Quantum game scenario deduction automatic driving planning method and system

Abstract

This application discloses a quantum game scenario extrapolation method and system for autonomous driving planning, belonging to the field of autonomous driving decision-making and planning technology. The method integrates multi-view images, scene semantics, vehicle state, and traffic participant states to generate corresponding latent space tokens; based on these tokens, an interactive game model is constructed, and the joint strategy probability distribution and vehicle decision results are obtained through quantum state entanglement, unitary transformation, anti-entanglement, and measurement; the probability distribution and decision results are encoded into game condition vectors, and multi-agent future scenario extrapolation is performed. The extrapolated trajectory is used to update the dynamic payoff function, re-solve the game, and repeat the extrapolation to output the final planned trajectory of the vehicle. The system includes modules for multimodal data acquisition, multimodal state encoding, quantum game decision-making, generative scenario extrapolation, feedback update, and planned trajectory output.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving decision-making and planning technology, and in particular to an autonomous driving planning method and system based on quantum game scenario deduction. Background Technology

[0002] As autonomous vehicles operate in scenarios that expand from structured roads to roundabouts, ramp merging areas, and areas with multiple traffic flows, vehicle planning systems need to perceive not only roads, lanes, and traffic participants, but also predict the future behavior of surrounding traffic participants and formulate the vehicle's planned trajectory based on the predictions. In these scenarios, there is an interaction between the vehicle's behavior and the behavior of surrounding traffic participants. When the vehicle accelerates, decelerates, or yields, the motion state of surrounding traffic participants may change accordingly, and their changed motion state will then constrain the vehicle's subsequent trajectory.

[0003] Traditional autonomous driving planning systems typically set environmental perception, behavior prediction, and trajectory planning as sequential processing modules. Detection or prediction errors from earlier modules may propagate to later modules, and the training objectives and processing results of different modules may lack unified constraints for the final planning task. In scenarios where the behavior of traffic participants is uncertain, a unidirectional serial processing approach struggles to fully represent the strategic interactions between the vehicle and surrounding traffic participants.

[0004] Chinese invention patent application CN 121212372 A, published on December 26, 2025, discloses a driving trajectory planning method and intelligent driving system based on a visual language model. This solution acquires multi-view images of the vehicle and driving intention text, directly generates the vehicle's future trajectory through a visual language model, and reduces model inference overhead through viewpoint or image patch filtering, visual token compression, and speculative decoding.

[0005] Chinese invention patent application CN 120182938 A, published on June 20, 2025, discloses a large-scale autonomous driving model framework based on 3D spatiotemporal perception and human-like decision-making reasoning. This scheme establishes a 3D spatiotemporal scene representation through joint encoding of full-view scene video streams and a global bird's-eye view. Visual and textual tags are input into a large-scale language model, and scene parsing, semantic action inference, and motion trajectory planning are performed through multi-task perception and classification decoding.

[0006] The existing solutions mentioned above improve multimodal scene perception, semantic reasoning, and vehicle trajectory generation capabilities, but their planning process mainly revolves around direct model output or rule-driven action inference. In interactive scenarios such as roundabouts and merging, existing solutions still have room for improvement in representing the joint influence between the vehicle's strategy and the strategies of multiple traffic participants, the uncertainty of traffic participants' intentions, and the coexistence of multiple feasible decision outcomes.

[0007] Some interactive planning methods use safety distance, collision time, or classical game theory models to evaluate traffic conflicts. Safety distance and collision time typically focus on one-way risk assessment; classical game theory methods usually select equilibrium strategies based on deterministic payoff data. When multiple equilibrium outcomes coexist, or when traffic participants do not act in a completely rational manner, deterministic strategy selection may not fully reflect the uncertainty of actual traffic behavior.

[0008] Generative scene prediction can continuously generate vehicle trajectories and traffic participant trajectories for multiple future moments based on current observations. However, the autoregressive generation process uses the prediction results of the previous moment as input for the next moment, and early prediction errors may gradually accumulate as the extrapolation time domain extends. If the generated results cannot be fed back to the decision model, the initial decision conditions are also difficult to correct based on future behavioral changes, which can easily lead to inconsistencies between the decision basis and the extrapolated scenario.

