A multi-net connected vehicle decision planning framework with diversity and interpretability

Abstract

The application relates to the fields of artificial intelligence and automatic driving, and specifically discloses a multi-net connected vehicle decision planning framework with diversity and interpretability, which comprises a decision module and a trajectory planning module; the decision module is used for solving the interactive behaviors among vehicles in a scene, and generates a rough and interpretable decision sequence for all the net connected vehicles; the decision module uses an improved Monte Carlo tree search algorithm; the trajectory planning module receives the discrete decision sequence output by the decision module, and generates a continuous kinematically feasible trajectory for each controlled net connected vehicle; the trajectory planning module adopts a distributed parallel architecture, that is, a planner is independently operated for each vehicle; the planner firstly predicts the future trajectories of other vehicles in the scene, and so on; the application can generate trajectories with diversity, interaction and interpretability in a mixed net connected vehicle and human-driven vehicle scene.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and autonomous driving, specifically a multi-connected vehicle decision-making and planning framework that is diverse and interpretable. Background Technology

[0002] The decision-making and planning module is a core component of the autonomous driving system. It receives the perception results of the upstream perception module on the surrounding environment (including the surrounding static environment and dynamic objects) and finally generates the vehicle's future driving trajectory through algorithms.

[0003] Decision-making and planning modules are currently widely used in single autonomous vehicles and are also known as single-vehicle decision-making and planning algorithms. Popular single-vehicle decision-making and planning methods can be divided into two categories: decision-heavy frameworks and decision-light frameworks. The former typically separates decision-making and motion planning to reduce computational burden and complexity, while decision-light methods improve the algorithm's adaptability to corner cases by weakening the decision-making module and increasing the planner's sampling range.

[0004] Furthermore, with the recent widespread application of vehicle-to-everything (V2X) technology in real-world scenarios, some approaches are beginning to consider joint decision-making and trajectory planning for multiple vehicles in traffic situations. Centralized multi-vehicle planners can effectively address complex vehicle interactions in traffic flow, including competition and cooperation. Currently, several types of multi-vehicle decision-making and planning methods exist in academia. Monte Carlo tree search-based methods use a tree structure to search for possible interactions between multiple vehicles and significantly improve algorithm efficiency by sampling the vehicle's action space. Machine learning-based methods directly obtain vehicle driving strategies from open datasets and predict and plan the trajectories of multiple vehicles through forecasting.

[0005] In existing technologies: single-vehicle decision-making and planning algorithms struggle to engage in game theory with other traffic participants in the real world, thus exhibiting insufficient understanding of inter-vehicle interactions; in multi-vehicle joint decision-making and planning technologies, Monte Carlo tree search methods typically directly use vehicle control variables as decision outputs, leading to collision risks when human-driven vehicles behave unexpectedly; furthermore, machine learning-based trajectories fail to clearly express driving intentions, making it difficult for passengers or surrounding traffic participants to understand the connected vehicle's driving status. While SUMO-generated vehicle trajectories in autonomous driving simulation software can avoid collisions, they lack vehicle kinematic constraints and employ instantaneous movement for lane changes, inconsistent with real-world scenarios. CARLA employs a relatively rigid vehicle behavior model, resulting in overly conservative and undiversified generated vehicles.

[0006] To address the aforementioned issues, this invention provides a multi-connected vehicle decision-making and planning framework that is diverse and interpretable. Summary of the Invention

[0007] The purpose of this invention is to provide a diverse and interpretable decision-making and planning framework for multiple connected vehicles, primarily addressing the problem of simultaneously making decisions and plans for multiple connected vehicles in scenarios where autonomous and human-driven vehicles coexist. This invention enables joint decision-making among these connected vehicles and generates interpretable behavioral sequences for each autonomous vehicle, explicitly conveying its intentions to the human-driven vehicles in the scenario. Furthermore, the trajectories planned for connected vehicles by this invention exhibit diversity at two levels: social behavior (how vehicles interact with other vehicles, such as yielding, overtaking, and accelerating) and driving habits (individual differences among drivers in terms of cruising comfort and safety).

