Trajectory determination methods, devices, equipment, and vehicles in autonomous driving scenarios

Through cascaded decision-making recursive steps and probabilistic chain propagation, the autonomous driving system generates target optimization trajectories in complex interactive scenarios, solving the problem of multi-obstacle interaction coupling, achieving a balance between traffic safety, riding experience and traffic efficiency, and improving the decision-making ability and cooperative driving ability of autonomous vehicles.

CN123104694APending Publication Date: 2026-11-17APOLLO INTELLIGENT DRIVING (BEIJING) TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611330789.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing autonomous driving systems struggle to achieve the optimal balance between traffic safety, passenger experience, and traffic efficiency in complex interactive scenarios. Furthermore, decision-making complexity is high in scenarios with multiple obstacles and interactions, and existing methods are ill-suited to handle multi-level, cascading game-theoretic decision-making problems.

Method used

A cascaded recursive decision-making process is adopted, which expands multiple decision nodes recursively. Based on the initial context information of the root node and the cascaded recursive decision-making strategy, key obstacles are identified and multiple candidate decision nodes are generated. Global and local decision probabilities are used for chain propagation to select decision nodes that meet the conditions and generate the target optimization trajectory.

Benefits of technology

It significantly enhances the system processing capabilities of autonomous vehicles in scenarios with multiple obstacles and interaction, enabling them to make reasonable decisions quickly, improve traffic efficiency, ensure traffic safety and passenger experience, adapt to complex road scenarios, and drive collaboratively with other vehicles.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123104694A_ABST
    Figure CN123104694A_ABST
Patent Text Reader

Abstract

This disclosure provides a trajectory determination method, apparatus, device, and vehicle for autonomous driving scenarios, involving fields such as artificial intelligence, autonomous driving, intelligent transportation, electric vehicles, trajectory planning, and deep learning. The specific implementation scheme is as follows: A cascaded decision recursive step is executed to obtain candidate trajectories of the vehicle corresponding to decision nodes that meet certain conditions; and an optimized trajectory of the vehicle is obtained based on the candidate trajectories. The cascaded decision recursive step includes: determining the i-th level key obstacle corresponding to the i-th level node, at least two (i+1)-th level candidate decision nodes, and the i-th local decision probability of the (i+1)-th level candidate decision nodes based on the candidate trajectories and obstacle state information in the i-th level context information of the i-th level node; and determining the (i+1)-th global decision probability of the (i+1)-th level candidate decision nodes based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of the (i+1)-th level candidate decision nodes, thereby determining the candidate trajectory of the (i+1)-th level decision nodes that meet the extended conditions.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This disclosure relates to the field of data processing technology, and in particular to the fields of artificial intelligence, autonomous driving, intelligent transportation, electric vehicles, trajectory planning and deep learning. Background Technology

[0002] In the field of autonomous driving, autonomous vehicles inevitably interact with other vehicles on the road, engaging in behaviors such as merging, overtaking, and yielding. Due to the inherent limitations of vehicle perception and prediction capabilities, and the highly uncertain and dynamic nature of the road environment, current autonomous driving systems struggle to make reasonable and accurate trajectory decisions in complex interaction scenarios. To ensure traffic safety, existing systems generally employ conservative yielding strategies, but this strategy not only severely impacts the passenger experience but also significantly reduces road efficiency and increases the risk of traffic congestion. Therefore, achieving the optimal balance between traffic safety, passenger experience, and traffic efficiency in complex interaction scenarios has become a core challenge that current autonomous driving technology urgently needs to address. Summary of the Invention

[0003] This disclosure provides a trajectory determination method, apparatus, device, and vehicle for autonomous driving scenarios.

[0004] According to one aspect of this disclosure, a trajectory determination method in an autonomous driving scenario is provided, comprising: The cascaded decision recursive steps are executed to obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and the target optimized trajectory of the target vehicle is obtained based on at least one linear candidate trajectory; the decision node is the node extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node; wherein, the cascaded decision recursive steps include: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle, the i-th level key obstacle corresponding to the i-th level node is determined; wherein, the i-th level node is the root node or the i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability and the state information of each obstacle; Based on the i-th level key obstacle, at least two i+1 level candidate decision nodes for the i-th level node are determined, as well as the i-th local decision probability of each i+1 level candidate decision node; where the i-th local decision probability represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the i+1 level candidate decision node. Based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th global decision probability of each (i+1)-th level candidate decision node is determined. Based on the (i+1)th global decision probability of each (i+1)th level candidate decision node, the (i+1)th level decision node that satisfies the expansion condition is determined, and the (i+1)th linear candidate trajectory of each (i+1)th level decision node is obtained.

[0005] According to another aspect of this disclosure, a trajectory determination device for an autonomous driving scenario is provided, comprising: A recursive processing unit is used to execute cascaded decision recursive steps, wherein the cascaded decision recursive steps include: determining the i-th level key obstacle corresponding to the i-th level node based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle; wherein the i-th level node is the root node or an i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle; and determining at least two (i+1)-th level candidate decision nodes for the i-th level node, and each (i+1)-th level candidate decision node, based on the i-th level key obstacle. The local decision probability of the i-th decision node; where the local decision probability of the i-th decision node represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the (i+1)-th level candidate decision node; based on the global decision probability corresponding to the i-th level node and the local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th global decision probability of each (i+1)-th level candidate decision node is determined; based on the (i+1)-th global decision probability of each (i+1)-th level candidate decision node, the (i+1)-th level decision node that meets the expansion condition is determined, and the (i+1)-th linear candidate trajectory of each (i+1)-th level decision node is obtained; The trajectory output unit is used to obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and to obtain the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory; the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0006] According to another aspect of this disclosure, an electronic device is provided, comprising: At least one processor; and The memory is communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform any of the methods described in the present disclosure.

[0007] According to another aspect of this disclosure, a non-transitory computer-readable storage medium is provided storing computer instructions, wherein the computer instructions are used to cause the computer to perform any of the methods according to embodiments of this disclosure.

[0008] According to another aspect of this disclosure, a computer program product is provided, including a computer program that, when executed by a processor, implements any of the methods according to embodiments of this disclosure.

[0009] According to another aspect of this disclosure, an autonomous driving vehicle is provided, wherein the autonomous driving vehicle includes the electronic devices described above.

[0010] In this way, the disclosed solution can decompose the complex multi-obstacle interaction decision problem into a recursive problem of processing key obstacles layer by layer by utilizing cascaded decision-making recursive steps, effectively reducing the complexity of multi-obstacle joint decision-making. At the same time, the resulting target vehicle's optimized trajectory has higher feasibility and consistency with driving behavior, and is more in line with actual driving needs. Thus, it significantly enhances the systematic processing capability of autonomous vehicles in the face of multi-obstacle interaction coupling problems, enabling them to make reasonable decisions quickly in complex scenarios such as overtaking, merging, and multi-vehicle games, thereby improving traffic efficiency while ensuring traffic safety and passenger experience, allowing autonomous vehicles to better adapt to complex road scenarios and cooperate with other social vehicles.

[0011] Furthermore, this disclosed solution obtains the probability of selecting a continuous interactive decision path from the root node to the current candidate decision node by chaining the global decision probability of the previous level node with the local decision probability of the current candidate decision node, which is the (i+1)th global decision probability. Further, based on this (i+1)th global decision probability, the (i+1)th level decision nodes that meet the expansion conditions are selected. In this way, the exponential expansion of the decision tree caused by low-probability branches is effectively avoided. At the same time, this disclosed solution can also generate candidate trajectories that match the preset interactive decisions for the retained (i+1)th level decision nodes. In this way, it ensures that each node not only has decision meaning, but also has trajectory results that can be used for subsequent recursion and final selection.

[0012] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this disclosure, nor is it intended to limit the scope of this disclosure. Other features of this disclosure will become readily apparent from the following description. Attached Figure Description

[0013] The accompanying drawings are provided to better understand this solution and do not constitute a limitation of this disclosure. Wherein: Figure 1 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 1 ; Figure 2 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 2 ; Figure 3 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 3 ; Figure 4 This is an illustrative diagram of a tree-like decision-making framework according to an embodiment of this application; Figure 5(a) is a flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of the present application in an example; Figure 5(b) is a schematic flowchart of the "two-level optimization" in the trajectory determination method in an autonomous driving scenario according to an embodiment of this application; Figure 6(a) is an illustrative diagram of a cascaded decision-making recursive mechanism for multiple obstacles according to an embodiment of the present application in a specific application scenario; Figure 6(b) is an illustrative diagram illustrating the application of the present disclosure scheme in an offline scenario diffusion process according to an embodiment of this application; Figure 7 This is a schematic diagram of the structure of a trajectory determination device in an autonomous driving scenario according to an embodiment of this application; Figure 8 This is a block diagram of an electronic device used to implement the trajectory determination method in an autonomous driving scenario according to embodiments of the present disclosure. Detailed Implementation

[0014] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0015] In this document, the term "and / or" merely describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. The term "at least one" in this document indicates any combination of at least two of a plurality of elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C. The terms "first" and "second" in this document refer to and distinguish between multiple similar technical terms, not to restrict the order or to limit there to only two. For example, "first feature" and "second feature" refer to two categories / two features; the first feature can be one or more, and the second feature can also be one or more.

[0016] Furthermore, to better illustrate this disclosure, numerous specific details are set forth in the following detailed description. Those skilled in the art will understand that this disclosure can still be practiced even without certain specific details. In some instances, methods, means, components, and circuits well known to those skilled in the art have not been described in detail in order to highlight the main points of this disclosure.

[0017] The following describes the related technologies of the embodiments of this disclosure. The following related technologies are optional solutions and can be combined with the technical solutions of the embodiments of this disclosure in any way, and they all fall within the protection scope of the embodiments of this disclosure.

[0018] In the field of autonomous driving, autonomous vehicles inevitably interact with other vehicles on the road, engaging in behaviors such as merging, overtaking, and yielding. Due to the inherent limitations of vehicle perception and prediction capabilities, and the highly uncertain and dynamic nature of the road environment, current autonomous driving systems struggle to make reasonable and accurate trajectory decisions in complex interaction scenarios. To ensure traffic safety, existing systems generally employ conservative yielding strategies, but this strategy not only severely impacts the passenger experience but also significantly reduces road efficiency and increases the risk of traffic congestion. Therefore, achieving the optimal balance between traffic safety, passenger experience, and traffic efficiency in complex interaction scenarios has become a core challenge that current autonomous driving technology urgently needs to address.

[0019] In addition, existing trajectory planning methods suffer from the following key technical bottlenecks: (1) Limitations of static prediction assumptions: Most existing methods treat surrounding vehicles as rigid obstacles moving along the predicted trajectory, ignoring the game-like nature of human-vehicle interaction in actual traffic scenarios, resulting in deviations between decision results and actual interaction behavior.

[0020] (2) The problem of multi-obstacle interaction coupling: When a vehicle needs to make interactive decisions with multiple key obstacles at the same time (such as giving way to vehicle A before overtaking vehicle B), existing methods are difficult to effectively handle this multi-level, cascaded game decision problem.

[0021] Based on this, the present disclosure provides a trajectory determination method for autonomous driving scenarios. This method can enhance the decision-making ability and interactivity of autonomous vehicles in complex road environments, and realize multi-level game interaction with other vehicles on the road. Specifically, within a decision cycle, this method can recursively expand multiple decision nodes based on the initial context information of the root node and a cascaded recursive decision strategy. For example, firstly, based on the i-th context information of the i-th level node, the i-th level key obstacle corresponding to the i-th level node is determined, and based on the i-th level key obstacle, at least two (i+1)-th level candidate decision nodes of the i-th level node are determined, as well as the i-th local decision probability of each (i+1)-th level candidate decision node; then, using the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th level candidate decision nodes are determined. The (i+1)th global decision probability of each decision node is used to determine the (i+1)th level decision node that meets the expansion conditions, and then the (i+1)th linear candidate trajectory of each (i+1)th level decision node is obtained. Thus, in road scenarios involving multiple obstacles, the target vehicle interacts with surrounding key obstacles layer by layer in a recursive cascade manner, according to two types of preset interaction decisions (e.g., yielding and overtaking), thereby expanding to multiple decision nodes. Finally, the target vehicle's optimized trajectory is determined based on the linear candidate trajectory corresponding to the decision node that meets the termination conditions. This significantly enhances the systematic processing capability of autonomous vehicles for multi-obstacle interaction coupling problems, enabling autonomous vehicles to quickly make reasonable decisions in complex scenarios such as overtaking, merging, and multi-vehicle game theory. This improves traffic efficiency while ensuring traffic safety and passenger comfort, allowing autonomous vehicles to better adapt to complex road scenarios and cooperate with other vehicles.

[0022] Specifically, Figure 1 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 1 This method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, or in-vehicle electronic devices. Alternatively, this method can be optionally applied to autonomous vehicles.

[0023] Furthermore, the method includes at least a portion of the following: For example... Figure 1 As shown, it includes: Step S101: Execute the cascaded decision recursive steps; that is, the cascaded decision recursive steps may specifically include: Step S101-1: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node (i is an integer greater than or equal to 0) and the state information of each obstacle, determine the i-th level key obstacle corresponding to the i-th level node.

[0024] Here, it should be noted that the i-th level key obstacle can be specifically understood as one of the obstacles that currently requires interactive decision-making with the target vehicle under the context information of the i-th level node, or as the obstacle that most currently requires interactive decision-making with the target vehicle.