[0009] When there is a lack of a unified data interface and iterative update conditions between generative inference results and game decisions, game payoffs, strategy probabilities, and inference trajectories are usually calculated independently at different processing stages, making it difficult to form a repeatable closed-loop update.

[0010] Therefore, there is a need for an autonomous driving planning method and system that can express the policy relationships between the vehicle and multiple traffic participants in complex interactive driving scenarios, handle policy uncertainty and multiple decision results, and dynamically correct planning decisions using future scenario projection results. Summary of the Invention

[0011] The main objective of this invention is to provide a quantum game scenario extrapolation method and system for autonomous driving planning. It aims to solve the problems of how to establish policy associations between the vehicle and traffic participants based on multimodal scenario information in complex interactive driving scenarios, how to use joint policy probabilities to express policy uncertainty, and how to feed the results of future scenario extrapolation back to the game decision-making process in order to reduce the inconsistency between initial decision conditions and future interaction trends.

[0012] To achieve the above objectives, this invention proposes a quantum game scenario deduction method for autonomous driving planning, comprising: Acquire multi-view camera image sequences, scene semantic descriptions, vehicle action questions, vehicle motion states, and the motion states of surrounding traffic participants. Extract visual features from the multi-view camera image sequences and linguistic features from the scene semantic descriptions and vehicle action questions. Fuse the visual features, linguistic features, vehicle motion states, and the motion states of surrounding traffic participants using a multimodal large language model to generate latent space tokens that represent the motion states of the vehicle and each surrounding traffic participant. At least one traffic participant is identified from the surrounding traffic participants, and the vehicle and the identified traffic participant are taken as game participants. A driving strategy space including two candidate driving strategies is determined for each game participant. An interactive game model with a dynamic payoff function is constructed based on the latent space token and the driving strategy space. Each player's two candidate driving strategies are mapped to two computational ground states of a qubit. The quantum states of each player form a joint quantum state. Entanglement operators, unitary operators of each player, and anti-entanglement operators are sequentially applied to the joint quantum state. The final joint quantum state after applying the anti-entanglement operator is measured to obtain the joint strategy probability distribution. The decision result of the self-vehicle game is determined based on the joint strategy probability distribution and the dynamic payoff function. The joint strategy probability distribution and the vehicle game decision result are encoded into a game condition vector. The game condition vector and the latent space token are input into a generative inference engine to predict the driving scenario at multiple future moments through autoregression, and generate the vehicle trajectory points and the trajectory points of each traffic participant at the corresponding moments. Based on the generated vehicle trajectory points and the trajectory points of each traffic participant, the behavioral evolution information in the prediction time domain is determined. The dynamic reward function is updated using the behavioral evolution information. Based on the updated dynamic reward function, quantum game processing is re-executed to obtain the updated joint strategy probability distribution and vehicle game decision results. The updated joint strategy probability distribution and vehicle game decision results are encoded into an updated game condition vector. Based on the updated game condition vector, driving scenario prediction is performed again. After iteration, the final vehicle planned trajectory is output.

[0013] Preferably, the vehicle's motion state includes its three-dimensional position, attitude angle, velocity vector, and acceleration vector in the global coordinate system; the motion state of each surrounding traffic participant includes traffic participant category, three-dimensional bounding box, historical motion trajectory, velocity vector, and acceleration vector; the scene semantic description includes scene category, key object information, and dynamic event description; the vehicle's action question is used to guide the multimodal large language model to generate vehicle driving decisions.

[0014] Preferably, the first Images from the perspective of a camera Input a visual Transformer with shared weights for each camera viewpoint to obtain the corresponding visual features. and in accordance with Multi-perspective visual features are formed along the spatial dimension, among which, The total number of camera viewpoints; the scene semantic description and the vehicle's motion problem are encoded into text embeddings. and in accordance with Cross-modal fusion is performed, in which, This indicates that the multi-view visual features are projected onto the visual-linguistic projection layer of the language embedding space. This indicates that the multimodal large language model, from the cross-modal fusion results Extract the initial representation of the vehicle token and the initial representation of the tokens of each surrounding traffic participant; where t is the current processing time, v is the camera view index, and v takes values ​​from 1 to V.