[0008] To achieve the above objectives, the present invention provides the following technical solution: a multi-connected vehicle decision-making and planning framework with diversity and interpretability, comprising a decision-making module and a trajectory planning module:

[0009] The decision module is used to solve the interaction behavior between vehicles in the scenario, and generates a coarse, interpretable decision sequence for all connected vehicles. The decision module uses an improved version of Monte Carlo tree search.

[0010] The trajectory planning module receives the discrete decision sequence output by the decision module and generates a continuous, kinematically feasible trajectory for each controlled connected vehicle. The trajectory planning module adopts a distributed parallel architecture, that is, each vehicle runs an independent planner. The planner first predicts the future trajectories of other vehicles in the scene. Then, it separates the different actions in the decision sequence and plans an optimal sub-trajectory for each action segment. Finally, the planner connects all the sub-trajectories to obtain the final kinematically feasible and continuous vehicle trajectory.

[0011] In a preferred embodiment of the present invention, the improved Monte Carlo tree search uses a meta-node instead of the node in the general Monte Carlo tree search method. This node can simultaneously simulate and generate decisions for multiple vehicles. For a meta-node extended to time step t, it receives the state of all vehicles at the current moment; vehicle V i The state at time step t can be described as x. i t =[s,d,v] Ts and d are the Frenzied coordinates of the vehicle, and v is the longitudinal velocity; each controlled vehicle defines a set of possible actions in one time step; the actions are five actions - maintain speed, accelerate, decelerate, change lanes to the left, and change lanes to the right; by selecting different actions, the vehicle can calculate its corresponding state in the next time step; as for uncontrolled human-driven vehicles in traffic, the default action is lane keeping; and an appropriate distance is always maintained from the vehicle in front during the decision-making process.

[0012] As a preferred embodiment of the present invention, the present invention also includes pruning the invention search tree.

[0013] As a preferred embodiment of the present invention, the vehicle in the scenario of the present invention should maintain a minimum safe distance Ds from the vehicle in front, where Ds consists of two parts: reaction distance and minimum clearance, defined as: D D =v·τ+MTH·Δv, where τ>0 and MTH>0 are time constants, representing the driver's reaction time and the shortest collision time, respectively; v represents the vehicle's current speed, and Δv = vv l This represents the speed difference between the vehicle itself and the vehicle in front; the speed constraints that a vehicle must satisfy at any given moment in the scenario are:

[0014]

[0015] Where Δs l and Δs f v represents the distance between the current vehicle and the vehicle in front / behind it, respectively. l and v f Representing the speeds of the preceding and following vehicles, in the meta node, if vehicle V i The state resulting from taking action a If the above equation is not satisfied, action a is invalid, and the corresponding new state is... The action combination of the meta node will be abandoned. For all valid state combinations within the meta node, it is necessary to check the collisions between the new states in each combination, as well as the conflicts in the process of reaching the state. If a conflict occurs, the action combination of the meta node will not be expanded in time step t+1.

[0016] In a preferred embodiment of the present invention, during the simulation phase of the Monte Carlo tree search algorithm, meta-nodes are continuously constructed, and only one valid action combination is randomly selected to enter the next time step until the simulation terminates. The simulation is terminated when all vehicles have traveled a certain distance in their target lanes, or when the simulation reaches the maximum time step but some vehicles have not yet completed their lane-changing intentions.

[0017] As a preferred embodiment of the present invention, the present invention further includes calculating a reward R for each vehicle i. i The behavioral reward R is calculated using the following formula:

[0018] R = R self +γR other

[0019] Here, γ = 1 indicates that the vehicle considers other vehicles to be of equal importance to itself when making decisions, and each vehicle i in the scenario has its unique factor γ. i R self The vehicle reward, R other As rewards are given to other vehicles, the rewards during the simulation phase are obtained by combining the behavioral rewards of all vehicles, and the formula is calculated as follows:

[0020]

[0021] As a preferred embodiment of the present invention, a cost function is introduced for each generated sub-trajectory in the trajectory planning module to evaluate its safety and comfort. The vehicle cost calculation method is as follows:

[0022]

[0023] Among them, J i,obs (t,j) Characterizing vehicle V j Compared to the current vehicle V i The position at time step t.