[0025] Here, the i-th level node is the root node or the i-th level decision node cascaded with the root node. For example, in one example, when i is 0, the i-th level node can be the root node, while when i is an integer greater than 0, such as 1, the i-th level node is the i-th level decision node cascaded with the root node.

[0026] Furthermore, in one example, each level of node, such as the i-th level node, contains complete context information for storing and managing the key data required for decision-making. In this way, through the standardized management of the context container, the efficient transmission of key information between parent and child nodes can be effectively achieved, ensuring that each level of decision-making can accurately obtain historical decision information from the upper level, avoiding duplicate calculations and information loss.

[0027] Furthermore, in one example, the i-th context information of the i-th level node includes the i-th initial candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle. Wherein, (1) The i-th linear candidate trajectory: If the i-th node is the root node, the i-th linear candidate trajectory can be specifically the initial linear trajectory; if the i-th level node is the decision node, the i-th linear candidate trajectory can be specifically obtained by optimizing the linear candidate trajectory of the next higher level node based on the preset interactive decision corresponding to the decision node. For example, in one example, the i-th linear candidate trajectory can be represented by a multi-dimensional state vector. At this time, the multi-dimensional state vector includes, but is not limited to: position, heading angle (θ), curvature (κ), velocity, acceleration, time, and jerk (that is, the rate of change of acceleration with respect to time, which is used to describe the smoothness of the vehicle's acceleration and deceleration process).

[0028] (2) The global decision probability of the i-th node can be specifically the path probability or decision confidence of extending from the root node to the i-th level node.

[0029] (3) The status information of each obstacle may specifically include at least one of the following: the obstacle's position, speed, acceleration, heading angle, size, type, historical trajectory, predicted trajectory, future occupied area, behavioral intention, updated trajectory information, or updated status information.

[0030] It should be noted that the state information of obstacles does not remain unchanged throughout the cascaded decision recursion process. After the target vehicle completes the interaction with the key obstacle in the previous node, the state information of the key obstacle may need to be updated with the interaction result. In this way, the environmental changes and state changes are dynamically reflected in the recursion, avoiding conflicting or invalid decisions made by the lower nodes based on outdated states.

[0031] Furthermore, in one example, the i-th context information of the i-th level node can also include the i-th label of each obstacle. In this case, the i-th label can specifically be one of unprocessed, processed, or background object. This provides a reference for subsequently identifying key obstacles.

[0032] Step S101-2: Based on the key obstacles of level i, determine at least two candidate decision nodes of level i+1 for the level i node, and the local decision probability of each candidate decision node of level i+1.

[0033] It should be noted that the local decision probability of the (i+1)th level candidate decision node can be specifically used to represent the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the (i+1)th level candidate decision node.

[0034] Furthermore, the (i+1)th level candidate decision node can be specifically understood as a node that may expand into the next level when the i-th level node is dealing with the i-th level key obstacle, and each (i+1)th level candidate decision node corresponds to a preset interaction decision, such as any interaction strategy in a traffic scenario like overtaking, merging, or yielding. For example, in one example, there is a correspondence between the (i+1)th level candidate decision nodes and the preset interaction strategies, such as a one-to-one correspondence, that is, one candidate decision node corresponds to one preset interaction strategy.

[0035] Furthermore, in one example, the preset interactive decisions of different candidate decision nodes at different levels (i+1) corresponding to the i-th level node are different. This ensures that autonomous vehicles can make reasonable decisions quickly in complex scenarios such as overtaking, merging, and multi-vehicle games, rather than making a single decision, thereby improving traffic efficiency while ensuring traffic safety and passenger experience.

[0036] Furthermore, in one example, the preset interaction strategy includes one of the following: overtaking, yielding, following, decelerating to avoid, accelerating through, and lane changing. For example, when the i-th level key obstacle is vehicle A, two (i+1)-th level candidate decision nodes can be generated for vehicle A, and the preset interaction strategies corresponding to the two candidate decision nodes can be "overtake vehicle A" and "yield to vehicle A", respectively.

[0037] Step S101-3: Based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, determine the (i+1)-th global decision probability of each (i+1)-th level candidate decision node.

[0038] Here, the (i+1)th global decision probability of the (i+1)th level candidate decision node specifically represents the probability that the continuous interactive decision path from the root node to the i-th level node is selected. In other words, the (i+1)th global decision probability of the (i+1)th level candidate decision node specifically characterizes the overall reliability of the cascaded interactive decision sequence corresponding to the (i+1)th level candidate decision node after the target vehicle starts from the root node, passes through one or more levels of decision nodes, and arrives at the (i+1)th level candidate decision node.

[0039] Step S101-4: Based on the global decision probability of each (i+1)th level candidate decision node, determine the (i+1)th level decision node that satisfies the expansion condition, and obtain the (i+1)th linear candidate trajectory of each (i+1)th level decision node.

[0040] For example, in one instance, the extended condition may specifically be that the global decision probability is greater than or equal to a preset probability threshold. It is understood that in practical applications, this extended condition may be based on actual traffic scenarios or actual autonomous driving rules, and this disclosure does not impose specific restrictions on it.

[0041] Step S102: After the cascaded decision recursive steps are completed, the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition is obtained, and the target optimized trajectory of the target vehicle is obtained based on at least one linear candidate trajectory.

[0042] It should be noted that the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0043] Furthermore, the root node can be specifically understood as the starting node of the cascading decision recursive process within the current decision cycle. For example, in one example, the root node usually corresponds to the initial state when the target vehicle has not yet made an interactive decision regarding any obstacle.

[0044] Furthermore, the candidate trajectory mentioned above may specifically include the state sequence of the target vehicle within a future time window, such as its position, speed, acceleration, heading angle, and curvature.

[0045] Furthermore, in one example, the initial context information may specifically include the target vehicle's initial linear trajectory, the initial global decision probability (e.g., a preset value), and the state information of each obstacle among multiple obstacles.

[0046] Furthermore, in another example, the initial context information may specifically include the initial labels for each obstacle. Here, the initial label may be an initialized value, such as initialized to unprocessed.

[0047] It should be noted that, in practical applications, the context information mentioned above (e.g., initial context information or i-th context information) may also specifically include: scene information of the scene where the target vehicle is located, such as a signed distance field, which is used to record the spatial distance field information of roads and obstacles in the scene where the target vehicle is located, so as to be used for fast calculation of obstacle avoidance constraints.

[0048] Furthermore, it should be noted that the signed distance field can be continuously updated during the recursive process. This ensures that autonomous vehicles can make reasonable decisions quickly in complex scenarios such as overtaking, merging, and multi-vehicle games, rather than making a single decision, thereby improving traffic efficiency while ensuring traffic safety and passenger comfort.

[0049] Furthermore, in another example, the context information mentioned above (e.g., initial context information or i-th context information) may also specifically include scene optimization results, such as the interaction results after the target scene interacts with the obstacle (e.g., the interaction process can be converted into an optimization problem, and the solution result can be used as the interaction result), such as convergence state, number of iterations, degree of constraint violation, etc.

[0050] It should be noted that the scene optimization result in the initial context information can also be a specific initial value, and the initial value can be updated based on the optimization result during subsequent optimization processes.

[0051] In this way, the disclosed solution can decompose the complex multi-obstacle interaction decision problem into a recursive problem of processing key obstacles layer by layer by utilizing cascaded decision-making recursive steps, effectively reducing the complexity of multi-obstacle joint decision-making. At the same time, the resulting target vehicle's optimized trajectory has higher feasibility and consistency with driving behavior, and is more in line with actual driving needs. Thus, it significantly enhances the systematic processing capability of autonomous vehicles in the face of multi-obstacle interaction coupling problems, enabling them to make reasonable decisions quickly in complex scenarios such as overtaking, merging, and multi-vehicle games, thereby improving traffic efficiency while ensuring traffic safety and passenger experience, allowing autonomous vehicles to better adapt to complex road scenarios and cooperate with other social vehicles.

[0052] Furthermore, this disclosed solution obtains the probability of selecting a continuous interactive decision path from the root node to the current candidate decision node by chaining the global decision probability of the previous level node with the local decision probability of the current candidate decision node, which is the (i+1)th global decision probability. Further, based on this (i+1)th global decision probability, the (i+1)th level decision nodes that meet the expansion conditions are selected. In this way, the exponential expansion of the decision tree caused by low-probability branches is effectively avoided. At the same time, this disclosed solution can also generate candidate trajectories that match the preset interactive decisions for the retained (i+1)th level decision nodes. In this way, it ensures that each node not only has decision meaning, but also has trajectory results that can be used for subsequent recursion and final selection.

[0053] Furthermore, in a specific example, the local decision probability corresponding to each (i+1)th level candidate decision node can be obtained in the following manner; specifically, the determination of the local decision probability corresponding to each (i+1)th level candidate decision node based on the i-th level key obstacle can specifically include: Search for at least two preset interaction decisions between the target vehicle and the i-th level key obstacle (e.g., perform a preset number of searches) to obtain search statistics for each preset interaction decision; Based on the search statistics results corresponding to each preset interactive decision, the probability of the i-th local decision corresponding to each (i+1)-th level candidate decision node is determined.

[0054] Here, the search statistics results may specifically include at least one of the following: relative access probability, number of visits, cumulative revenue, average revenue, and decision confidence level, corresponding to each preset interactive decision.

[0055] For example, taking overtaking and yielding as two preset interaction decisions corresponding to the next level node of the i-th level node, and taking the local decision probability as the relative access probability as an example, the next level node of the i-th level node can be deduced N times: in each deduction process, the interaction decision between the target vehicle and the i-th level key obstacle is searched, and the next level node of the i-th level node (i.e., the i+1-th level candidate decision node) is expanded based on the searched interaction decision, and the interaction decision is used as the preset interaction decision of the next level node of the i-th level node; further, N times After the deduction results, the number of times the interaction decision between the target vehicle and the i-th level key obstacle is "overtake" (i.e., the number of visits) (which can be denoted as N_over) is obtained, which is also the search statistics of the i+1 level candidate decision nodes for "overtake"; and the number of times the interaction decision between the target vehicle and the i-th level key obstacle is "yield" (i.e., the number of visits) (which can be denoted as N_yield, where N_over+N_yield is less than or equal to N), which is also the search statistics of the i+1 level candidate decision nodes for "yield".

[0056] Furthermore, based on the search statistics N_over of the (i+1)th level candidate decision node corresponding to "overtake", the local decision probability of "overtaking" in all lower-level nodes of the i-th level node can be obtained, for example, N_over / N; furthermore, based on the search statistics N_yield of the (i+1)th level candidate decision node corresponding to "yield", the local decision probability of "yielding" in all lower-level nodes of the i-th level node can be obtained, for example, N_yield / N.

[0057] It should be noted that the above is only an illustrative example. In actual applications, the search statistics results can be specifically at least one of the following: the number of visits, cumulative revenue, average revenue, and decision confidence level corresponding to each preset interactive decision.

[0058] In this way, the disclosed solution can determine the local decision probability of each (i+1)th level candidate decision node through the search statistics results corresponding to each preset interactive decision, thereby making the probability source of each (i+1)th level candidate decision node interpretable, and can transform the statistical results (such as the number of visits) in the search process into the local probability required for subsequent global probability chain propagation. Thus, it provides favorable support for the cascaded decision recursion when autonomous vehicles face multiple obstacles.

[0059] Furthermore, in one example, the above-described method of determining the (i+1)th global decision probability of each (i+1)th level candidate decision node based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node can specifically include: The i-th global decision probability corresponding to the i-th level node is multiplied by the i-th local decision probability of the (i+1)-th level candidate decision node to obtain the (i+1)-th global decision probability of the (i+1)-th level candidate decision node.

[0060] It should be noted that the global decision probability corresponding to the i-th level node is obtained by multiplying the global decision probability corresponding to the parent node cascaded with the i-th level node by the local decision probability of the i-th level node. Here, for the root node, its global decision probability is the initial global decision probability, which can be set to 1.

[0061] Alternatively, in another example, the above-described method of determining the (i+1)th global decision probability of each (i+1)th level candidate decision node based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)th level candidate decision node can specifically include: The global decision probability of each node extended to the (i+1)th level candidate decision node is multiplied together with the (i)th local decision probability of the (i+1)th level candidate decision node to obtain the (i+1)th global decision probability of the (i+1)th level candidate decision node.

[0062] In other words, the global decision probability of the (i+1)th level candidate decision node can also be understood as the product of the global decision probabilities of all nodes in the path from the root node to the (i+1)th level candidate decision node, and then the product of the local decision probability of the (i+1)th level node.

[0063] In this way, the proposed solution achieves chain accumulation of decision path probabilities through a probability multiplication mechanism, so that the global decision probability at level i+1 carries the historical confidence of the entire path from the root node to the current node. This accurately quantifies the long-sequence interactive decision results, providing reliable data support for subsequent filtering of low-confidence branches, thereby effectively suppressing the exponential expansion of the decision tree from the source. Moreover, the multiplication logic is highly compatible with the cascaded recursive architecture, and no additional complex global evaluation function is required. The probability path optimization can be automatically completed during the recursive process, achieving both decision accuracy and decision efficiency.

[0064] Figure 2 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 2 This method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, or in-vehicle electronic devices. Alternatively, this method can be optionally applied to autonomous vehicles. It is understood that the above... Figure 1 The methods shown can also be applied to this example, and the related content will not be elaborated further in this example.

