142] The first simulation reward consists of several components: safety reward, efficiency reward, acceleration smoothness reward, jerk (jerk) smoothness reward, and action consistency reward. The safety reward is quantified by calculating the shortest collision time (TTC) between the vehicle and the game objective. When the TTC is greater than a preset safety threshold, the reward value approaches its maximum; when the TTC is less than the threshold, the reward decays exponentially to reflect the sharp increase in collision risk. The efficiency reward is calculated based on the ratio of the vehicle's theoretical estimated arrival time at the target point to the actual remaining time. If the vehicle completes the path at a speed close to the theoretical optimal speed, it receives a high reward; otherwise, it is penalized. The acceleration reward and jerk reward constrain the amplitude of longitudinal acceleration and jerk, respectively: when the acceleration or jerk value is below a set comfort threshold, the reward is positive; otherwise, it is zero, thus suppressing abrupt acceleration or deceleration. The action consistency reward is calculated using an exponential decay function by comparing the change in jerk between the current and previous moments, reducing decision jitter and improving behavioral continuity. This reward mechanism ensures that the vehicle does not sacrifice its driving quality due to excessive compromise during the game.

[0143] The aforementioned second simulated reward refers to the quantified value of the indirect impact of the vehicle's decision-making behavior on the game objective during the simulation, particularly the negative disturbances to the objective's comfort, efficiency, and naturalness of movement. The second simulated reward is negative, used to constrain the vehicle's behavior from causing unnecessary fright, sudden stops, or forced changes in movement patterns to vulnerable road users.

[0144] The factors considered in determining the second simulated reward include: the change in the target's speed, the change in its direction, and the target's current speed. The speed change Δv is triggered only when the VRU is forcibly decelerated; that is, when Δv ≤ 0, an exponential decay function is used to calculate the penalty. The change in direction measures the angular deviation of the VRU's forced turn to avoid the voluntary vehicle; the larger the deviation, the stronger the penalty, directly reflecting the interference with the natural movement of the target. The target's current speed is normalized and used as an efficiency compensation term. The introduction of the second simulated reward allows the voluntary vehicle to proactively consider the psychological and physiological impact on vulnerable road users while pursuing its own safety and efficiency, achieving "human-centered" collaborative decision-making.

[0145] In one optional embodiment, the first simulated reward and the second simulated reward are calculated independently during the simulation phase of MCTS, corresponding to the autonomous vehicle's performance and the degree of influence on the VRU, respectively. When the simulated path extends to the game termination condition (such as time running out, both parties remaining stationary, or the interaction being completed), the two are nonlinearly weighted and fused according to a preset behavioral habit parameter α to obtain the simulated reward. This parameter α is a dynamic adjustment factor with a value range of [0, π / 2]. Its value is inferred from the VRU's motion characteristics (such as the historical average acceleration) and is used to characterize the autonomous vehicle's "empathy" towards the VRU's behavior pattern. When α approaches 0, it indicates that the autonomous vehicle prioritizes its own interests and hardly considers the VRU's experience, at which point the weight of the second simulated reward is close to 0; when α approaches π / 2, it indicates that the autonomous vehicle pays close attention to the VRU's comfort, the weight of the second simulated reward reaches its maximum, and the autonomous vehicle tends to actively yield or slow down; when α is π / 4, the two weights are balanced, reflecting a compromise strategy of group collaboration.

[0146] Optionally, during the calculation of the second simulated reward, two types of VRUs are distinguished: pedestrians and non-motorized vehicles. For pedestrians, due to their low inertia and fast response, the sensitivity of Δv and Δθ (change in direction) is set higher; for non-motorized vehicles (such as electric bicycles), their mass and inertia are greater, and the penalty function for Δv adopts a slower decay rate. In addition, during the simulation phase, if the VRU is forced to interrupt its original trajectory due to the behavior of the vehicle (such as turning from a pedestrian crossing to a motor vehicle lane), an additional structured penalty is applied. This penalty is based on the pre-set legal passage area in the scene semantic map. If the VRU is forced into an unexpected area, a fixed negative reward is superimposed to guide the vehicle to respect traffic rules and path expectations.