[0024] As a preferred embodiment of the present invention, the present invention further includes a warning zone, the shape of which surrounds the vehicle V. i A rectangle is defined, with the length of the warning zone in front of the vehicle being the shortest safe distance Ds, the length of the warning zone behind the vehicle being 1.5 times the length of the vehicle body, and the width of the warning zone being 1.5 times the width of the vehicle body; Vehicle V j In the scene, relative to V i The location is divided into three different types: outside the warning zone, inside the warning zone without collision, and with V. i There is a collision at the current location; vehicle Vj is safe outside the warning zone, so the cost is 0; while vehicle V... j With vehicle V i A direct collision is unacceptable within the warning zone; vehicle V j The closer to V i The higher the cost, the more expensive the J. i,obs (t,j) The calculation formula is as follows:

[0025]

[0026] Where |D| x and |D| y They represent the vehicles V respectively j To vehicle V i The shortest distance along the length and width of the warning zone, l and d represent the length and width of the vehicle body, C z >0 is a cost constant;

[0027] All the above cost items are ultimately summarized into the cost function of the trajectory: J i =∑w(·)J i ,(·). Among them, J i ,(·) represents all the above cost items, and w(·) is the weight parameter corresponding to each cost item.

[0028] As a preferred embodiment of the present invention, the present invention further includes a trajectory weight set W containing M weight vectors, and the algorithm is as follows:

[0029] W = {W1, W2, W3…W} M}

[0030] Each weight vector W in the trajectory weight set W i A unique weighted combination that includes all cost items.

[0031] Compared with existing technologies, the advantages of this invention are: This invention overcomes the shortcomings of existing technologies, generating diverse, interactive, and interpretable trajectories in mixed connected vehicle and human-driven vehicle scenarios. By introducing vehicle cooperation factors into the decision-making module and trajectory weight sets into the planning module, the vehicle trajectories generated by the framework proposed in this invention are diverse in terms of both social interaction and individual driving habits. The framework proposed in this invention performs joint decision-making for all vehicles in the traffic flow and ensures real-time response to dynamic environments through a high-frequency replanning method. Furthermore, the decision-making module of this invention generates interpretable decision sequences, which are used to inform passengers of the current situation and explicitly convey the driver's intentions to the surrounding human drivers. Attached Figure Description

[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention.

[0033] Figure 1 This is a schematic diagram of a multi-connected vehicle decision-making and planning framework with diversity and interpretability according to the present invention;

[0034] Figure 2This is a schematic diagram of the MCTS method of the present invention;

[0035] Figure 3 This is an extended diagram of the MCTS meta-nodes of the present invention;

[0036] Figure 4 This is a flowchart of the trajectory planning method in this invention. Detailed Implementation

[0037] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.

[0038] Example

[0039] Please see Figures 1-4 ,in Figure 1 This is a schematic diagram of our proposed framework, where vehicles with IDs 2 and 3 are connected vehicles controlled by our framework, while vehicle ID 1 is an uncontrolled human-driven vehicle. The left box represents the decision-making module, and the right box represents the trajectory planning module. The connected vehicles mentioned in this invention refer to autonomous, networked vehicles.

[0040] This invention proposes a two-stage multi-vehicle decision-making and planning framework that includes decision-making and trajectory planning modules. The input to this framework is a perceptible environment, which includes road network topology, route information, vehicle states, and trajectory predictions for uncontrolled vehicles.

[0041] The first stage of this invention framework is a decision-making module designed to address the interaction behaviors between vehicles in a given scenario. This module generates a coarse, interpretable decision sequence for all connected vehicles simultaneously. It utilizes an improved version of Monte Carlo Tree Search (MCTS).