[0065] Furthermore, the method includes at least a portion of the following: For example... Figure 2 As shown, it includes: Step S201: Based on the current decision data of the target vehicle in the current decision cycle, determine the initial context information corresponding to the root node.

[0066] Here, in one example, the current decision data of the target vehicle in the current decision period may specifically include the state information of the target vehicle, the road environment information, and the state information of each obstacle among the multiple obstacles.

[0067] Furthermore, in one example, the target vehicle's state information may specifically include the target vehicle's current position, speed, acceleration, heading angle, curvature, vehicle size, etc. Alternatively, in another example, the road environment information may specifically include road geometry information, lane center lines, lane boundaries, road curvature, speed limit information, traffic rule information, etc.

[0068] Furthermore, the initial context information includes the target vehicle's initial linear trajectory (i.e., the 0th initial linear trajectory), the initial global decision probability, the state information of each obstacle, and the initial label of each obstacle.

[0069] For example, in one instance, the initial linear trajectory is estimated by combining the root node with the current decision data of the target vehicle within the current decision period.

[0070] Furthermore, in one example, the initial label of each obstacle can be used to indicate the processing status of each obstacle at the root node stage. For example, in one example, the initial label of an obstacle can be initialized to "unprocessed". Furthermore, in the recursive process, this initialized label can be updated at any time according to the actual processing situation. Moreover, in practical applications, the key obstacles processed at each level are not in a fixed order or are not fixed, but are dynamically determined based on the candidate trajectory of the current node and the label of the obstacle. In this way, cascaded recursive processing of multiple obstacles is realized, which significantly enhances the systematic processing capability of autonomous vehicles in the face of interactive coupling problems involving multiple obstacles, enabling rapid and reasonable decisions in complex scenarios such as overtaking, merging, and multi-vehicle game.

[0071] Furthermore, when the i-th level node is the root node (i.e., i=0), the i-th linear candidate trajectory is the initial linear trajectory for the target vehicle. And / or, the i-th global decision probability is the initial global decision probability.

[0072] In this way, the proposed solution can provide a unified data starting point for cascading decision recursion through the initial context information corresponding to the root node, so that subsequent nodes at all levels have a consistent recursive basis in terms of trajectory, probability, obstacle status and label.

[0073] Step S202: Execute the cascaded decision recursive steps, which include: Step S202-1: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node (i is an integer greater than or equal to 0) and the state information of each obstacle, determine the i-th level key obstacle corresponding to the i-th level node.

[0074] Here, the i-th level node is the root node or the i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle.

[0075] It should be noted that the relevant content regarding the i-th level node, the i-th context information, and the i-th level key obstacle can be referred to the above example, and will not be repeated here.

[0076] Step S202-2: Based on the key obstacles of level i, determine at least two candidate decision nodes of level i+1 for the level i node, and the local decision probability of each candidate decision node of level i+1.

[0077] Here, the i-th local decision probability represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the i+1-th level candidate decision node.

[0078] It should be noted that the relevant content regarding the (i+1)th level candidate decision node and the preset interaction strategy can be found in the example above, and will not be repeated here.

[0079] Step S202-3: Based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, determine the (i+1)-th global decision probability of each (i+1)-th level candidate decision node.

[0080] It should be noted that the relevant content regarding the global decision probability of the (i+1)th time can be found in the example above, and will not be repeated here.

[0081] Step S202-4: Determine whether the global decision probability of the (i+1)th level candidate decision node satisfies the preset probability condition; if yes, proceed to step S202-5; otherwise, proceed to step S202-7.

[0082] In one example, the preset probability condition can be specifically defined as the global decision probability being greater than or equal to a preset probability threshold. Here, the preset probability threshold is an empirical value; in practical applications, it can be determined based on actual computing resources, scenario complexity, etc. This disclosure does not impose specific restrictions on the specific value of the preset probability threshold.

[0083] Step S202-5: Determine the (i+1)th level candidate decision node as the next level node of the i-th level node to obtain the (i+1)th level decision node. Execute step S202-6.

[0084] In other words, in one example, if the global decision probability of the (i+1)th level candidate decision node meets the preset probability condition, the (i+1)th level candidate decision node still has a high decision probability or confidence level, and its branch can be retained. Then, the (i+1)th level candidate decision node can be used as the next level node of the (i)th level node and continue to expand downward.

[0085] Step S202-6: Based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, perform initial optimization on the i-th linear candidate trajectory of the i-th level node to obtain the (i+1)th linear candidate trajectory of the (i+1)th level decision node.

[0086] Step S202-7: Prune the (i+1)th candidate decision node and return to step S202-4 to process the other (i+1)th candidate decision node according to the above logic.

[0087] In other words, in one example, if the global decision probability of the (i+1)th level candidate decision node does not meet the preset probability condition, the decision probability or confidence of the (i+1)th level candidate decision node is low. In other words, when the target vehicle interacts with the i-th level obstacle, the probability of adopting the preset interaction decision corresponding to the (i+1)th level candidate decision node is low. At this time, the (i+1)th level candidate decision node can be pruned and no longer used as a node for subsequent recursive expansion. That is, the branch where the (i+1)th level candidate decision node is located stops expanding and no longer processes the next key obstacle.

[0088] Here, after traversing all (i+1)th candidate decision nodes, we return to step S202-1 to continue processing the next level node, such as the (i+1)th level decision node, and so on until the termination condition is met.

[0089] Step S203: Obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and obtain the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory.

[0090] Here, the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0091] It should be noted that the relevant content regarding the root node and candidate trajectory can be found in the example above, and will not be repeated here.

[0092] In this way, the proposed solution, based on the obtained global decision probabilities of each (i+1)th level candidate decision node and combined with preset probability conditions, selects the (i+1)th level candidate decision node with a high global decision probability as the (i+1)th level decision node, or prunes the (i+1)th level candidate decision node with a low global decision probability. This process preserves multiple reasonable interaction strategies while suppressing the computational bloat caused by low-probability branches. This significantly enhances the decision-making ability and interactivity of autonomous vehicles in complex road environments, enabling them to engage in game theory with key obstacles on the road. It allows autonomous vehicles to quickly make reasonable decisions in complex scenarios such as overtaking, merging, and multi-vehicle game theory, thereby improving traffic efficiency while ensuring traffic safety and passenger comfort. This allows autonomous vehicles to better adapt to complex road scenarios and cooperate with other vehicles.

[0093] Furthermore, the present invention utilizes the preset interactive decision corresponding to the (i+1)th level decision node to optimize the i-th linear candidate trajectory of the i-th level node, thereby obtaining the (i+1)-th linear candidate trajectory of the (i+1)-th level decision node. In this way, the obtained (i+1)-th linear candidate trajectory not only inherits the planning basis of the i-th linear candidate trajectory, but also reflects the preset interactive decision corresponding to the (i+1)-th level decision node, thereby significantly improving the interaction and decision-making capabilities of autonomous vehicles in the process of playing against obstacles.

[0094] Moreover, by pruning candidate decision nodes with low global probability, the present invention can reduce the number of branches in the recursive decision tree, reduce computational complexity, and improve the real-time solution capability within the current decision cycle, thereby significantly improving the ability of autonomous vehicles to make rapid decisions in complex scenarios with multiple obstacles.

[0095] Further, in a specific example, the initial candidate trajectory of the (i+1)th level decision node can be obtained in the following manner; specifically, the above-described optimization of the i-th linear candidate trajectory of the i-th level node based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle to obtain the (i+1)th linear candidate trajectory of the (i+1)th level decision node (e.g., step S202-6) can specifically include: Step S202-6-1: Based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, perform initial optimization on the i-th linear candidate trajectory of the i-th level node to obtain the (i+1)-th initial candidate trajectory.

[0096] In other words, in this example, the i+1 level decision node's preset interaction decision and the i-th level key obstacle can be combined to perform coarse-grained optimization of the i-th linear candidate trajectory of the i-th level node, that is, to perform single-agent interaction, thereby obtaining the coarse-grained i+1-th initial candidate trajectory.

[0097] Step S202-6-2: Determine the interaction constraints or interaction costs of the target vehicle relative to the i-th level key obstacle based on the preset interaction decision corresponding to the (i+1)-th level decision node.

[0098] It should be noted that, compared to single-agent interaction, the interaction constraints and costs are the result of multi-agent interaction, thus providing reliable data support for effectively improving the accuracy of trajectory optimization results.

[0099] Here, in one example, interaction constraints can be specifically used to define "conditions that cannot be violated," which are hard requirements. For example, in one example, interaction constraints can specifically transform a certain decision intention into constraints in the optimization problem. In other words, interaction constraints are used to limit the interaction relationships that the target vehicle and key obstacles need to satisfy, including spatial relationships, temporal relationships, and safety relationships, so that the optimized candidate trajectory meets the feasible conditions corresponding to the preset interaction decision.

[0100] For example, in one example, the interaction constraints of the target vehicle relative to the i-th level key obstacle may specifically include at least one of the following constraints: relative temporal constraints, relative position constraints, safety distance constraints, and conflict zone passage order constraints. In this way, the above constraints are used to ensure that the optimized candidate trajectory meets the feasible conditions corresponding to the preset interaction decision.

[0101] Furthermore, in another example, the interaction cost can be specifically used to evaluate "which solution is better," which is a soft optimization objective. In other words, the interaction cost is used to evaluate how well the target vehicle conforms to the preset interaction decision. By adjusting the optimization objective function, the optimized candidate trajectory is made to better match the driving intent corresponding to the preset interaction decision.

[0102] For example, in another example, the interaction cost of the target vehicle relative to the i-th level critical obstacle may specifically include at least one of the following: the cost of deviating from the preset interaction decision (e.g., the cost of deviating from the overtaking intention or the cost of deviating from the yielding intention), the cost of being too close to the critical obstacle, and the cost of unreasonable timing in passing through the conflict zone.

[0103] Step S202-6-3: Based on the interaction constraints or interaction costs, and the preset vehicle motion constraints, perform linear optimization on the (i+1)th initial candidate trajectory to obtain the (i+1)th linear candidate trajectory that matches the preset interaction decision.

[0104] Here, the preset vehicle motion constraints may specifically include speed constraints, acceleration constraints, jerk constraints, curvature constraints, rate of change of curvature constraints, vehicle dynamics constraints, or road boundary constraints, etc., which can be set based on actual scenario requirements. This disclosure does not impose specific restrictions on them.

[0105] In other words, in one example, the interaction constraints of the target vehicle relative to the i-th level key obstacle can be determined based on the preset interaction decision corresponding to the i-th level decision node. Then, based on the interaction constraints and the preset vehicle motion constraints, the coarse-grained i+1 initial candidate trajectory can be linearly optimized to obtain the i+1 linear candidate trajectory that matches the preset interaction decision.

[0106] Alternatively, in another example, based on the preset interaction decision corresponding to the i-th level decision node, the interaction cost of the target vehicle relative to the i-th level key obstacle is determined. Then, based on the interaction cost and the preset vehicle motion constraints, the coarse-grained i+1-th initial candidate trajectory is linearly optimized to obtain the i+1-th linear candidate trajectory that matches the preset interaction decision.

[0107] Alternatively, in another example, based on the preset interaction decision corresponding to the (i+1)th level decision node, the interaction cost and interaction constraints of the target vehicle relative to the (i)th level key obstacle are determined. Then, based on the interaction cost and interaction constraints, as well as the preset vehicle motion constraints, the coarse-grained (i+1)th initial candidate trajectory is linearly optimized to obtain the (i+1)th linear candidate trajectory that matches the preset interaction decision.

[0108] In other words, in one example, the disclosed solution can also employ a two-level trajectory optimization strategy. For instance, after identifying key obstacles and expanding to the next-level candidate decision nodes, coarse-grained trajectory optimization is first performed based on the preset interaction decisions corresponding to the expanded next-level candidate decision nodes. For example, taking the expansion from the i-th level node (e.g., the root node or decision node) to the (i+1)-th level decision node as an example, the i-th linear candidate trajectory of the i-th level node can be coarsely optimized based on the preset interaction decisions corresponding to the (i+1)-th level decision node, resulting in the (i+1)-th initial candidate trajectory of the (i+1)-th level decision node. This (i+1)-th initial candidate trajectory is a coarse-grained trajectory. Furthermore, to ensure the trajectory is smooth and that the identified key obstacles are more accurate and better meet the scenario requirements, nonlinear programming (NLP) can be used to finely re-optimize the obtained (i+1)-th initial candidate trajectory, thereby obtaining the (i+1)-th linear candidate trajectory. In this case, the optimized (i+1)-th linear candidate trajectory is smoother and easier to use directly for vehicle execution.

[0109] In this way, the present invention introduces the interactive decision results (such as the i-th linear candidate trajectory) into the lower-level trajectory optimization process, combining discrete interactive decisions (such as preset interactive strategies) with continuous trajectory optimization, avoiding the problem of inconsistency between interactive decision results and trajectory execution process, and improving the decision reliability and trajectory executability of the target vehicle in complex interactive scenarios.

[0110] It should be noted that the current linear candidate trajectory of each decision node in this disclosed scheme is not simply inherited from the initial linear trajectory, nor is it generated out of thin air. Rather, it is the trajectory optimization or trajectory update result after interacting with the key obstacle selected by the previous node according to the preset interactive decision corresponding to the decision node, based on the linear candidate trajectory of the previous node. The trajectory optimization or trajectory update result matches the preset interactive decision represented by the decision node.