[0147] Optionally, regarding the calculation of simulated rewards, since the algorithm modifies the low-speed VRU laterally, its predicted trajectory cannot be used for future extrapolation. Therefore, the reward between the vehicle and the low-speed VRU needs to be redesigned, dividing it into a first simulated reward related to the vehicle. and the second simulated reward associated with the vehicle's low-speed VRU Habits of using obstacles (Based on acceleration) Combining the two, specifically expressed as:

[0148] (19)

[0149] when When the value is π / 2, maximize the reward for obstacles. When the value is π / 4, the interests of the group are prioritized, and the joint reward is maximized. When the value is 0, the reward for the vehicle itself is maximized, and the rewards for other low-speed VRUs are ignored.

[0150] First Simulated Reward This includes process rewards and endpoint rewards. Process rewards include: safety rewards, efficiency rewards, acceleration rewards, jerk rewards, and action rewards.

[0151] Based on the speed and direction of the low-speed VRU, the time to reach the vehicle's path is calculated and used as the time to time (TTC) to quantify the safety reward. The safety reward calculation method is as follows:

[0152] (20)

[0153] in, For the calculated collision time, For the maximum collision time, This is the maximum reward.

[0154] Assuming theoretical speed is used as the evaluation criterion, the theoretical time can be calculated based on the distance between the game points. If the leader's remaining time is less than This represents the maximum reward value. Specifically, the efficiency reward is calculated as follows:

[0155] (twenty one)

[0156] in, This represents the estimated actual remaining time for the vehicle to reach the destination.

[0157] Optionally, if the vehicle's current speed is greater than the acceleration threshold, the acceleration bonus is 0; otherwise, the acceleration bonus is calculated as follows:

[0158] (twenty two)

[0159] in, For theoretical acceleration, This is the actual acceleration.

[0160] Optionally, if the vehicle's jerk is greater than the jerk threshold, the jerk reward is 0; otherwise, the jerk reward is calculated as follows:

[0161] (twenty three)

[0162] in, For theoretical jerk, For the actual jerk.

[0163] Optionally, the formula for calculating action rewards is:

[0164] (twenty four)

[0165] Furthermore, the formula for calculating process rewards is as follows:

[0166] (25)

[0167] For the endpoint reward (calculated only at leaf nodes), check if the game outcome reaches the desired state. If the speeds of both the autonomous vehicle and the VRU are close to 0, it is considered that the desired state has not been reached, and a penalty (calibration parameter) is applied; otherwise, the endpoint reward is applied. It is 0.

[0168] Ultimately, the first simulated reward It can be represented as:

[0169] (26)

[0170] Where a and b are constants.

[0171] The second simulated reward assesses the adverse effects of the vehicle's behavior on the low-speed VRU; it is a negative value and can be designed according to requirements. For example, in a scenario where the vehicle's acceleration forces the low-speed VRU to switch from walking to running, the primary consideration is the magnitude of the change in VRU speed. (Measurement of comfort), change in velocity direction (Measures comfort), speed v (measures efficiency). Second simulation reward The calculation is as follows:

[0172] (27)

[0173] in, This represents the maximum speed for pedestrians or non-motorized vehicles, which can be uniformly set to 10 m / s, where c, d, and e are constants. This is only considered during VRU deceleration. , It is expressed as follows:

[0174] (28)

[0175] in, It is a constant.

[0176] When backtracking, the cumulative reward value of the nodes on the backtracking path can be represented as:

[0177] (29)

[0178] in, To simulate rewards, This refers to the cumulative rewards before any retrospective review.

[0179] Optionally, based on the decision result, a control command is sent to the target vehicle, including the following steps:

[0180] Step S131: Based on the decision results, determine the target game path, wherein the cumulative reward value of the path corresponding to the target game path is greater than the cumulative reward value of the path corresponding to the remaining game paths in the game decision tree.

[0181] Step S132: Interpolate and complete the target game path to obtain the target game trajectory;

[0182] Step S133: Based on the target game trajectory, send control commands to the target vehicle.

[0183] The aforementioned objective game path refers to the road behavior trajectory path corresponding to the complete action sequence with the highest cumulative reward value after multiple node selections, expansions, simulations, and backtrackings in the decision tree constructed by Monte Carlo tree search.

[0184] Optionally, after completing the preset maximum number of iterations or reaching the maximum depth, MCTS traverses all leaf nodes and their backtracking paths of the entire decision tree, calculating the total reward value corresponding to each path. This total reward value is obtained by summing the cumulative reward values ​​of each node on that path.