[0042] Monte Carlo Search (MCTS) is a heuristic search algorithm for decision-making processes. It analyzes the search space using Monte Carlo sampling to find the most promising actions. A typical MCTS algorithm consists of four phases in a single search iteration: selection, expansion, simulation, and backpropagation. However, traditional MCTS methods usually only allow decision-making for a single agent / vehicle. In multi-vehicle scenarios, it's necessary to assign priorities to vehicles and proceed with decisions sequentially from highest to lowest priority, with later decisions requiring compliance with earlier ones. In contrast, our method considers all vehicles in the traffic scenario, including human-driven vehicles, and eliminates the priority system among vehicles during the decision-making process.

[0043] like Figures 2-3As shown, Figure 2 This is a schematic diagram of the MCTS method, where the selection phase is based on the existing search tree. The algorithm recursively selects the most valuable child nodes using a tree-like strategy until it reaches a node that is not fully expanded. The selected node is then expanded using one of the remaining possible actions to obtain a new leaf node. From this node, the algorithm simulates using a default strategy (also known as expansion) until it reaches a terminal node. In the backpropagation phase, the algorithm extracts rewards from the simulated samples and uses them to update the average reward of all passed nodes. After reaching a predetermined maximum number of iterations, the solution can be obtained by greedily selecting the child nodes with the maximum reward at each level. Figure 3 This demonstrates an extension of the MCTS meta-node in this invention, where vehicles 2 and 3 are controlled by our framework, while vehicle 1 is an uncontrolled vehicle. Each vehicle has five actions—maintain speed (KS), accelerate (AC), decelerate (DC), change lanes left (LCL), and change lanes right (LCR), while a human-driven vehicle has only one action (maintain lane KL).

[0044] This invention replaces the nodes in the general MCTS method with a metanode, which can simultaneously simulate and generate decisions for multiple vehicles. For a metanode extended to time step t, it receives the states of all vehicles at the current time. The state of vehicle Vi at time step t can be described as x. i t =[s,d,v] T Let s and D be the Frenzied coordinates of the vehicle, and v be the longitudinal velocity. This invention defines a set of possible actions for each controlled vehicle in one time step. There are five actions—maintain speed (KS), accelerate (AC), decelerate (DC), change lanes to the left (LCL), and change lanes to the right (LCR). By selecting different actions, the vehicle can calculate its corresponding state in the next time step. As for uncontrolled human-driven vehicles in traffic flow, we assume they are performing lane-keeping (KL) actions. This action dynamically adjusts their speed and maintains an appropriate distance from the vehicle in front throughout the decision-making process.

[0045] As the MCTS evaluates the entire interaction process, the search tree grows exponentially. Therefore, this invention also prunes the search tree. Although each vehicle has five optional actions, these actions are not feasible at every moment; some actions may lead to potential collisions. For example, vehicles on either side of the road cannot continue to change lanes to cross the boundary line. Furthermore, vehicles in the scenario should maintain the shortest possible safe distance D from the vehicle in front. S D S It consists of two parts: the reaction distance and the minimum spacing, defined as: DS =v·τ+MTH·Δv, where τ>0 and MTH>0 are time constants, representing the driver's reaction time and the shortest collision time, respectively; v represents the vehicle's current speed, and Δv = τ - v l This represents the speed difference between the vehicle itself and the vehicle in front; the speed constraints that a vehicle must satisfy at any given moment in the scenario are:

[0046]

[0047] Where Δs l and Δs f v represents the distance between the current vehicle and the vehicle in front / behind it, respectively. l and v f Representing the speeds of the preceding and following vehicles, in the meta node, if vehicle V i The state resulting from taking action a If the above equation is not satisfied, action a is invalid, and the corresponding new state is... The action combination of the meta node will be discarded. For all valid state combinations within the meta node, it is necessary to check for collisions between new states in each combination, as well as conflicts in the process of reaching a state. If a conflict occurs (e.g., a collision occurs), the action combination of the meta node will not be expanded in time step t+1.