[0111] For example, the current linear candidate trajectory of the first-level decision node is the optimized trajectory obtained after the target vehicle interacts with the key obstacles selected by the root node according to the preset interaction decisions corresponding to the first-level decision node, based on the initial linear trajectory of the root node. Similarly, the current linear candidate trajectory of the second-level decision node is the optimized trajectory obtained after the target vehicle continues to interact with the key obstacles selected by the first-level decision node according to the preset interaction decisions corresponding to the second-level decision node, based on the current linear candidate trajectory of the first-level decision node. This process continues, with the linear candidate trajectory of each level being updated layer by layer along the candidate decision path. Clearly, this approach is more reasonable and can effectively improve the decision reliability and trajectory executability of the target vehicle in complex interaction scenarios.

[0112] In this way, the proposed solution can utilize the interaction constraints or interaction costs of the target vehicle relative to the i-th level key obstacle to perform two-level optimization on the i-th linear candidate trajectory, thereby obtaining the i+1-th linear candidate trajectory that satisfies the preset interaction decision corresponding to the i+1-th level decision node. Thus, the i+1-th linear candidate trajectory obtained in the proposed solution not only inherits the planning basis of the i-th linear candidate trajectory, but also reflects the preset interaction decision corresponding to the i+1-th level decision node, effectively improving the consistency between the candidate trajectory and the interaction decision. At the same time, it significantly enhances the interaction and decision-making capabilities of autonomous vehicles in the process of playing with obstacles, providing favorable support for achieving the optimal balance between traffic safety, riding experience and traffic efficiency.

[0113] Furthermore, in a specific example, after obtaining the (i+1)th linear candidate trajectory of the (i+1)th level decision node, for example, after step S101-4 or step S202-6, the cascaded decision recursive step further includes: Based on the i-th context information, the i+1-th global decision probability of the i+1-th level decision node, the i+1-th linear candidate trajectory, and the preset interactive decision corresponding to the i+1-th level decision node, the i+1-th context information of the i+1-th level decision node is determined.

[0114] Here, the (i+1)th context information includes the (i+1)th linear candidate trajectory, the (i+1)th global decision probability, the state information of each obstacle, and the preset interaction decision corresponding to the (i+1)th level decision node. It should be noted that in practical applications, apart from the initial context information of the root node, the context information of all other decision nodes can specifically include the preset interaction decision corresponding to that decision node, thus providing data support for unified decision-making in subsequent autonomous driving.

[0115] Alternatively, the (i+1)th context information can further include the (i+1)th label of each obstacle. The (i+1)th label can be one of unprocessed, processed, or background object, thus providing a reference for subsequently identifying key obstacles.

[0116] In other words, when the (i+1)th level decision node becomes a new node in the cascaded decision recursive steps, its (i+1)th context information will also serve as the current context information of that node, so as to participate in the recursive processes such as determining the next key obstacle, generating candidate decision nodes, calculating local probabilities, propagating global probabilities, and updating candidate trajectories.

[0117] Furthermore, it should be noted that, considering the recursive processing logic of cascaded decision-making, the (i+1)th context information can inherit the information of the ith context information and is a further update of the ith context information. Based on this, the (i+1)th context information may also include scene information such as the scene where the target vehicle is located, as detailed above, and will not be repeated here.

[0118] In this way, the disclosed solution determines the i+1 context information of the i+1 level decision node based on the i+1 context information, the i+1 global decision probability of the i+1 level decision node, the i+1 linear candidate trajectory, and the preset interactive decision corresponding to the i+1 level decision node. This realizes the transmission of key information between parent and child nodes, avoids repeated calculations or loss of historical decision information in subsequent levels, and provides data support for the next decision iteration.

[0119] Furthermore, in a specific example, the i-th context information also includes the i-th label of each obstacle, which is one of unprocessed, processed, and background object.

[0120] Here, "unprocessed" means that the obstacle has not yet been used as a key obstacle in the current recursive path for interaction decision-making; "processed" means that the obstacle has been used as a key obstacle in the parent node for interaction decision-making; "background object" means that the obstacle is not used as the current interaction decision object, but can be used as an obstacle avoidance constraint, road constraint, or environmental constraint to participate in trajectory optimization.

[0121] At this point, the above-described method of determining the i-th level key obstacle corresponding to the i-th level node based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle (e.g., step S202-1 or step S101-1) can specifically include: Step S202-1-1: Based on the i-th linear candidate trajectory, the i-th label of each obstacle, and the state information of each obstacle, determine at least one obstacle from the plurality of obstacles whose label satisfies the interaction conditions and is not processed.

[0122] Step S202-1-2: Based on the fact that the label satisfying the interaction condition is at least one unprocessed obstacle, determine the i-th level critical obstacle.

[0123] Here, in a specific example, the interaction conditions may specifically include at least one of the following: there is a spatiotemporal conflict between the obstacle and the i-th linear candidate trajectory, the expected collision time between the obstacle and the target vehicle is less than a time threshold, and the obstacle is located within the planned passage area of ​​the target vehicle.

[0124] For example, in one example, obstacles labeled as unprocessed are first identified, then the earliest conflict time between the i-th linear candidate trajectory and the predicted trajectories of each unprocessed obstacle is calculated, and the obstacle with the earliest conflict time and labeled as unprocessed is identified as the i-th level critical obstacle.

[0125] It should be noted that in practical applications, interaction conditions can be set based on specific real-world scenario requirements, and this publicly available solution does not impose any specific restrictions on this.

[0126] Thus, this disclosed solution provides a specific method for determining the i-th level key obstacle. This method can quickly identify the obstacle that the current node most needs to interact with based on the interaction conditions and labels. This can improve the efficiency and consistency of cascade decision recursion, thereby providing favorable support for subsequent multi-vehicle games and determining the optimal trajectory.

[0127] Furthermore, in a specific example, after determining the (i+1)th level decision node that satisfies the expansion condition and obtaining the (i+1)th linear candidate trajectory of each (i+1)th level decision node, the cascaded decision recursive step further includes updating the label and state information of the (i)th level key obstacle. For example, in one example, the cascaded decision recursive step further specifically includes: Update the label of the i-th critical obstacle to "processed", and use the updated label as the i+1-th label; Determine the state response information of the i-th level key obstacle under the preset interactive decision corresponding to the (i+1)-th level decision node, and update the state information of the i-th level key obstacle based on the determined state response information of the i-th level key obstacle.

[0128] For example, when the target vehicle interacts with the i-th level critical obstacle, the preset interaction decision corresponding to the (i+1)-th level decision node can be adopted, such as the decision to surpass the i-th level critical obstacle. The response of the i-th level critical obstacle under the above preset interaction decision can be predicted, that is, the state response information can be predicted, such as trajectory information, occupied area information, behavior information, time series, etc. Then, based on the obtained state response information, the cached state information of the i-th level critical obstacle can be updated.

[0129] In one example, while updating the state information of the i-th level critical obstacle, this disclosed solution can also obtain the latest state information of other obstacles, so as to update the state information of other obstacles based on the determined latest state information. This provides data support for subsequently determining the next level of critical obstacles.

[0130] In this example, the determination of the (i+1)th context information of the (i+1)th level decision node based on the (i+1)th context information, the (i+1)th global decision probability of the (i+1)th level decision node, the (i+1)th linear candidate trajectory, and the preset interaction decision corresponding to the (i+1)th level decision node can specifically include: Based on the i-th context information, the i+1-th global decision probability and i+1-th linear candidate trajectory of the i+1-th level decision node, the preset interactive decision corresponding to the i+1-th level decision node, the updated i+1-th label of the i-th level key obstacle, and the updated state information of the i-th level key obstacle, the i+1-th context information is obtained.

[0131] For example, in one example, the (i+1)th context information includes the (i+1)th linear candidate trajectory, the (i+1)th global decision probability, the state information of each obstacle (including the updated state information of the i-th level key obstacle), the preset interactive decision corresponding to the (i+1)th level decision node, and the (i+1)th label of the updated i-th level key obstacle.

[0132] In other words, the state information of key obstacles does not remain in its initial state throughout the recursive process. For example, for obstacles that have already been processed, their new state information under the corresponding preset interaction decision can be used. Thus, in subsequent levels, although processed obstacles no longer participate in interaction decisions as new key obstacles, their updated state information can still be used as background constraints or obstacle avoidance constraints to participate in the target vehicle trajectory optimization.

[0133] In this way, by simultaneously updating the label and status information of the i-th level key obstacle, the interactive information (such as the status information of the obstacle) in the previous level node can be passed to the next level node, avoiding inconsistencies in the interactive assumptions caused by the subsequent decision still using the original status information of the processed obstacle. In other words, the interactive decision of each level is based on the interactive result of the previous level, thereby improving the consistency, interpretability and security of the cascaded interactive decision of multiple obstacles.

[0134] Figure 3 This is an illustrative flowchart of a trajectory determination method in an autonomous driving scenario according to an embodiment of this application. Figure 3 This method can be optionally applied to electronic devices, such as personal computers, servers, server clusters, or in-vehicle electronic devices. Alternatively, this method can be optionally applied to autonomous vehicles. It is understood that the above... Figure 1 and Figure 2 The methods shown can also be applied to this example, and the related content will not be elaborated further in this example.

[0135] Furthermore, the method includes at least a portion of the following: For example... Figure 3 As shown, it includes: Step S301: Execute the cascaded decision recursive steps.

[0136] Here, the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0137] It should be noted that the relevant content of the root node, candidate trajectory and initial context information can be referred to the above example, and will not be repeated here.

[0138] Furthermore, the above-mentioned cascaded decision recursive steps specifically include: Step S301-1: Determine the node that needs to be processed, for example, the i-th level node.

[0139] Step S301-2: Determine whether the i-th level node meets the termination condition; if the termination condition is met, proceed to step S301-3; otherwise, proceed to step S301-4.

[0140] Here, the termination condition may specifically include one of the following conditions: Condition 1: The i-th level key obstacle has not been identified based on the i-th level context information of the i-th level node; Condition 2: The i-th level node reaches the preset maximum recursion level, or the execution time of the cascaded decision recursive step reaches the preset time, or all the i+1 level candidate decision nodes corresponding to the i-th level node are pruned, or the i-th linear candidate trajectory of the i-th level node does not meet the trajectory feasibility condition.

[0141] Step S301-3: Determine the i-th level node as the decision node that satisfies the termination condition. Then return to step S301-1.

[0142] Step S301-4: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node (i is an integer greater than or equal to 0) and the state information of each obstacle, determine the i-th level key obstacle corresponding to the i-th level node. Proceed to step S301-5.

[0143] Here, the i-th level node is the root node or the i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle.

[0144] It should be noted that the relevant content regarding the i-th level node, the i-th context information, and the i-th level key obstacle can be referred to the above example, and will not be repeated here.

[0145] Step S301-5: Based on the key obstacles of level i, determine at least two candidate decision nodes of level i+1 for the level i node, and the local decision probability of each candidate decision node of level i+1. Proceed to step S301-6.

[0146] Here, the i-th local decision probability represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the i+1-th level candidate decision node.

[0147] It should be noted that the relevant content regarding the (i+1)th level candidate decision node and the preset interaction strategy can be found in the example above, and will not be repeated here.

[0148] Step S301-6: Based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, determine the (i+1)-th global decision probability of each (i+1)-th level candidate decision node. Proceed to step S301-7.

[0149] Step S301-7: Based on the global decision probability of each (i+1)th level candidate decision node, determine the (i+1)th level decision node that satisfies the expansion condition, and obtain the (i+1)th linear candidate trajectory of each (i+1)th level decision node.

[0150] It should be noted that the relevant content regarding extended conditions can be found in the above example, and will not be repeated here.

[0151] Step S302: After the cascaded decision recursive steps are completed, the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition is obtained, and the target optimized trajectory of the target vehicle is obtained based on at least one linear candidate trajectory.

[0152] In this way, by setting termination conditions, the proposed solution can obtain at least one decision node that meets the termination conditions, as well as the candidate trajectories of the target vehicle corresponding to each decision node. Then, based on the candidate trajectories, the target optimized trajectory of the target vehicle can be obtained. Thus, the recursive expansion of cascaded decisions can be stopped in a timely manner when there is no further interaction requirement, computational resources are limited, branches are pruned, or trajectories are infeasible. This ensures the closed-loop nature and real-time performance of the cascaded decision recursive steps, enabling autonomous vehicles to respond quickly in complex interaction scenarios and achieve the optimal balance between traffic safety, passenger experience, and traffic efficiency.

[0153] Furthermore, in a specific example, the target optimized trajectory of the target vehicle can be obtained in the following manner; specifically, the above-described method of obtaining the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory can specifically include: For the linear candidate trajectories corresponding to at least one decision node that meets the termination condition, trajectory feasibility screening is performed to obtain at least one feasible candidate trajectory. Based on the global decision probability of the decision nodes corresponding to each feasible candidate trajectory and the trajectory evaluation information of each feasible candidate trajectory, the comprehensive evaluation result of each feasible candidate trajectory is determined. Based on the comprehensive evaluation results, the target optimized trajectory is determined from the at least one feasible candidate trajectory.

[0154] It should be noted that, in one example, the trajectory feasibility screening may specifically include at least one of the following: solution state screening, collision detection screening, kinematic constraint screening, road boundary screening, and safety distance screening. For example, only candidate trajectories that have successfully optimized the trajectory may be retained. This disclosure does not impose specific restrictions on the specific screening strategy; in practical applications, it can be set based on actual needs.