[0185] Because MCTS uses a discrete action space (such as fixed-step acceleration, jerk, and angular velocity sampling) when expanding nodes, the generated target game path is essentially a discrete sequence of points. Therefore, interpolation must be performed on the discrete path to generate a continuous, abrupt, high-density trajectory.

[0186] The aforementioned target game trajectory refers to transforming a discrete sequence of actions into a continuous, smooth, and vehicle-dynamically constrained spatiotemporal trajectory for direct execution by downstream control modules. Interpolation completion refers to using spline interpolation or polynomial fitting methods to generate a high-density, continuously varying sequence of trajectory points for the path nodes at discrete time steps (e.g., one state point every 0.2 seconds) output by the MCTS, ensuring that it meets the requirements of continuity and physical feasibility for velocity, acceleration, and jerk.

[0187] Optionally, cubic spline interpolation is used to spatially smooth the position coordinates, ensuring continuous trajectory curvature and avoiding sudden changes in angular velocity during turning. Simultaneously, fifth-order polynomial interpolation is used for velocity and acceleration to ensure that acceleration and jerk are continuous in the time domain, satisfying dynamic feasibility. For orientation, quaternion interpolation or minimum rotation angle interpolation is used to prevent orientation reversal or circling.

[0188] The aforementioned control commands refer to the underlying execution signals used to drive the target vehicle to travel along the target game trajectory, including longitudinal acceleration commands, lateral steering angle commands, or direct wheel-end torque / steering angle commands.

[0189] In one optional embodiment, the target game trajectory obtained after interpolation is used as a reference input. Based on model predictive control, the optimal control quantity that satisfies the vehicle dynamics model and physical constraints is calculated. For example, based on the deviation between the current vehicle state (position, speed, heading angle) and the target game trajectory, the trajectory error in the next few time steps is predicted, a cost function (including tracking error, rate of change of control quantity, comfort penalty, etc.) is constructed, and the optimal control sequence is solved through numerical optimization, outputting the acceleration and steering angle commands at the first moment.

[0190] Optionally, the game targets include: non-motorized vehicle drivers and pedestrians; the game scenarios include: left turn scenario, right turn scenario, straight-ahead scenario, and U-turn scenario; and the game constraint parameters include: game duration, the target vehicle's speed limit and acceleration limit, and the game target's speed limit and acceleration limit.

[0191] The aforementioned game objectives encompass two typical categories of participants: non-motorized vehicle drivers and pedestrians. Non-motorized vehicle drivers refer to riders using bicycles, electric bicycles, electric scooters, etc., on roads, exhibiting high mobility, non-linear trajectories, and frequent lateral changes of direction. Pedestrians refer to individuals moving along roads or crosswalks using gait patterns; their behavior is highly uncertain and often influenced by psychological expectations, group effects, and environmental guidance. Neither category possesses the same traffic rule compliance capabilities as autonomous vehicles, and their motion is susceptible to perception errors and prediction biases. Therefore, proactive prediction and collaborative decision-making through game theory models are necessary.

[0192] The aforementioned game scenarios include four typical urban road behavior patterns: left-turn, right-turn, straight-ahead, and U-turn. A left-turn scenario refers to a vehicle changing lanes to the left or turning into a side road at an intersection or road segment, which requires lateral space competition with oncoming pedestrians or non-motorized vehicles. A right-turn scenario refers to a vehicle turning right into a side road, often resulting in conflicts with non-motorized vehicles traveling to the right in the same direction or pedestrians crossing a crosswalk. A straight-ahead scenario refers to a vehicle maintaining a straight line in its current lane, but needing to deal with non-motorized vehicles and pedestrians cutting in laterally or suddenly decelerating; this is common at crosswalks without traffic lights or in mixed lanes. A U-turn scenario refers to a vehicle completing a 180° turn in a specific road segment, usually requiring it to occupy the oncoming lane, creating multi-dimensional spatial conflicts with oncoming vehicles, pedestrians, and non-motorized vehicles.

[0193] Optionally, semantic information for each scenario is jointly provided by a high-precision map and a navigation system, and is identified in real time by matching the road topology with the vehicle's current driving intention. The current scenario type is dynamically matched based on the vehicle's positioning and route planning results.