[0015] Preferably, the initial representation of the vehicle token is input into the encoder of the variational autoencoder to obtain the mean value. And the logarithmic variance, and obtain the standard deviation based on the logarithmic variance. ,according to Generate a hidden space token for the vehicle; Initial representation of tokens from surrounding traffic participants The input traffic participant motion prediction head's encoding branch is a three-layer multilayer perceptron with shared weights among surrounding traffic participants, and is processed according to... Generate corresponding hidden space tokens for traffic participants, where... , , This is the weight matrix of the three-layer multilayer perceptron. , , Let t be the corresponding bias vector; where t is the current processing time, i is the index of the surrounding traffic participants, I is the identity matrix, and ReLU represents the linear rectified activation function.

[0016] Preferably, when there are at least two surrounding traffic participants, the two closest surrounding traffic participants to the vehicle are selected to form a three-player game with the vehicle; when there is only one surrounding traffic participant, the surrounding traffic participant and the vehicle form a two-player game. The method is applied to roundabout scenarios or merging scenarios. Each player has two candidate driving strategies, namely an acceleration strategy and a deceleration strategy. The latent space tokens of each player are input into the strategy classification head, and the strategy classification head determines the feasible driving strategy of the corresponding player from the acceleration strategy and deceleration strategy.

[0017] Preferably, for relatively close speeds The The first game participant and the second Each participant in the game, according to Determine the collision time, where, For the two game participants in terms of strategy and strategy The distance below; for the joint strategy profile , No. The security benefit for each game participant is The comfort benefits are Efficiency gains are ,in, For indicator functions, The collision time threshold, For the first The acceleration of each participant in the game. The acceleration standard deviation parameter, For the expected arrival time, The ideal arrival time; the dynamic revenue function is ,in, , and All are located in the interval And satisfy Where i and j are the game participant indices, a is the joint strategy profile, and the acceleration standard deviation parameter is greater than 0; the efficiency gain is calculated according to the above formula only when the expected arrival time is greater than 0, and the efficiency gain is 0 when the expected arrival time is less than or equal to 0.

[0018] Preferably, for the first Each game participant is assigned a qubit, and the acceleration and deceleration strategies are mapped to the computational ground state, respectively. and and with Indicates the first The quantum states of each participant in the game, where... and Let be a complex probability amplitude, and satisfy . ;according to form The initial joint quantum state of each game participant, and through entanglement operators Entanglement processing is performed on the initial joint quantum state, wherein, For entanglement parameters, To act on the first Each game participant corresponds to a Pauli-X operator for a qubit. is the imaginary unit; n is an integer not less than 2, and i is the index of the game participant, with a value from 0 to n-1.

[0019] Preferably, the first The unitary operators of each game participant are ,in, ;according to Obtain the final joint quantum state and cross-reference the joint policy in the final joint quantum state. Modulus square of corresponding component The probability of being identified as the joint strategy profile.

[0020] Preferably, determining the decision result of the self-vehicle game includes: judging whether there is a strictly dominant strategy for the self-vehicle, when the self-vehicle strategy... For any other autonomous vehicle strategy and strategy profiles of any other game participants All meet At that time, the aforementioned autonomous vehicle strategy The outcome of the game is determined as the outcome of the self-vehicle game; when there is no strictly dominant strategy for the self-vehicle, determine whether there exists a unique pure strategy Nash equilibrium, when the joint strategy profile... For each player in the game, the following conditions are met: When the pure strategy Nash equilibrium is unique, the vehicle's strategy in that pure strategy Nash equilibrium is determined as the vehicle's game decision outcome; when no pure strategy Nash equilibrium exists or multiple pure strategy Nash equilibria exist, the strategy is determined according to... Calculate the quantum expected utility of each vehicle's strategy, and according to... Determine the decision outcome of the self-vehicle game; where A_i is the driving strategy space of the i-th player, A_-i is the joint strategy space of the other players besides the i-th player, A_0 is the driving strategy space of the self-vehicle, and A_-0 is the joint strategy space of the other players besides the self-vehicle; when the quantum expected utility of multiple self-vehicle strategies is the same, determine the decision outcome of the self-vehicle game according to the pre-fixed arrangement order in the self-vehicle driving strategy space A_0.

[0021] Preferably, the joint policy probability data representing the probabilities of each joint policy profile is used. Flattened into a one-dimensional probability vector, the one-dimensional probability vector is then compared with the game decision result of the autonomous vehicle. The one-hot encoded vectors are concatenated and then... Generate the game condition vector, where, , and Let be the projection matrix. and The corresponding bias vector is used, and the game condition vector remains unchanged in each inference step of the same round of driving scenario prediction; wherein, the joint strategy probability data is flattened according to the fixed arrangement order of each joint strategy profile in the joint strategy space.