[0155] Furthermore, in another example, the trajectory evaluation information may specifically include at least one of the following: safety evaluation information, smoothness evaluation information, traffic efficiency evaluation information, comfort evaluation information, rule compliance evaluation information, road reference line tracking evaluation information, and constraint violation degree. This disclosure does not impose specific limitations on the specific trajectory evaluation information; in practical applications, it can be set based on actual needs.

[0156] For example, in one example, multiple feasible candidate trajectories can be obtained from the linear candidate trajectories corresponding to at least one decision node that meets the termination condition. Then, the global decision probability of the decision node corresponding to the feasible candidate trajectory and the trajectory score of the feasible candidate trajectory are weighted and summed to obtain the comprehensive score of the feasible candidate trajectory. Then, based on the comprehensive score, the feasible candidate trajectory with the highest comprehensive score is selected as the target optimization trajectory of the target vehicle from the multiple feasible candidate trajectories.

[0157] It is understood that the target optimized trajectory obtained above is only an illustrative example. In actual applications, it can be set based on the actual driving scenario and driving conditions. This disclosed solution does not impose any specific restrictions on this.

[0158] In this way, the proposed solution determines the target vehicle's optimal trajectory from multiple candidate trajectories by integrating the global decision probability and trajectory evaluation information of the decision nodes. This achieves a balance between decision confidence and trajectory feasibility, ensuring that the final target optimal trajectory has good executability, safety, and decision consistency in multi-obstacle interaction scenarios. This provides favorable support for achieving the optimal balance between traffic safety, passenger experience, and traffic efficiency in complex interaction scenarios.

[0159] The following specific examples further illustrate the proposed solution. This proposed solution provides an interactive trajectory decision-making scheme based on joint optimization using Monte Carlo Tree Search (MCTS). When an autonomous vehicle faces multiple obstacles, this scheme constructs a recursive game decision tree and decomposes the complex multi-obstacle interaction problem into layer-by-layer two-vehicle game sub-problems in a "cascaded recursive" manner. This effectively controls the complexity of the decision-making process, thereby improving traffic efficiency while ensuring traffic safety and passenger comfort. It also enables autonomous vehicles to better adapt to complex traffic scenarios and cooperate with other vehicles.

[0160] Specifically, Figure 4 This is the tree-like decision-making framework of the disclosed solution; specifically, as Figure 4 As shown, the present invention constructs a recursive game-theoretic decision tree within this tree-like decision framework to organize and manage hierarchical decision-making processes in multi-obstacle interaction scenarios. Here, each node in the decision tree represents a specific interactive decision state, encapsulating complete environmental perception information, historical decision records, and pending decision tasks within that state.

[0161] Furthermore, in one example, the recursive game decision tree contains two types of nodes: root nodes and decision nodes, each with different functional responsibilities: (1) For the root node, as the starting point of the recursive game decision tree, it is responsible for the initialization of the entire cascading decision recursive process. Furthermore, the root node performs the following three core tasks, specifically including: (a) Scene data parsing and environment modeling: First, the scene data provided by the upstream perception and prediction module is parsed, and the vehicle configuration parameters of the target vehicle (such as vehicle size, dynamic constraints, etc.) are obtained; second, the geometric information model of the road is constructed, and key road features such as lane centerline, lane boundary, and road curvature are extracted; further, a reference speed curve is generated and used as a speed guide for trajectory optimization; finally, the background obstacle list is filtered and managed, and the signed distance field corresponding to the obstacles in the obstacle list is constructed for subsequent obstacle avoidance constraint calculation.

[0162] (b) Single-agent nonlinear programming optimization: After completing environment modeling, the root node performs a single-agent nonlinear programming optimization task. This task is guided by key road features (e.g., road reference lines) and reference speed curves (e.g., reference lines and reference speeds), with smoothness and safety as objectives, and outputs the initial optimized trajectory of the target vehicle (e.g., corresponding to the initial linear trajectory described above). Furthermore, the initial optimized trajectory can be extended to a sufficient time range to provide ample search space for subsequent obstacle interaction detection.

[0163] (c) Key obstacle identification and first round of MCTS game search: Based on the initial optimized trajectory of the target vehicle, key obstacles that may conflict with the target vehicle can be identified, and the first round of MCTS game search is performed for the first key obstacle to generate at least two candidate decision branches, such as overtaking and yielding, to drive the subsequent expansion of the decision tree.

[0164] For example, such as Figure 4 As shown, for the root node in the recursive game decision tree, after completing the scene data parsing and environment modeling, initial optimization is performed with reference lines and reference speeds as guides and smoothness and safety as objectives, to output the initial optimized trajectory of the target vehicle (corresponding to the initial linear trajectory mentioned above). Furthermore, based on the initial optimized trajectory of the target vehicle, the level 0 critical obstacles that have potential conflicts with the target vehicle can be identified, and the first round of MCTS game search is performed for the level 0 critical obstacles to generate two candidate decision branches: overtaking and yielding.

[0165] (2) For decision nodes, there are two types: overtaking and yielding. That is, the preset interactive decision corresponding to the decision node can be overtaking or yielding. Among them, a decision node with the preset interactive decision of overtaking indicates that the target vehicle adopts the overtaking decision for a certain key obstacle. Correspondingly, a decision node with the preset interactive decision of yielding indicates that the target vehicle adopts the yielding decision for a certain key obstacle.

[0166] Furthermore, each decision node performs the following two core tasks: (a) Scene re-optimization task: The trajectory output by the previous level node (also known as the parent node) (corresponding to the linear candidate trajectory mentioned above) is the processing object of the two-level optimization. At this time, the trajectory output by the previous level node can be initially optimized first, and then the initially optimized trajectory can be used as the initial guess. Through nonlinear programming, and under the determined interaction decision (such as overtaking or yielding), the initial guess is refined and re-optimized to output a smooth trajectory that satisfies the kinematic constraints.

[0167] (b) Next round of MCTS search task: Based on the re-optimization result (i.e., a smooth trajectory that satisfies kinematic constraints), identify the next key obstacle, execute a new round of MCTS game search, and generate a new decision branch, thereby driving the decision tree to continue to expand downward.

[0168] It should be noted that the recursive game decision tree in this disclosed scheme also supports a context information relationship mechanism. Specifically, any tree node (such as the root node or decision node) contains a complete context information container to store and manage the key data required for decision-making. In this way, through the standardized management of the context container, efficient transmission of key information between parent and child nodes is achieved, ensuring that each level of decision-making can accurately obtain historical decision information from the upper level, avoiding redundant calculations and information loss. Here, the context information may specifically include: (a) Trajectory results: The trajectory output by the NLP optimization of the current node (such as the (i+1)th linear candidate trajectory) contains a complete multidimensional state vector, which includes position, heading angle, curvature, velocity, acceleration and jerk. (b) Coarse-grained trajectory: The decision trajectory (such as the (i+1)th initial candidate trajectory) of the node corresponding to the node output by the MCTS search after the trajectory output by the previous node is optimized by the preset interaction decision. This is used as the initial guess for NLP optimization. (c) Signed distance field: Spatial distance field information of roads and obstacles, used for fast calculation of obstacle avoidance constraints; (d) Obstacle label management: Classification labels of obstacles (such as key interactive objects, already played objects, background objects, etc.) and their corresponding status information; (e) Scenario optimization results: Detailed results of the optimization problem solution, including convergence status, number of iterations, degree of constraint violation, etc.

[0169] Furthermore, such as Figure 4 As shown in Figures 5(a) and 5(b), the steps of the interactive trajectory decision-making scheme provided by this disclosure include: Step S501: Based on the current decision data of the target vehicle within the current decision period, determine the initial context information corresponding to the root node.

[0170] Here, the current decision data includes the target vehicle's status information (such as vehicle configuration parameters, driving status, etc.), road environment information, and the status information of each obstacle among multiple obstacles predicted by the obstacle prediction.

[0171] Furthermore, the initial context information corresponding to the root node may specifically include the target vehicle's initial linear trajectory, initial global decision probability, state information of each obstacle, and initial label of each obstacle.

[0172] Step S502: Perform cascaded decision recursion steps to expand to obtain at least one decision node that satisfies the termination condition.

[0173] Here, the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0174] Furthermore, taking the i-th level node as an example, the cascaded decision recursive steps include: Step S502-1: Determine the node that needs to be processed, for example, the i-th level node (i is an integer greater than or equal to 0).

[0175] Step S502-2: Determine whether the i-th level node meets the termination condition. If yes, proceed to step S502-3; otherwise, proceed to step S502-4.

[0176] Here, the termination condition in this disclosed solution may be that the key obstacle has not been identified (corresponding to condition one above), or that the MCTS search has failed (i.e., no effective decision can be found) or that the nonlinear programming problem has failed to be solved (i.e., the problem has no solution) (corresponding to condition two above). This disclosed solution does not impose specific restrictions on the specific extended conditions, which can be determined based on the actual needs of the scenario.

[0177] Step S502-3: Determine the i-th level node as the decision node that satisfies the termination condition. Then return to step S502-1.

[0178] Step S502-4: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node, the i-th label of each obstacle, and the state information of each obstacle, determine at least one obstacle from the plurality of obstacles that satisfies the interaction conditions and whose label is unprocessed.

[0179] Here, the i-th level node is the root node or the i-th level decision node cascaded with the root node.

[0180] Furthermore, the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, the state information of each obstacle, and the i-th label of each obstacle, wherein the i-th label is one of unprocessed, processed, or background object.

[0181] It should be noted that for the case where i is 0, that is, the i-th level node is the root node, please refer to the above explanation, which will not be repeated here.

[0182] Furthermore, it should be noted that the cascading decision-making recursive process of this disclosed solution supports a three-category obstacle classification mechanism. This mechanism divides obstacles in a multi-obstacle road scenario into three categories: (a) key interaction objects, i.e., the game objects at the current decision level, which are the direct interaction targets of MCTS search; (b) already played objects, i.e., obstacles that have completed the game in the previous level node, whose optimized trajectories can be used as background processing and will not participate in optimization again; (c) background objects, i.e., other obstacles, whose original predicted trajectories can be used, and obstacle avoidance constraints can be processed by constructing a signed distance field, without participating in interactive decision-making. In this way, through this classification mechanism, the effective transmission and accumulation of interactive information is achieved, ensuring that the decision at each level can utilize the game results of the previous level.

[0183] Step S502-5: Based on the fact that the label satisfying the interaction condition is at least one unprocessed obstacle, determine the i-th level key obstacle.

[0184] Here, the disclosed solution provides a key obstacle identification mechanism. This mechanism calculates the earliest conflict time between the linear trajectory of the target vehicle and the predicted trajectory of each obstacle, and selects the obstacle with the earliest conflict time that has not been processed as the key interaction object (i.e., key obstacle) at the current level. In this way, the mechanism can ensure that the most urgent interaction risks are dealt with first.

[0185] For example, in one example, obstacles labeled as unprocessed are first obtained from multiple obstacles. Then, through the obstacle decision-maker, the earliest conflict time between the target vehicle's i-th linear candidate trajectory at the i-th level node and the predicted trajectories of each obstacle labeled as unprocessed is calculated. The obstacles are sorted according to the urgency of the conflict, and then the obstacle with the earliest conflict time is selected as the key interaction object of the current level (i.e., the i-th level key obstacle).

[0186] Furthermore, a Stackelberg game model is constructed for key obstacles. In this model, the target vehicle is designated as the leader, possessing the initiative to act first; key obstacles are designated as followers, responding based on the target vehicle's interactive decisions; other obstacles are treated as background objects and do not participate in the game interaction. This asymmetric game model better reflects the interaction logic in real-world driving scenarios, where the target vehicle, as the active decision-maker, needs to consider the possible responses of key obstacles (such as key vehicles).

[0187] Step S502-6: Based on the key obstacles of level i, determine the two candidate decision nodes of level i+1 of level i.

[0188] For example, given the identification of a key obstacle at level i, a Monte Carlo tree game search task is performed at the level i node to explore a pre-defined interaction strategy for the key obstacle at level i. The Monte Carlo tree game search task can specifically include: (1) Node expansion: Select the overtake or yield action to create a new child node for the i-th level node (such as the i+1-th level candidate decision node). (2) Sampling: Randomly sample the response behavior of the i-th level key obstacle to simulate its driving state in the future preset time. For example, random sampling is performed based on the Stackelberg game model constructed above to simulate the response trajectory of the i-th level key obstacle in the future preset time. (3) Evaluation: Perform forward simulation (e.g., MCTS simulation) and calculate the cumulative benefit of the interaction trajectory between the target vehicle and the i-th level key obstacle according to the benefit function. Here, the benefit function can take into account dimensions such as safety, efficiency, and comfort, and can be set according to actual needs. (4) Backtracking: Propagate the cumulative revenue back along the search path and update the statistical information of all nodes (number of visits, cumulative revenue, etc., corresponding to the search statistics results mentioned above).

[0189] It should be noted that, in one example, the Monte Carlo tree game search task is repeated N times to obtain the number of visits for "overtaking" (denoted as N_over) and the number of visits for "yielding" (denoted as N_yield) for the child nodes of the i-th level node. At this time, at least two child nodes of the i-th level node are determined, and the number of visits for each child node is also determined.