[0194] The aforementioned game-related constraint parameters are a set of boundary conditions set to constrain the game decision space and improve computational efficiency and security. They cover core elements such as game duration, speed and acceleration limits for both the autonomous vehicle and the target vehicle, and the target's trajectory. Game duration refers to the maximum decision-making time window allowed for both the autonomous vehicle and the target VRU during a single game. It is set based on the urgency of the traffic scenario and the vehicle's dynamic response delay, typically ranging from 2 to 5 seconds. Too short a value will lead to insufficient decision-making, while too long a value will increase the computational burden and reduce real-time performance.

[0195] The target vehicle's speed limit and acceleration limit refer to the maximum longitudinal speed and acceleration boundaries that the vehicle is allowed to execute during the game. Their values ​​are set based on vehicle dynamics performance, road speed limits, and comfort standards. For example, in urban roads, the vehicle's speed limit is 30 km / h, and the longitudinal acceleration limit is ±2.5 m / s², in order to avoid sudden braking or acceleration that could cause passenger discomfort or the risk of rear-end collisions.

[0196] The speed and acceleration limits for the game objectives are set separately for different VRU types: the maximum speed for pedestrians is set at 1.5 m / s (approximately 5.4 km / h), and the maximum acceleration is ±1.2 m / s², to conform to the physiological limits of human movement; the maximum speed for non-motorized vehicles is set at 6 m / s (approximately 21.6 km / h), and the maximum acceleration is ±2.0 m / s², to reflect the acceleration capability of electric bicycles.

[0197] The aforementioned limiting parameters together form the physical boundary of the state space, ensuring that the generated action sequence conforms to the feasibility of a real traffic environment.

[0198] Figure 2 This is a flowchart of another vehicle control method according to an embodiment of this application, such as... Figure 2 As shown, the vehicle control method can be implemented as follows: game obstacle screening, game scenario judgment, game parameter update, initial state calculation, game subject execution, game result filtering, and target predicted trajectory correction.

[0199] Optionally, pedestrians and non-motorized vehicle targets whose trajectories overlap with the path considering the vehicle's width are filtered out; combined with navigation and map information, it is determined whether the vehicle is turning left, right, going straight, or making a U-turn; the game time limit, the vehicle's speed and acceleration limit, the game target's speed and acceleration limit, the vehicle's path, and the semantics of the game scenario are updated; the initial state calculation includes: the position, speed, acceleration, orientation, length and width of the vehicle and the game target, and the position of the collision point also needs to be calculated.

[0200] Furthermore, the executing game players, Figure 3 This is an example diagram of the game entities according to embodiments of this application, such as... Figure 3As shown, the execution of the game involves four steps: node selection, node expansion, simulation, and backtracking. Starting from the root node, the optimal child node is selected according to the UCB formula. During node expansion, the Stackelberg game is used as the basis to update the target state. Node expansion mainly includes expanding the vehicle's state, expanding the target state, calculating the target state cost, and calculating the optimal expanded target state. A random strategy is used for simulation, with nodes randomly selected each time during expansion, until one side reaches the game's end point. If the speeds of both sides are close to 0, the game is considered lost. After the simulation, the simulated reward value is backtracked, updating the cumulative reward value and visit count of each node. If the game's exit condition is met (e.g., exceeding the game's total time limit, finding a leaf node), all nodes are traversed, the node sequence with the largest cumulative reward is found, and returned.

[0201] In addition, the obtained game results need to be filtered, and invalidation flags need to be added to unreasonable game branches.

[0202] Finally, the target game result (i.e. the target game path) is selected from the filtered game results, and the result is interpolated and completed to obtain the new predicted trajectory of the target and then updated.

[0203] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation entry points are provided for users to choose to authorize or refuse.

[0204] According to an embodiment of this application, a system embodiment of a vehicle control system is provided. It should be noted that this system can be used to execute the above-described vehicle control method.