[0048] During the MCTS simulation phase, the algorithm continues to construct meta-nodes, but only randomly selects a valid action combination to proceed to the next time step until the simulation terminates. The simulation terminates when all vehicles have traveled a certain distance in their target lanes, or when the simulation reaches its maximum time step but some vehicles have not yet completed their lane-changing intentions. Subsequently, the algorithm calculates the reward for this simulation based on the results. We first calculate the reward Ri for each vehicle i in the traffic flow, where the reward should be within the range of 0 to 1. To make vehicles exhibit diverse social behaviors, each vehicle's cooperative tendency needs to be represented by a reward. Therefore, this invention introduces a behavioral reward R = Ri, which considers both self-reward and reward for others. self +γR other Here, γ∈[0,1] is the vehicle cooperation factor. γ=0 means the vehicle is selfish and does not consider the behavior of other vehicles, while γ=1 means the vehicle considers other vehicles to be of equal importance to itself when making decisions. Each vehicle i in the scenario has its unique factor γ. i Since each meta-node in the search tree handles the actions of all vehicles, the reward for the simulation phase is obtained by combining the rewards for the actions of all vehicles, calculated using the following formula:

[0049]

[0050] Furthermore, such as Figure 4 As shown, Figure 4 A flowchart of the trajectory planning method in this invention is shown. It receives a decision sequence from the decision module. The planner first predicts the future trajectories of other vehicles in the scene. Then, it separates the different actions in the decision sequence and plans an optimal sub-trajectory for each action segment. Finally, the planner connects all the sub-trajectories to obtain the final kinematically feasible and continuous vehicle trajectory.

[0051] The second stage of this invention is a trajectory planning module that receives the discrete decision sequence output from the first stage and generates a continuous, kinematically feasible trajectory for each controlled connected vehicle. This planning module employs a distributed parallel architecture, meaning that a planner runs independently for each vehicle, which is closer to real-world human driving. The planner maintains a high replanning frequency and continuously predicts surrounding vehicles to avoid potential collisions.

[0052] For a connected vehicle V under this framework control i ∈V C It receives its decision sequence from the decision module and assigns it to a trajectory planner. This planner first predicts the future trajectories of other vehicles in the scene. Then, it segments the different actions in the decision sequence and plans an optimal sub-trajectory for each action segment. Finally, the planner connects all the sub-trajectories to obtain the final kinematically feasible, continuous vehicle trajectory. For the sub-trajectory planner, any currently common autonomous driving single-vehicle trajectory planning algorithm can be adapted to our framework with minor modifications.

[0053] For each generated sub-trajectory, we should calculate a cost function to evaluate its safety and comfort. Specifically, the cost function includes several cost terms: curve energy cost (cumulative sum of squared trajectory curvature), orientation difference cost (cumulative sum of squared differences between the vehicle's heading and the lane centerline), lane departure cost (cumulative sum of squared perpendicular distances between the vehicle and the lane centerline), acceleration cost (cumulative sum of squared trajectory acceleration), jerk cost (cumulative sum of squared trajectory jerk), and dynamic obstacle cost.

[0054] Dynamic obstacle cost takes into account the distance between a vehicle traveling along a trajectory and dynamic obstacles on the road (i.e., surrounding vehicles). The cost of vehicle Vi is defined as:

[0055]

[0056] Among them, Ji ,obs (t,j) Characterizing vehicle V j Compared to the current vehicle V i The position at time step t.