[0190] Step S502-7: Determine the local decision probability of each (i+1)th level candidate decision node, and multiply the global decision probability of the (i)th level node and the local decision probability of the (i+1)th level candidate decision node to obtain the global decision probability of the (i+1)th level candidate decision node, and then obtain the global decision probability of each (i+1)th level candidate decision node.

[0191] It should be noted that the cascaded decision recursive process of this disclosed scheme supports a probabilistic chain propagation mechanism. This mechanism first calculates the local decision probability of "overtaking" at the (i+1)th level candidate decision node based on the number of visits for "overtaking" (N_over) and "yielding" (N_yield) output by the MCTS search: And the local decision probability of "yielding" in the child nodes of the current level node is calculated as follows: Furthermore, the global decision probability of the (i+1)th level candidate decision node (corresponding to the (i+1)th global decision probability mentioned above) = the global decision probability of the i-th level node (corresponding to the i-th global decision probability mentioned above) × the local decision probability of the (i+1)th level candidate decision node (corresponding to the i-th local decision probability mentioned above). In this way, the global decision probability of the (i+1)th level candidate decision node can be calculated, providing a basis for the creation and pruning of subsequent nodes.

[0192] In practical applications, we can also multiply all probabilities along the path from the root node to the child nodes of the current level node to obtain the global decision probability corresponding to the child node of the current level node. For example, the global decision probability of the (i+1)th level candidate decision node = the initial global decision probability corresponding to the root node × the global decision probability corresponding to the first level decision node × × Global decision probability corresponding to the i-th level decision node × Local decision probability of the (i+1)-th level candidate decision node.

[0193] Step S502-8: Determine whether the global decision probability of the (i+1)th level candidate decision node satisfies the preset probability condition. If yes, proceed to step S502-9; otherwise, proceed to step S502-11.

[0194] Step S502-9: Determine the (i+1)th level candidate decision node as the next level node of the ith level node to obtain the (i+1)th level decision node. Proceed to step S502-10.

[0195] Here, the proposed solution uses the global decision probability of the candidate decision node satisfying the preset probability condition as the node creation condition. That is, if the global decision probability of the (i+1)th level candidate decision node satisfies the preset probability condition, the (i+1)th level candidate decision node will be used as the next level node of the (i)th level node.

[0196] For example, in one instance, a preset probability threshold can be set based on empirical values. If the value exceeds this threshold, the preset probability condition is considered met; otherwise, it is not. The above is merely an illustrative example. In practical applications, the preset probability condition can be determined based on actual needs, and this disclosure does not impose specific restrictions on it.

[0197] Step S502-10: Based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, perform joint optimization on the i-th linear candidate trajectory of the i-th level node to obtain the (i+1)-th linear candidate trajectory of the (i+1)-th level decision node. Then return to step S502-1.

[0198] Specifically, firstly, based on the preset interaction decisions corresponding to the (i+1)th level decision node and the key obstacles at the (i)th level, the i-th linear candidate trajectory of the i-th level node is initially optimized to obtain the (i+1)-th initial candidate trajectory (i.e., generating a coarse-grained trajectory). Then, based on the preset interaction decisions corresponding to the (i+1)-th level decision node, using the (i+1)-th initial candidate trajectory as an initial guess, the initial guess is refined and re-optimized through nonlinear programming (NLP), combined with interaction constraints or interaction costs and preset vehicle motion constraints, to obtain the (i+1)-th linear candidate trajectory that satisfies kinematic constraints and has smooth executability. In other words, a multi-agent problem is constructed and jointly optimized to obtain a fine-grained trajectory that can be executed by the target vehicle.

[0199] For example, when the MCTS outputs a preset interactive decision for the i-th level obstacle, the linear trajectory output by the i-th level node, optimized by the preset interactive decision (i.e., the (i+1)-th initial candidate trajectory), is used as the initial guess. A nonlinear programming problem is then constructed and refined using the Interior Point Optimizer (IPOPT) solver. The specific process is as follows: (a) Initial value discretization: Discretize the coarse-grained trajectory (e.g., the i+1th initial candidate trajectory) into initial values ​​of variables for the NLP problem, including initial estimates of state variables such as position, velocity, and heading.

[0200] Here, the aforementioned state variables are seven-dimensional state vectors, which can specifically include: position, heading angle, curvature, velocity, acceleration, and jerk.

[0201] (b) Construction of the Nonlinear Programming Problem: First, a cost function for the nonlinear programming problem (corresponding to the interaction cost described above) is constructed. This cost function includes trajectory smoothness cost, reference line tracking cost, and obstacle avoidance cost. The trajectory smoothness cost penalizes the curvature and rate of change of jerk of the target vehicle in the optimized trajectory. The reference line tracking cost penalizes the lateral and velocity deviations between the optimized trajectory and the reference trajectory (e.g., a pre-planned, ideal geometric path and motion state sequence). The obstacle avoidance cost penalizes the collision risk between the target vehicle and obstacles in the optimized trajectory. Further, constraints for the nonlinear programming problem are set (corresponding to the interaction constraints described above). These constraints include forward Euler dynamics constraints, dynamic obstacle collision constraints, and static obstacle collision constraints. The forward Euler dynamics constraints ensure that the trajectory satisfies the vehicle kinematic equations; the dynamic obstacle collision constraints ensure a safe distance between the target vehicle and moving obstacles in the trajectory; and the static obstacle collision constraints ensure a safe distance between the target vehicle and road boundaries or fixed obstacles in the trajectory.

[0202] (c) IPOPT solution: Use the IPOPT solver to solve the above nonlinear programming problem.

[0203] Here, the proposed scheme uses the (i+1)th initial candidate trajectory as the initial guess and combines it with the preset interactive decision output by MCTS to achieve a hot start for nonlinear programming, thereby accelerating convergence. Furthermore, in practical applications, the maximum number of iterations (e.g., 1000 times) and the convergence tolerance (e.g., 1e-2) can be set to control the solution time and accuracy.

[0204] (d) Refined trajectory output: Output a smooth, executable trajectory that satisfies the kinematic constraints, which is the (i+1)th linear candidate trajectory.

[0205] Here, after determining the (i+1)th level decision node and its (i+1)th linear candidate trajectory, the label of the (i)th level key obstacle can be updated to "processed", and the updated label can be used as the (i+1)th label. Based on the state response information of the (i+1)th level key obstacle under the preset interactive decision corresponding to the (i+1)th level decision node, the state information of the (i)th level key obstacle can be updated.

[0206] Furthermore, the i+1 context information of the i-th level decision node can be constructed based on the i+1-th global decision probability and i+1-th linear candidate trajectory of the i+1-th level decision node, the preset interaction decision corresponding to the i+1-th level decision node, the updated i+1-th label of the i-th level key obstacle, and the updated state information of the i-th level key obstacle. This provides data support for the next cascaded decision recursive step.

[0207] Step S502-11: Prune the (i+1)th candidate decision node.

[0208] For example, in one example, if the global decision probability of the (i+1)th level candidate decision node is lower than a preset probability threshold (e.g., 0.1), then the (i+1)th level candidate decision node is pruned and the branch containing the (i+1)th level candidate decision node is no longer expanded, thereby effectively controlling the search space and avoiding exponential branch explosion.

[0209] It should be noted that, in one example, after obtaining the (i+1)th linear candidate trajectory of the (i+1)th decision node (such as its interaction decision corresponding to overtaking or yielding), the process returns to step S502-1, and then determines the (i+1)th level decision node as the node to be processed (i.e., the (i+1)th level node). At this time, it is determined whether the (i+1)th level node meets the termination condition. If the termination condition is met, the process proceeds to step S503. Otherwise, that is, if the termination condition is not met, the processing flow from step S502-4 to step S502-11 is executed. For example, a new key obstacle (i.e., the (i+1)th level key obstacle) is re-determined, and a new round of Monte Carlo tree search is performed for the (i+1)th level key obstacle to determine the next level (i.e., the (i+2)th level candidate decision node).

[0210] Step S503: Obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition.

[0211] Step S504: Perform trajectory feasibility screening on the linear candidate trajectories corresponding to at least one decision node that meets the termination condition to obtain feasible candidate trajectories.

[0212] Step S505: Determine the comprehensive evaluation result of the feasible candidate trajectory based on the global decision probability of the decision node corresponding to the feasible candidate trajectory and the trajectory evaluation information of the feasible candidate trajectory.

[0213] Step S506: Based on the comprehensive evaluation results, determine the target optimized trajectory from the feasible candidate trajectories.

[0214] In other words, after the recursive game decision tree is fully expanded, the trajectory selection module in this publicly available solution can select the optimal decision from all leaf nodes. For example, in one example, the trajectory selection module can employ a multi-stage screening and scoring mechanism, comprehensively considering factors such as the safety, feasibility, and probability of the trajectory, to output the optimal decision. Specifically, it includes the following four steps: Step 1: Depth-First Search (DFS) Candidate Collection Phase Here, in this stage, the decision tree can be traversed in a depth-first manner to extract linear candidate trajectories from each leaf node, and each linear candidate trajectory contains the following key information: (a) Node identifier: A unique identifier for the leaf node corresponding to the linear candidate trajectory; (b) Optimized trajectory mapping for each agent: optimized trajectories of the target vehicle and each interacting agent (i.e., the obstacles that need to be interacted with); (c) Path cumulative probability: the global decision probability from the root node to the current leaf node; (d) NLP solution status: Whether the NLP problem has successfully converged; (e) Score: The overall score of the linear candidate trajectories.

[0215] Step 2: Multi-stage Filtering Here, in this stage, a filter chain can be used to perform multi-level filtering on the extracted linear candidate trajectories, eliminating infeasible linear candidate trajectories. It should be noted that this filter chain supports flexible expansion and can be configured with multiple filters. The filter chain in this disclosed solution is configured with the following filters: (a) Solving the state filter: Only retaining linear candidate trajectories that have been successfully solved by NLP optimization; (b) Collision detection filter: Verifies the safe distance between the linear candidate trajectory and each obstacle (including dynamic and static obstacles); (c) Kinematic constraint filter: Verify whether the linear candidate trajectory satisfies the vehicle's kinematic constraints.

[0216] Thus, through the various filters described above, linear candidate trajectories that meet the filtering requirements (e.g., the feasible candidate trajectories mentioned above) can be obtained and entered into the scoring stage to ensure that the finally selected trajectory is executable.

[0217] Step 3: Multi-dimensional scoring stage Here, in this stage, a weighted summation scorer can be used to implement a multi-dimensional scoring strategy, and the scoring dimensions used by the multi-dimensional scoring strategy include: (a) Path cumulative probability score: The decision confidence of the linear candidate trajectory is evaluated based on the global decision probability corresponding to the decision stage corresponding to the linear candidate trajectory; (b) Trajectory smoothness score: Based on the curvature and jerk in the linear candidate trajectory, evaluate the ride comfort of the target vehicle when performing the linear candidate trajectory; (c) Efficiency score: Based on the travel time and distance of the linear candidate trajectory, evaluate the traffic efficiency of the target vehicle when executing the linear candidate trajectory; (d) Safety score: The safety of the target vehicle is evaluated based on the minimum distance between the target vehicle and the obstacle when executing a linear candidate trajectory.

[0218] Furthermore, the four scores of the linear candidate trajectory and the weights of each score dimension are weighted to obtain the total score of the linear candidate trajectory.

[0219] In practical applications, the scoring dimensions and their weights used in the multi-dimensional scoring strategy can be flexibly configured according to the actual needs of the scenario. This public solution does not impose specific restrictions on this.

[0220] Step 4: Output stage.

[0221] Here, in this stage, based on the total score of each linear candidate trajectory obtained in step 3, the linear candidate trajectory with the highest total score is determined from multiple linear candidate trajectories as the final decision trajectory. At the same time as outputting the final decision trajectory, the key information carried by the final decision trajectory is also output: the predicted trajectory of each interactive agent (i.e. the trajectory after interaction), the decision path sequence for each interactive agent (including overtaking or yielding), and the decision confidence of the decision path sequence (i.e., the global decision probability).

[0222] Furthermore, the aforementioned final decision trajectory can be directly sent to the target vehicle to execute the final decision trajectory, or the target vehicle can transmit the final decision trajectory to the upstream module for visualization of the decision results, or offline scenarios can be generated based on the aforementioned final decision trajectory and the context information of the node it is located in.

[0223] It should be noted that this disclosed solution can encapsulate the core functions of the aforementioned tree-like decision-making framework, making its core processing logic clearer, easier to understand and maintain, and enabling users to call various functions in the framework through custom information interfaces, thereby reducing the possibility of errors.

[0224] Furthermore, in practical applications, users can adjust the MCTS search parameters (such as the maximum number of simulations, exploration constant, preset probability threshold, etc.), NLP optimization parameters (such as the maximum number of iterations, convergence tolerance), and cost weights (such as the weights corresponding to trajectory smoothness cost, reference line tracking cost, and obstacle avoidance cost smoothing) according to their own needs. This disclosed solution does not impose specific restrictions on these aspects.

[0225] In this way, the present solution provides real-time data information through a custom information interface, enabling users to understand the operation of the tree-structured decision-making framework in a timely manner. This allows for targeted adjustments and optimizations when problems arise, thereby continuously improving the system's performance and effectiveness.

[0226] It should also be noted that this disclosed solution has high flexibility and scalability. For example, it can be deployed in online real-time decision-making mode and offline scenario extension mode according to actual application needs, so as to meet the development and deployment needs at different stages.