[0205] According to another aspect of the embodiments of this application, a vehicle control system is also provided. Figure 4 This is a structural block diagram of a vehicle control system according to an embodiment of this application, such as... Figure 4As shown, the vehicle control system 400 includes: an acquisition module 401, used to acquire the game objective and game scenario corresponding to the target vehicle, wherein the game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle, and the game scenario is used to characterize the current driving scenario type of the target vehicle; a determination module 402, used to determine the initial state of the game based on the vehicle data of the target vehicle, the target data of the game objective, and the game scenario; an execution module 403, used to execute interactive game decisions with the initial state of the game as the root node according to the master-slave progressive strategy response mechanism, and obtain the decision result, wherein the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game objective, and the decision result includes the target vehicle action sequence represented by a tree structure and the game objective action sequence corresponding to the target vehicle action sequence; and a sending module 404, used to send control commands to the target vehicle based on the decision result.

[0206] Optionally, the execution module 403 is further configured to: construct a root node based on the initial state of the game; cyclically execute an interactive game loop, wherein the interactive game loop includes: starting from the root node, selecting nodes to determine the path to be expanded; for the child nodes to be expanded on the path to be expanded, with the target vehicle as the interactive leader and the game target as the interactive follower, performing node expansion according to game constraint parameters and a predefined set of target vehicle expansion actions to obtain the expansion path, wherein the child node to be expanded is the endpoint of the path to be expanded; starting from the expanded child nodes on the expansion path, performing random action simulation to obtain a simulated reward, wherein the expanded child node is the endpoint of the expansion path; backtracking the simulated reward along the expansion path to update the cumulative reward value and access count corresponding to the nodes on the expansion path; and stopping the execution of the interactive game loop in response to the fulfillment of the loop termination condition to obtain a decision result.

[0207] Optionally, the execution module 403 is further configured to: start from the root node, use an upper confidence interval selection strategy to select nodes and obtain a candidate node set, wherein the upper confidence interval selection strategy is used to measure the scalability of nodes; and determine the path to be expanded based on the candidate node set.

[0208] Optionally, the target vehicle's extended action set includes: a first vehicle action and a second vehicle action. The execution module 403 is further configured to: construct extended sub-nodes connected to the sub-nodes to be extended on the path to be extended, based on the game constraint parameters and the target vehicle's extended action set, to obtain the extended path. The sub-nodes to be extended include: the target vehicle's first vehicle action and first vehicle state, and the game target's first target action and first target state; the extended sub-nodes include: the target vehicle's second vehicle action and second vehicle state, and the game target's second target action and second target state. The first vehicle state represents the state of the target vehicle after performing the first vehicle action, the second vehicle state represents the state of the target vehicle after performing the second vehicle action, the first target action represents the game target's first response action to the first vehicle action, and the second target action represents the game target's second response action to the second vehicle action. The first vehicle action, the second vehicle action, the first target action, and the second target action satisfy the constraints of the game constraint parameters.

[0209] Optionally, the execution module 403 is further configured to: start from the extended child node and execute a simulation loop, wherein the simulation loop includes: randomly selecting any vehicle action from the target vehicle's extended action set as the target vehicle's current execution action; randomly generating the game target's response action to the current execution action, wherein the target vehicle's extended action set includes multiple vehicle actions; stopping the execution of the simulation loop in response to the simulation termination condition being met, and obtaining the simulation result; and calculating the simulation reward based on the simulation result.

[0210] Optionally, the execution module 403 is further configured to: determine a first simulated reward and a second simulated reward based on the simulation results, wherein the first simulated reward is used to characterize the reward of the target vehicle for its own game behavior, and the second simulated reward is used to characterize the reward of the target vehicle's own game behavior on the game target; and perform weighted fusion of the first simulated reward and the second simulated reward to obtain the simulated reward.

[0211] Optionally, the sending module 404 is further configured to: determine the target game path based on the decision result, wherein the cumulative reward value of the path corresponding to the target game path is greater than the cumulative reward value of the path corresponding to the remaining game paths in the game decision tree; perform interpolation to complete the target game path to obtain the target game trajectory; extract the target action sequence of the target vehicle from the target game trajectory; and send control commands to the target vehicle based on the target action sequence.

[0212] Optionally, the game targets include: non-motorized vehicle drivers and pedestrians; the game scenarios include: left turn scenario, right turn scenario, straight-ahead scenario, and U-turn scenario; and the game constraint parameters include: game duration, the target vehicle's speed and acceleration limit, the game target's speed and acceleration limit, and the target's trajectory.