[0057] This invention introduces an alert zone, which is a rectangle surrounding the vehicle Vi. The length of this zone in front of the vehicle is the shortest safe distance D. s As defined by the formula above, the length of the warning zone behind the vehicle is 1.5 times the length of the vehicle body. The width of the warning zone is 1.5 times the width of the vehicle body. The position of vehicle Vj relative to Vi in the scene can be divided into three different types: outside the warning zone, inside the warning zone without collision, and colliding with Vi's current position. Vehicle Vj is safe outside the warning zone, so the cost is 0; however, a direct collision between vehicle Vj and vehicle Vi is unacceptable. Within the warning zone, the closer vehicle Vj is to Vi, the higher the cost. i,obs (t,j) is specifically defined as follows:

[0058]

[0059] Where |D| x and |D| y They represent the vehicles V respectively j To vehicle V i The shortest distance along the length and width of the warning zone, l and d represent the length and width of the vehicle body, C z >0 is a cost constant.

[0060] All the above cost items are ultimately summarized into the cost function of the trajectory: J i =∑w(·)J i ,(·). Among them, J i ,(·) represents all the above cost items, and w(·) is the weight parameter corresponding to each cost item. Different vehicles in the traffic flow have different weights for these trajectory cost items. For example, some vehicles may be more concerned with driving comfort, while others prefer to maintain a longer safe distance from other vehicles. This invention defines a trajectory weight set W containing M weight vectors. The algorithm is as follows: W={W1,W2,W3…W M}

[0061] Each weight vector W in the trajectory weight set W i A unique weight combination encompassing all cost items. In this scenario, connected vehicles select different weight vectors to generate their trajectories, and combined with cooperative factors considering social behavior in the decision-making module, this brings diversity to the entire traffic flow.

[0062] In summary, the advantages of this invention are as follows:

[0063] -Trajectory diversity. This invention incorporates vehicle cooperation factors into the decision-making module and introduces trajectory weight sets into the planning module, thereby enabling the generated multi-vehicle trajectories to exhibit diversity at both the social interaction and individual driving habit levels.

[0064] - Decision interpretability. The decision module in the framework proposed in this invention generates an interpretable sequence of behaviors for each connected vehicle, enabling passengers on board to understand the current situation and clearly communicate driving intentions to surrounding human-driven vehicles through brake lights and turn signals.

[0065] - Vehicle Interaction. This invention takes into account all vehicles, including human-driven vehicles, when making decisions, thus enabling the exploration of complex interactions between vehicles.

[0066] - Bottom-line safety. The planning module of this invention ensures the real-time response of the entire framework to dynamic environments through high-frequency re-planning. Even if the decision module generates some practically infeasible decisions, the trajectory planning module can avoid potential collision risks by predicting other vehicles in the vicinity.

[0067] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element.

[0068] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A multi-connected vehicle decision-making and planning framework with diversity and interpretability, characterized in that, Includes a decision-making module and a trajectory planning module: The decision module is used to solve the interaction behavior between vehicles in the scenario, and generates a coarse, interpretable decision sequence for all connected vehicles. The decision module uses an improved version of Monte Carlo tree search. The trajectory planning module receives the discrete decision sequence output by the decision module and generates a continuous, kinematically feasible trajectory for each controlled connected vehicle. The trajectory planning module employs a distributed parallel architecture, meaning each vehicle runs an independent planner. The planner first predicts the future trajectories of other vehicles in the scenario. Then, it separates different actions in the decision sequence, plans an optimal sub-trajectory for each action segment, and finally connects all the sub-trajectories to obtain the final kinematically feasible, continuous vehicle trajectory. The improved Monte Carlo tree search uses a meta-node instead of the node in the general Monte Carlo tree search method; this node can simultaneously simulate and generate decisions for multiple vehicles. For extensions to more time steps… The meta node receives the current state of all vehicles; vehicles In time steps The state can be described as x i t =[s,d,v] T , and These are the vehicle's Frenét coordinates. It refers to longitudinal speed; each controlled vehicle defines a set of possible actions in a time step; these actions include five actions—maintaining speed, accelerating, decelerating, changing lanes to the left, and changing lanes to the right; by selecting different actions, the vehicle can calculate its corresponding state in the next time step t+1; as for uncontrolled human-driven vehicles in traffic flow, the default action is lane keeping; and they always maintain an appropriate distance from the vehicle in front during the decision-making process; it also includes a... The algorithm for the trajectory weight set W of weight vectors is as follows: Each weight vector in the trajectory weight set W A unique combination of weights for all cost items; It also includes a behavioral reward that considers both self-reward and reward for others. ,in, ∈[0,1] is the vehicle cooperation factor. =0 means the vehicle is selfish and doesn't consider the behavior of other vehicles, while =1 indicates that the vehicle considers other vehicles to be of equal importance to itself when making decisions.