[0227] (a) Online real-time decision-making mode Here, the online real-time decision-making mode is the core application scenario of this disclosed solution, aiming to realize real-time interactive decision-making of autonomous vehicles in real traffic environments. In this application scenario, this disclosed solution, through the introduction of a series of accelerated solution techniques, can meet the real-time requirements of autonomous driving systems for decision-making and planning, and is suitable for handling complex scenarios including overtaking, merging, and multi-vehicle interactive game.

[0228] (1) Real-time solution acceleration technology To meet the real-time requirements of online operation, this disclosed solution integrates several solution acceleration technologies: (1) a probability-driven pruning strategy, which can terminate the expansion when the global decision probability is lower than a preset probability threshold, thus avoiding invalid search; (2) a hot-start optimization strategy, which uses the initial optimization result of the trajectory output by the parent node as the initial guess of the child node to accelerate NLP convergence; (3) a parallel MCTS simulation strategy, which can execute the Monte Carlo tree simulation task of MCTS in parallel using multiple threads, such as two decision nodes at the same level performing the Monte Carlo tree simulation task of MCTS in parallel; (4) an early stopping mechanism, which can terminate the search in advance when the global decision probability is significantly better than other branches, thus avoiding unnecessary waste of computing resources.

[0229] (2) Typical online application scenarios Scenario 1: In ramp merging or lane reduction scenarios, the target vehicle needs to coordinate its entry timing with other vehicles on the main road. In this scenario, the proposed solution first explores various merging strategies (such as preemptive entry, waiting for yield, and coordinated entry) through MCTS, and evaluates the safe interaction space with other vehicles under different merging strategies. Then, through "two-level optimization," it optimizes and generates an executable trajectory that meets the merging time window constraints, ensuring safe entry during gaps between other vehicles.

[0230] Scenario 2: When the target vehicle faces multiple key obstacles simultaneously (e.g., yielding to a fast vehicle on the right before overtaking a slow vehicle in front). In this scenario, as shown in Figure 6(a), the proposed solution decomposes the complex multi-vehicle interaction problem into layer-by-layer two-vehicle game subproblems through a cascaded decision-making recursive mechanism for multiple obstacles. Moreover, the decision tree in the proposed solution processes each obstacle sequentially according to the urgency of the interaction between the target vehicle and each obstacle. Each decision node at each level will output the optimal strategy for overtaking or yielding. At the same time, a probability propagation mechanism is used to ensure the consistency and optimality of the global decisions from the root node to the decision nodes in the decision tree, ultimately outputting a complete trajectory scheme that coordinates all interactive obstacles.

[0231] (ii) Offline Scene Extension Mode Here, the offline scene extension mode is another innovative application of this disclosed solution. This mode uses the MCTS-NLP joint optimization framework as a tool for scene generation and extension, and applies it to the offline development, testing, and verification phases of autonomous driving systems. In this way, the mode fully leverages the global exploration capabilities of MCTS to generate diverse interactive scenarios and trajectory samples, providing rich data support for the training and testing of autonomous driving systems.

[0232] (1) Functional positioning of scene expander In the development process of autonomous driving systems, high-quality scene data is crucial for algorithm development and testing. However, collecting real-world road data is costly and struggles to cover various extreme and edge scenarios. This publicly available solution serves as an offline scene extender, capable of automatically generating a large number of derived scenarios and diverse trajectories based on a small number of complex real-world cases, thus effectively addressing the scarcity of scene data.

[0233] Furthermore, the core functions of the aforementioned offline scene extender include: (a) Target vehicle behavior generation: Generate various driving behavior trajectories such as overtaking, yielding, and following, covering diverse driving strategies; (b) Obstacle behavior generation: Generate the response behavior trajectory of obstacles (such as acceleration, deceleration, etc.) to simulate the uncertainty of real interaction with the target vehicle; (c) Decision branch extraction: Extract the trajectory of each branch in the decision tree and retain all candidate trajectory schemes to support the comparative analysis of multiple trajectory schemes.

[0234] (2) Offline scene extension process Here, as shown in Figure 6(b), the offline scene extension process specifically includes: Step 1: Input one or more complex seed scenarios, which are usually derived from typical interaction cases in real road data (such as real accident scenarios, dangerous scenarios, etc.).

[0235] Step 2: Perform the complete MCTS-NLP decision (i.e. joint optimization algorithm) process on the seed scene described above, and configure it to "full branch retention mode", that is, do not perform probabilistic pruning, retain all branches of the decision tree, and output all possible trajectories between the target vehicle and obstacles.

[0236] Step 3: Automated annotation of the generated trajectory, including: decision type labels (such as overtaking / yielding / following, etc.), safety assessment (such as collision risk, safety margin, etc.), efficiency assessment (such as travel time, distance, etc.), and comfort assessment (such as jerk, lateral acceleration, etc.); then, based on the above extended scenarios, abnormal behaviors (such as heavy braking, collision, etc.) are removed to obtain the final traffic scenario model.

[0237] (3) Typical offline application scenarios This publicly available solution serves as an offline scene extender, enabling the extraction of specific types of trajectories from a complete decision tree to meet diverse testing needs. (a) Overtaking trajectory set: Extract the trajectories with the decision type "overtaking" from all leaf nodes to test the system's aggressive driving capability; (b) Yield Trajectory Set: Extract the trajectories with the decision type "yield" from all leaf nodes to test the conservative safety strategy of the system; (c) Hybrid decision trajectory set: Extract complete trajectories containing multiple types of decisions to test the system's comprehensive decision-making capabilities.

[0238] In summary, this disclosed solution has the following significant advantages compared to existing solutions: (1) Enhance the decision-making ability of multi-agent interaction Most existing autonomous driving planning solutions adopt single-agent planning methods (such as rule-based state machines, simple optimization methods, or end-to-end learning methods). When facing multi-vehicle interaction scenarios, other vehicles are usually treated as static or obstacles moving along predicted trajectories, ignoring the game-theoretic interaction between vehicles. However, the proposed solution uses the Stackelberg game model to explicitly model the master-slave game relationship between vehicles, thereby possessing true multi-agent interactive reasoning capabilities. In this way, it can predict the response behavior of other vehicles to the target vehicle's decisions and make more reasonable decisions in complex scenarios such as overtaking, merging, and multi-vehicle game, thereby significantly improving the throughput and decision-making quality in dense traffic flow.

[0239] (2) Scalable processing capability for multi-obstacle scenarios In scenarios involving multi-obstacle interactions, existing solutions typically employ exhaustive combination strategies or heuristic simplification strategies. The former's computational complexity increases exponentially with the number of obstacles, while the latter sacrifices decision quality. The cascaded recursive game mechanism proposed in this disclosure transforms the multi-obstacle interaction problem into layered two-car game subproblems through hierarchical decomposition of a decision tree. Combined with probabilistic chain propagation and pruning strategies, it effectively controls decision complexity. Thus, this mechanism can efficiently handle simultaneous interactions of three or more key obstacles while only increasing computational resources linearly, significantly reducing computational resource consumption.

[0240] (3) Iterative advantages brought by modular architecture Existing solutions often employ tightly coupled architectures, making it difficult to replace or upgrade their core algorithms independently. In contrast, this publicly available solution offers rich user-defined interfaces, and key components such as node tasks, filters, and scorers all support independent replacement and expansion. This highly modular architecture design enables rapid response to diverse needs and application scenarios, significantly shortening product iteration cycles, reducing custom development costs, and thus creating a sustainable competitive advantage at the product level.

[0241] (4) The unique value of offline scene expansion Existing solutions typically focus only on online decision-making functionality, while this disclosed solution not only supports online decision-making but also offline scenario expansion capabilities. The offline scenario expansion capability of this disclosed solution allows for the automatic generation of a large amount of diverse interaction trajectory data based on a small number of real-world seed scenarios. In this way, the massive amounts of generated labeled data are used to train and validate other algorithm modules, continuously improving the overall system performance and thus forming a data barrier and technological moat.

[0242] This disclosure provides a trajectory determination device for autonomous driving scenarios, such as... Figure 7 As shown, it includes: The recursive processing unit 701 is used to execute cascaded decision recursive steps, wherein the cascaded decision recursive steps include: determining the i-th level key obstacle corresponding to the i-th level node based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle; wherein the i-th level node is the root node or an i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle; and determining at least two (i+1)-th level candidate decision nodes of the i-th level node, and each (i+1)-th level candidate obstacle, based on the i-th level key obstacle. The local decision probability of the decision node; where the local decision probability of the i-th local decision probability represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the (i+1)-th level candidate decision node; based on the global decision probability corresponding to the i-th level node and the local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th global decision probability of each (i+1)-th level candidate decision node is determined; based on the (i+1)-th global decision probability of each (i+1)-th level candidate decision node, the (i+1)-th level decision node that meets the expansion condition is determined, and the (i+1)-th linear candidate trajectory of each (i+1)-th level decision node is obtained; The trajectory output unit 702 is used to obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and to obtain the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory; the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

[0243] In a specific example of the disclosed solution, the recursive processing unit is specifically used for: If the global decision probability of the (i+1)th level candidate decision node satisfies the preset probability condition, the (i+1)th level candidate decision node is determined as the next level node of the i-th level node to obtain the (i+1)th level decision node; based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, the i-th linear candidate trajectory of the i-th level node is optimized to obtain the (i+1)th linear candidate trajectory of the (i+1)th level decision node. or, If the global decision probability of the (i+1)th candidate decision node does not meet the preset probability condition, the (i+1)th candidate decision node is pruned.

[0244] In a specific example of the disclosed solution, the recursive processing unit is specifically used for: Based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, the i-th linear candidate trajectory of the i-th level node is initially optimized to obtain the (i+1)-th initial candidate trajectory. Based on the preset interaction decision corresponding to the (i+1)th level decision node, determine the interaction constraints or interaction costs of the target vehicle relative to the i-th level key obstacle; Based on interaction constraints or interaction costs and preset vehicle motion constraints, linear optimization is performed on the (i+1)th initial candidate trajectory to obtain the (i+1)th linear candidate trajectory that matches the preset interaction decision.

[0245] In a specific example of the scheme disclosed herein, the recursive processing unit is further configured to: Based on the current decision data of the target vehicle within the current decision cycle, the initial context information corresponding to the root node is determined; wherein, the current decision data includes the state information of the target vehicle, road environment information, and the state information of each obstacle among the multiple obstacles; the initial context information includes the initial linear trajectory of the target vehicle, the initial global decision probability, the state information of each obstacle, and the initial label of each obstacle; Wherein, when the i-th level node is the root node, the i-th linear candidate trajectory is the initial linear trajectory, and / or, the i-th global decision probability is the initial global decision probability.

[0246] In a specific example of the scheme disclosed herein, the recursive processing unit is further configured to: After obtaining the (i+1)th linear candidate trajectory of the (i+1)th level decision node, the (i+1)th context information of the (i+1)th level decision node is determined based on the (i+1)th context information, the (i+1)th global decision probability of the (i+1)th level decision node, the (i+1)th linear candidate trajectory, and the preset interaction decision corresponding to the (i+1)th level decision node. The (i+1)th context information includes the (i+1)th linear candidate trajectory, the (i+1)th global decision probability, the state information of each obstacle, and the preset interaction decision corresponding to the (i+1)th level decision node.

[0247] In a specific example of the disclosed scheme, the i-th context information also includes the i-th label of each obstacle, wherein the i-th label is one of unprocessed, processed, and background object; Specifically, the recursive processing unit is used for: Based on the i-th linear candidate trajectory, the i-th label of each obstacle, and the state information of each obstacle, at least one obstacle that has not been processed is identified from the plurality of obstacles as having a label that satisfies the interaction conditions. Based on the fact that the label satisfying the interaction conditions is at least one unprocessed obstacle, the i-th level critical obstacle is determined.

[0248] In a specific example of the scheme disclosed herein, the recursive processing unit is further configured to: Update the label of the i-th critical obstacle to "processed", and use the updated label as the i+1-th label; Determine the state response information of the i-th level key obstacle under the preset interactive decision corresponding to the i+1-th level decision node, and update the state information of the i-th level key obstacle based on the determined state response information of the i-th level key obstacle. Based on the i-th context information, the i+1-th global decision probability and i+1-th linear candidate trajectory of the i+1-th level decision node, the preset interactive decision corresponding to the i+1-th level decision node, the updated i+1-th label of the i-th level key obstacle, and the updated state information of the i-th level key obstacle, the i+1-th context information is obtained.

[0249] In a specific example of the scheme disclosed herein, the trajectory output unit is specifically used for: If the i-th level key obstacle is not identified based on the i-th level context information of the i-th level node, the i-th level node is identified as a decision node that meets the termination condition. or, If the i-th level node reaches the preset maximum recursion level, or the execution time of the cascaded decision recursion step reaches the preset time, or all the i+1 level candidate decision nodes corresponding to the i-th level node are pruned, or the i-th linear candidate trajectory of the i-th level node does not meet the trajectory feasibility condition, the i-th level node is determined as a decision node that meets the termination condition.

[0250] In a specific example of the scheme disclosed herein, the trajectory output unit is specifically used for: For the linear candidate trajectories corresponding to at least one decision node that meets the termination condition, trajectory feasibility screening is performed to obtain at least one feasible candidate trajectory. Based on the global decision probability of the decision nodes corresponding to each feasible candidate trajectory and the trajectory evaluation information of each feasible candidate trajectory, the comprehensive evaluation result of each feasible candidate trajectory is determined. Based on the comprehensive evaluation results, the target optimized trajectory is determined from the at least one feasible candidate trajectory.