[0213] According to another aspect of the embodiments of this application, a vehicle is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0214] Optionally, in this embodiment, the executable program performs the following steps when it runs:

[0215] Step S10: Obtain the game objective and game scenario corresponding to the target vehicle. The game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle. The game scenario is used to characterize the current driving scenario type of the target vehicle.

[0216] Step S11: Determine the initial state of the game based on the target vehicle's own vehicle data, the target data of the game objective, and the game scenario;

[0217] Step S12: According to the master-slave progressive strategy response mechanism, execute the interactive game decision with the initial state of the game as the root node and obtain the decision result. In this case, the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game target. The decision result includes the target vehicle action sequence represented by a tree structure and the game target action sequence corresponding to the target vehicle action sequence.

[0218] Step S13: Based on the decision result, send control commands to the target vehicle.

[0219] According to another aspect of the embodiments of this application, an electronic device is also provided, including: a memory storing an executable program; and a processor for running the program, wherein the program executes the methods in various embodiments of this application when it runs.

[0220] Optionally, Figure 5 This is a schematic diagram of an electronic device according to an embodiment of this application, such as... Figure 5 As shown, the electronic device 500 may include a memory 510 and a processor 520, wherein the memory 510 is used to store an executable program; and the processor 520 is used to run the program stored in the memory 510, and the program executes the methods in various embodiments of this application when it runs.

[0221] According to another aspect of the embodiments of this application, a computer-readable storage medium is also provided, the computer-readable storage medium including a stored executable program, wherein, when the executable program is running, it controls the device where the computer-readable storage medium is located to perform the methods of various embodiments of this application.

[0222] Optionally, in this embodiment, the executable program can be configured to store an executable program for performing the following steps:

[0223] Step S10: Obtain the game objective and game scenario corresponding to the target vehicle. The game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle. The game scenario is used to characterize the current driving scenario type of the target vehicle.

[0224] Step S11: Determine the initial state of the game based on the target vehicle's own vehicle data, the target data of the game objective, and the game scenario;

[0225] Step S12: According to the master-slave progressive strategy response mechanism, execute the interactive game decision with the initial state of the game as the root node and obtain the decision result. In this case, the interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game target. The decision result includes the target vehicle action sequence represented by a tree structure and the game target action sequence corresponding to the target vehicle action sequence.

[0226] Step S13: Based on the decision result, send control commands to the target vehicle.

[0227] According to another aspect of the embodiments of this application, a computer program product is also provided, including a computer program that, when executed by a processor, implements the methods of various embodiments of this application.

[0228] According to another aspect of the embodiments of this application, a computer program product is also provided, including a non-volatile computer-readable storage medium for storing a computer program, which, when executed by a processor, implements the methods in various embodiments of this application.

[0229] According to another aspect of the embodiments of this application, a computer program is also provided, which, when executed by a processor, implements the methods described in the various embodiments of this application.

[0230] In the above embodiments of this application, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0231] In this application, "multiple" refers to two or more.

[0232] In this application, unless otherwise expressly defined, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0233] The terms “first,” “second,” “third,” “fourth,” etc., in this application (if any) are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0234] In this application, the term "and / or" is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, in this application, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0235] In the embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0236] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0237] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0238] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0239] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A vehicle control method, characterized in that, include: Obtain the game objective and game scenario corresponding to the target vehicle, wherein the game objective is a traffic participation target that overlaps with the target running trajectory of the target vehicle, and the game scenario is used to characterize the current driving scenario type of the target vehicle; Based on the vehicle data of the target vehicle, the target data of the game objective, and the game scenario, the initial state of the game is determined. According to the master-slave progressive strategy response mechanism, an interactive game decision is executed with the initial state of the game as the root node to obtain the decision result. The interactive leader corresponding to the master-slave progressive strategy response mechanism is the target vehicle, and the interactive follower corresponding to the master-slave progressive strategy response mechanism is the game target. The decision result includes the target vehicle action sequence represented by a tree structure and the game target action sequence corresponding to the target vehicle action sequence. Based on the decision result, a control command is sent to the target vehicle.