2. The multi-connected vehicle decision-making and planning framework with diversity and interpretability according to claim 1, characterized in that, It also includes pruning the invention search tree.

3. The multi-connected vehicle decision-making and planning framework with diversity and interpretability according to claim 2, characterized in that, Vehicles in the scenario should maintain the shortest safe distance Ds from the vehicle in front, where Ds consists of two parts. Components: reaction distance and minimum spacing, defined as: ,in, and These are all time constants, representing the driver's reaction time and the shortest collision time, respectively. Indicates the vehicle's current speed. This represents the speed difference between the vehicle itself and the vehicle in front; the speed constraints that a vehicle must satisfy at any given moment in the scenario are: in and These represent the distances between the current vehicle and the vehicle in front / behind it, respectively. and Representing the speeds of the preceding and following vehicles, in the meta node, if the vehicles... Action was taken The resulting state When the above equation is not satisfied, the action... Invalid, corresponding new state The state will be discarded; for all valid state combinations within the metanode, it is necessary to check for collisions between new states in each combination, as well as conflicts during the process of reaching a state. If a conflict occurs, it will be resolved at the time step. In +1, the action combinations of meta nodes will not be expanded.

4. The multi-connected vehicle decision-making and planning framework with diversity and interpretability according to claim 3, characterized in that, In the simulation phase of the Monte Carlo tree search algorithm, meta-nodes are continuously constructed, and only one valid action combination is randomly selected to enter the next time step until the simulation terminates. The simulation is terminated when all vehicles have traveled a certain distance in their target lanes, or when the simulation reaches the maximum time step but there are still vehicles that have not completed their lane-changing intentions.

5. A multi-connected vehicle decision-making and planning framework with diversity and interpretability as described in claim 4, characterized in that, It also includes each vehicle Calculate rewards separately and behavioral rewards The calculation formula is as follows: Among them, and This means that when making decisions, each vehicle considers other vehicles to be of equal importance to itself. Each has its unique factors , The vehicle was awarded. As rewards are given to other vehicles, the rewards during the simulation phase are obtained by combining the behavioral rewards of all vehicles, and the formula is calculated as follows: 。 6. The multi-connected vehicle decision-making and planning framework with diversity and interpretability according to claim 5, characterized in that, For each generated sub-trajectory in the trajectory planning module, a cost function is introduced to evaluate its safety and comfort. The vehicle cost calculation method is as follows: in, Characterizing vehicles Compared to the current vehicle At time step The position at that time.

7. The multi-connected vehicle decision-making and planning framework with diversity and interpretability according to claim 6, characterized in that, It also includes a warning zone, the shape of which surrounds the vehicle. A rectangle is defined, with the length of the warning zone in front of the vehicle being the shortest safe distance Ds, the length of the warning zone behind the vehicle being 1.5 times the length of the vehicle body, and the width of the warning zone being 1.5 times the width of the vehicle body; In the scene relative to The location is divided into three different types: outside the warning zone, inside the warning zone without collision, and with... There is a collision at the current location; vehicle It is safe outside the restricted area, therefore the cost is zero; while the vehicle... With vehicles A direct collision is unacceptable within the warning zone; vehicles... The closer The higher the cost, the more expensive it is; The calculation formula is as follows: in and Representing vehicles To the vehicle Along the shortest distance of the length and width of the warning zone, and Indicates the length and width of the vehicle body. 0 is a cost constant; All the above cost items are ultimately summarized into the cost function of the trajectory: ,in representing all the above cost items, w( ) is the weight parameter corresponding to each cost item.

Images