[0251] In a specific example of the disclosed solution, the recursive processing unit is specifically used for: Search for at least two preset interaction decisions between the target vehicle and the i-th level key obstacle, and obtain the search statistics results corresponding to each preset interaction decision; Based on the search statistics results corresponding to each preset interactive decision, the probability of the i-th local decision corresponding to each (i+1)-th level candidate decision node is determined.

[0252] In a specific example of the disclosed solution, the recursive processing unit is specifically used for: The i-th global decision probability corresponding to the i-th level node is multiplied by the i-th local decision probability of the (i+1)-th level candidate decision node to obtain the (i+1)-th global decision probability of the (i+1)-th level candidate decision node. or, The global decision probability of each node extended to the (i+1)th level candidate decision node is multiplied by the i-th local decision probability of the (i+1)th level candidate decision node to obtain the (i+1)th global decision probability of the (i+1)th level candidate decision node.

[0253] For a description of the specific functions and examples of each unit of the apparatus in this disclosure embodiment, please refer to the relevant descriptions of the corresponding steps in the above method embodiments, which will not be repeated here.

[0254] The acquisition, storage, and application of user personal information involved in the technical solution disclosed herein comply with the provisions of relevant laws and regulations and do not violate public order and good morals.

[0255] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, a computer program product, and an autonomous vehicle.

[0256] Figure 8 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0257] like Figure 8 As shown, device 800 includes a computing unit 801, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 802 or a computer program loaded from storage unit 808 into random access memory (RAM) 803. RAM 803 may also store various programs and data required for the operation of device 800. The computing unit 801, ROM 802, and RAM 803 are interconnected via bus 804. Input / output (I / O) interface 805 is also connected to bus 804.

[0258] Multiple components in device 800 are connected to I / O interface 805, including: input unit 806, such as keyboard, mouse, etc.; output unit 807, such as various types of monitors, speakers, etc.; storage unit 808, such as disk, optical disk, etc.; and communication unit 809, such as network card, modem, wireless transceiver, etc. Communication unit 809 allows device 800 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0259] The computing unit 801 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as trajectory determination methods in autonomous driving scenarios. For example, in some embodiments, the trajectory determination method in an autonomous driving scenario can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on device 800 via ROM 802 and / or communication unit 809. When the computer program is loaded into RAM 803 and executed by the computing unit 801, one or more steps of the trajectory determination method in an autonomous driving scenario described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform trajectory determination methods in an autonomous driving scenario by any other suitable means (e.g., by means of firmware).

[0260] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0261] The program code used to implement the methods of this disclosure may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing apparatus, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0262] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0263] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0264] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0265] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0266] It should be understood that the various forms of processes shown above can be used to rearrange, add, or delete steps. For example, the steps described in this disclosure can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this disclosure can be achieved, and this is not limited herein.

[0267] Furthermore, the autonomous driving vehicle provided in this disclosure includes a vehicle body and any of the electronic devices described in the above embodiments. The autonomous driving vehicle may also include a power system, a steering system, a braking system, a perception system, etc. The electronic devices may be located in the vehicle's central controller or independently within the vehicle's autonomous driving control unit, and are used to execute one or more steps of the trajectory determination method in the autonomous driving scenario described above. The electronic devices may be connected to one or more components such as cameras, lidar, millimeter-wave radar, inertial navigation devices, global positioning systems, and vehicle actuators to receive vehicle operation data and / or environmental perception data.

[0268] In some implementations, the electronic device can generate control commands and send the control commands to the vehicle's steering system, braking system, and power system to control the vehicle to perform one or more automated driving operations such as path planning, obstacle avoidance, automatic following, automatic lane changing, and automatic parking.

[0269] Those skilled in the art should understand that the electronic device can be implemented in hardware, software, or a combination of both. Its specific structure and working principle can be found in the relevant descriptions of the above embodiments, and will not be repeated here.

[0270] The specific embodiments described above do not constitute a limitation on the scope of protection of this disclosure. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this disclosure should be included within the scope of protection of this disclosure.

Claims

1. A trajectory determination method in an autonomous driving scenario, comprising: Execute the cascaded decision recursive steps to obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and obtain the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory. Decision nodes are nodes cascaded from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node; the cascaded decision recursive steps include: Based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle, the i-th level key obstacle corresponding to the i-th level node is determined; wherein, the i-th level node is the root node or the i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability and the state information of each obstacle; Based on the i-th level key obstacle, at least two i+1 level candidate decision nodes for the i-th level node are determined, as well as the i-th local decision probability of each i+1 level candidate decision node; where the i-th local decision probability represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the i+1 level candidate decision node. Based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th global decision probability of each (i+1)-th level candidate decision node is determined. Based on the (i+1)th global decision probability of each (i+1)th level candidate decision node, the (i+1)th level decision node that satisfies the expansion condition is determined, and the (i+1)th linear candidate trajectory of each (i+1)th level decision node is obtained.

2. The method according to claim 1, wherein, The process of determining the (i+1)th level decision nodes that satisfy the expansion conditions based on the (i+1)th level global decision probability of each (i+1)th level candidate decision node, and obtaining the (i+1)th level linear candidate trajectory of each (i+1)th level decision node, includes: If the global decision probability of the (i+1)th level candidate decision node satisfies the preset probability condition, the (i+1)th level candidate decision node is determined as the next level node of the i-th level node to obtain the (i+1)th level decision node; based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, the i-th linear candidate trajectory of the i-th level node is optimized to obtain the (i+1)th linear candidate trajectory of the (i+1)th level decision node. or, If the global decision probability of the (i+1)th candidate decision node does not meet the preset probability condition, the (i+1)th candidate decision node is pruned.

3. The method according to claim 2, wherein, The optimization of the i-th linear candidate trajectory of the i-th level node based on the preset interactive decision corresponding to the i+1-th level decision node and the i-th level key obstacle, to obtain the i+1-th linear candidate trajectory of the i+1-th level decision node, includes: Based on the preset interactive decision corresponding to the (i+1)th level decision node and the i-th level key obstacle, the i-th linear candidate trajectory of the i-th level node is initially optimized to obtain the (i+1)-th initial candidate trajectory. Based on the preset interaction decision corresponding to the (i+1)th level decision node, determine the interaction constraints or interaction costs of the target vehicle relative to the i-th level key obstacle; Based on interaction constraints or interaction costs and preset vehicle motion constraints, linear optimization is performed on the (i+1)th initial candidate trajectory to obtain the (i+1)th linear candidate trajectory that matches the preset interaction decision.

4. The method according to claim 1, 2, or 3, further comprising, prior to the cascaded decision recursion step: Based on the current decision data of the target vehicle within the current decision cycle, the initial context information corresponding to the root node is determined; wherein, the current decision data includes the state information of the target vehicle, road environment information, and the state information of each obstacle among the multiple obstacles; the initial context information includes the initial linear trajectory of the target vehicle, the initial global decision probability, the state information of each obstacle, and the initial label of each obstacle; Wherein, when the i-th level node is the root node, the i-th linear candidate trajectory is the initial linear trajectory, and / or, the i-th global decision probability is the initial global decision probability.

5. The method according to claim 4, after obtaining the (i+1)th linear candidate trajectory of the (i+1)th level decision node, the cascaded decision recursive step further includes: Based on the i-th context information, the i+1-th global decision probability of the i+1-th level decision node, the i+1-th linear candidate trajectory, and the preset interaction decision corresponding to the i+1-th level decision node, the i+1-th context information of the i+1-th level decision node is determined; wherein, the i+1-th context information includes the i+1-th linear candidate trajectory, the i+1-th global decision probability, the state information of each obstacle, and the preset interaction decision corresponding to the i+1-th level decision node.

6. The method according to claim 5, wherein, The i-th context information also includes the i-th label of each obstacle, where the i-th label is one of unprocessed, processed, or background object; The step of determining the i-th level key obstacle corresponding to the i-th level node based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle includes: Based on the i-th linear candidate trajectory, the i-th label of each obstacle, and the state information of each obstacle, at least one obstacle that has not been processed is identified from the plurality of obstacles as having a label that satisfies the interaction conditions. Based on the fact that the label satisfying the interaction conditions is at least one unprocessed obstacle, the i-th level critical obstacle is determined.

7. The method according to claim 6, wherein, The cascaded decision recursive steps also include: Update the label of the i-th critical obstacle to "processed", and use the updated label as the i+1-th label; Determine the state response information of the i-th level key obstacle under the preset interactive decision corresponding to the i+1-th level decision node, and update the state information of the i-th level key obstacle based on the determined state response information of the i-th level key obstacle. The step of determining the (i+1)th context information of the (i+1)th level decision node based on the (i+1)th context information, the (i+1)th global decision probability of the (i+1)th level decision node, the (i+1)th linear candidate trajectory, and the preset interactive decision corresponding to the (i+1)th level decision node includes: Based on the i-th context information, the i+1-th global decision probability and i+1-th linear candidate trajectory of the i+1-th level decision node, the preset interactive decision corresponding to the i+1-th level decision node, the updated i+1-th label of the i-th level key obstacle, and the updated state information of the i-th level key obstacle, the i+1-th context information is obtained.

8. The method according to any one of claims 1 to 7, wherein, At least one decision node that meets the termination condition is determined in the following way: If the i-th level key obstacle is not identified based on the i-th level context information of the i-th level node, the i-th level node is identified as a decision node that meets the termination condition. or, If the i-th level node reaches the preset maximum recursion level, or the execution time of the cascaded decision recursion step reaches the preset time, or all the i+1 level candidate decision nodes corresponding to the i-th level node are pruned, or the i-th linear candidate trajectory of the i-th level node does not meet the trajectory feasibility condition, the i-th level node is determined as a decision node that meets the termination condition.

9. The method according to claim 8, wherein, The process of obtaining the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory includes: For the linear candidate trajectories corresponding to at least one decision node that meets the termination condition, trajectory feasibility screening is performed to obtain at least one feasible candidate trajectory. Based on the global decision probability of the decision nodes corresponding to each feasible candidate trajectory and the trajectory evaluation information of each feasible candidate trajectory, the comprehensive evaluation result of each feasible candidate trajectory is determined. Based on the comprehensive evaluation results, the target optimized trajectory is determined from the at least one feasible candidate trajectory.

10. The method according to any one of claims 1 to 9, wherein, Determine the local decision probability of each (i+1)th level candidate decision node, including: Search for at least two preset interaction decisions between the target vehicle and the i-th level key obstacle, and obtain the search statistics results corresponding to each preset interaction decision; Based on the search statistics results corresponding to each preset interactive decision, the probability of the i-th local decision corresponding to each (i+1)-th level candidate decision node is determined.

11. The method according to any one of claims 1 to 10, wherein, The step of determining the (i+1)th global decision probability of each (i+1)th level candidate decision node based on the i-th global decision probability corresponding to the i-th level node and the i-th local decision probability of each (i+1)-th level candidate decision node includes: The i-th global decision probability corresponding to the i-th level node is multiplied by the i-th local decision probability of the (i+1)-th level candidate decision node to obtain the (i+1)-th global decision probability of the (i+1)-th level candidate decision node. or, The global decision probability of each node extended to the (i+1)th level candidate decision node is multiplied by the i-th local decision probability of the (i+1)th level candidate decision node to obtain the (i+1)th global decision probability of the (i+1)th level candidate decision node.

12. A trajectory determination device for an autonomous driving scenario, comprising: A recursive processing unit is used to execute cascaded decision recursive steps, wherein the cascaded decision recursive steps include: determining the i-th level key obstacle corresponding to the i-th level node based on the i-th linear candidate trajectory contained in the i-th context information of the i-th level node and the state information of each obstacle; wherein the i-th level node is the root node or an i-th level decision node cascaded with the root node; the i-th context information includes the i-th linear candidate trajectory of the target vehicle, the i-th global decision probability, and the state information of each obstacle; and determining at least two (i+1)-th level candidate decision nodes for the i-th level node, and each (i+1)-th level candidate decision node, based on the i-th level key obstacle. The local decision probability of the i-th decision node; where the local decision probability of the i-th decision node represents the probability that the target vehicle and the i-th level key obstacle will interact according to the preset interaction decision corresponding to the (i+1)-th level candidate decision node; based on the global decision probability corresponding to the i-th level node and the local decision probability of each (i+1)-th level candidate decision node, the (i+1)-th global decision probability of each (i+1)-th level candidate decision node is determined; based on the (i+1)-th global decision probability of each (i+1)-th level candidate decision node, the (i+1)-th level decision node that meets the expansion condition is determined, and the (i+1)-th linear candidate trajectory of each (i+1)-th level decision node is obtained; The trajectory output unit is used to obtain the linear candidate trajectory of the target vehicle corresponding to at least one decision node that satisfies the termination condition, and to obtain the target optimized trajectory of the target vehicle based on at least one linear candidate trajectory; the decision node is a node that is extended from the root node after the target vehicle interacts with multiple obstacles based on the initial context information of the root node.

13. An electronic device, comprising: At least one processor; as well as A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-11.

14. A non-transitory computer-readable storage medium storing computer instructions, wherein, The computer instructions are used to cause the computer to perform the method according to any one of claims 1-11.

15. A computer program product comprising a computer program that, when executed by a processor, implements the method according to any one of claims 1-11.

16. An autonomous vehicle, including the electronic equipment as claimed in claim 13.