2. The vehicle control method according to claim 1, characterized in that, According to the master-slave progressive strategy response mechanism, the interactive game decision with the initial state of the game as the root node is executed to obtain the decision result, including: Based on the initial state of the game, construct the root node; An interactive game loop is executed cyclically, comprising: starting from the root node, selecting nodes to determine the path to be expanded; for the child nodes to be expanded on the path to be expanded, with the target vehicle as the interaction leader and the game target as the interaction follower, performing node expansion according to game constraint parameters and a predefined set of target vehicle expansion actions to obtain the expansion path, wherein the child node to be expanded is the endpoint of the path to be expanded; starting from the expanded child node on the expansion path, performing random action simulation to obtain a simulated reward, wherein the expanded child node is the endpoint of the expansion path; and backtracking the simulated reward along the expansion path to update the cumulative reward value and access count corresponding to the nodes on the expansion path. In response to the fulfillment of the loop termination condition, the execution of the interactive game loop body is stopped, and the decision result is obtained.

3. The vehicle control method according to claim 2, characterized in that, Starting from the root node, the node selection process is performed to determine the path to be expanded, including: Starting from the root node, a node selection strategy with an upper confidence interval is adopted to select nodes and obtain a candidate node set. The upper confidence interval selection strategy is used to measure the expandability of the nodes. Based on the candidate node set, the path to be extended is determined.

4. The vehicle control method according to claim 2, characterized in that, The target vehicle expansion action set includes: a first vehicle action and a second vehicle action. For the sub-nodes to be expanded on the path to be expanded, with the target vehicle as the interaction leader and the game objective as the interaction follower, the node expansion is performed according to the game constraint parameters and the predefined target vehicle expansion action set to obtain the expanded path, including: Based on the game constraint parameters and the target vehicle's extended action set, extended sub-nodes connected to the sub-nodes to be extended on the path to be extended are constructed to obtain the extended path. The sub-nodes to be extended include: the first vehicle action and first vehicle state of the target vehicle, and the first target action and first target state of the game objective. The extended sub-nodes include: the second vehicle action and second vehicle state of the target vehicle, and the second target action and second target state of the game objective. The first vehicle state represents the state of the target vehicle after performing the first vehicle action, the second vehicle state represents the state of the target vehicle after performing the second vehicle action, the first target action represents the first response action of the game objective to the first vehicle action, and the second target action represents the second response action of the game objective to the second vehicle action. The first vehicle action, the second vehicle action, the first target action, and the second target action satisfy the constraints of the game constraint parameters.

5. The vehicle control method according to claim 2, characterized in that, Starting from the extended child node on the extended path, the random action simulation is performed to obtain the simulated reward, including: Starting from the extended child node, a simulation loop is executed, wherein the simulation loop includes: randomly selecting any vehicle action from the target vehicle extended action set as the current action of the target vehicle; randomly generating the game target's response action to the current action, wherein the target vehicle extended action set includes multiple vehicle actions; In response to the simulation termination condition being met, the execution of the simulation loop body is stopped, and the simulation result is obtained; Based on the simulation results, the simulated reward is calculated.

6. The vehicle control method according to claim 5, characterized in that, Based on the simulation results, the simulated reward is calculated, including: Based on the simulation results, a first simulation reward and a second simulation reward are determined respectively, wherein the first simulation reward is used to characterize the reward of the target vehicle for its own game behavior, and the second simulation reward is used to characterize the reward of the target vehicle's own game behavior on the game target. The simulated reward is obtained by weighted fusion of the first simulated reward and the second simulated reward.

7. The vehicle control method according to claim 2, characterized in that, Based on the decision result, the control command is sent to the target vehicle, including: Based on the decision result, a target game path is determined, wherein the cumulative reward value of the target game path is greater than the cumulative reward value of the remaining game paths in the game decision tree. The target game path is interpolated and completed to obtain the target game trajectory; Based on the target game trajectory, the control command is sent to the target vehicle.

8. The vehicle control method according to claim 2, characterized in that, The game objectives include: non-motorized vehicle drivers and pedestrians; the game scenarios include: left turn scenario, right turn scenario, straight-ahead scenario, and U-turn scenario; and the game constraint parameters include: game duration, the speed and acceleration limit of the target vehicle, and the speed and acceleration limit of the game objective.

9. A vehicle, characterized in that, include: Memory, which stores executable programs; A processor for running the executable program, wherein the executable program, when running, performs the vehicle control method according to any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored executable program, wherein, when the executable program is executed, it controls the device on which the computer-readable storage medium is located to perform the vehicle control method according to any one of claims 1 to 8.

Images