[0022] Preferably, for the first Each step of the deduction process, according to Construct the input sequence of the generative inferrer, wherein, For the vehicle's hidden space token, A hidden space token for all traffic participants. Let be the game condition vector. For scene semantic description, For the problem of vehicle motion; the multimodal large language model predicts the first... The latent space tokens of the vehicle, each traffic participant, the scene semantic description, and the vehicle's action response for each inference step are used. The decoder of the variational autoencoder decodes the predicted latent space tokens of the vehicle into vehicle trajectory points. The decoding branch of the traffic participant motion prediction head decodes the predicted latent space tokens of the traffic participants into trajectory points and 3D bounding boxes for the corresponding traffic participants. The predicted latent space tokens of the vehicle and the latent space tokens of the traffic participants are then used as the first inference step. The input for each inference step is used, and the predicted scene semantic description and the vehicle's action response are used as the text context for the next inference step to construct the scene semantic description and vehicle action question for the next inference step; where k is a non-negative integer and Embed represents the text embedding operation.

[0023] Preferably, the collision frequency in the prediction time domain is determined based on the generated vehicle trajectory points and the trajectory points of each traffic participant. ,according to and The safety and efficiency benefit weights are updated, while the comfort benefit weight remains unchanged in the update formula. Then, the updated safety, comfort, and efficiency benefit weights are normalized so that the sum of the three benefit weights is 1. To update the step size, the update step size is set such that both the updated safety gain weight and efficiency gain weight are within the interval [range]. The dynamic payoff function is reconstructed using the updated payoff weights, the autonomous vehicle game decision result and game condition vector are redefined, and driving scenario prediction is performed again based on the redefined game condition vector. The driving scenario prediction, payoff weight update, and redefined autonomous vehicle game decision result and game condition vector constitute one iteration, and three iterations are performed. The trajectory set {τ} consists of the autonomous vehicle trajectory and the trajectory of each traffic participant generated in the prediction time domain, and CollisionRate({τ}) is the proportion of the number of simulation times corresponding to the collision event to the total number of simulation times evaluated.

[0024] This application also discloses a quantum game scenario deduction autonomous driving planning system, including a multimodal data acquisition module, a multimodal state encoding module, a quantum game decision-making module, a generative scenario deduction module, a feedback update module, and a planned trajectory output module; the multimodal data acquisition module is connected to a vehicle-mounted camera, a vehicle state acquisition unit, and a traffic participant state acquisition unit, respectively, and is used to acquire multi-view camera image sequences, scene semantic descriptions, vehicle action problems, vehicle motion states, and surrounding traffic participant motion states, and send the data associated according to timestamps to the multimodal state encoding module; the multimodal... The state encoding module extracts visual and linguistic features, fuses these features with the vehicle's motion state and the motion states of surrounding traffic participants using a multimodal large language model, and generates latent space tokens representing the motion states of the vehicle and each surrounding traffic participant, respectively. The quantum game decision module is data-connected to the multimodal state encoding module and determines the game participants, driving strategy space, and dynamic payoff function based on the latent space tokens, forming a joint quantum state. Entanglement processing, unitary transformation of each game participant, and deentanglement processing are then sequentially performed, and the final joint quantum state is measured to obtain the joint strategy probability distribution. The system determines the vehicle game decision result based on the joint strategy probability distribution and dynamic payoff function. The generative scenario deduction module is data-connected to the multimodal state encoding module and the quantum game decision module, respectively. It encodes the joint strategy probability distribution and the vehicle game decision result into a game condition vector, and predicts the driving scenario at multiple future moments based on the game condition vector and latent space token autoregression, generating the vehicle trajectory point and the trajectory points of each traffic participant at the corresponding moments. The feedback update module is data-connected to the generative scenario deduction module and the quantum game decision module, respectively, and is used to determine the vehicle game decision result based on the generated vehicle trajectory points. The system determines the behavioral evolution information within the prediction time domain based on the trajectory points of each traffic participant. It then uses this behavioral evolution information to update the dynamic payoff function and triggers the quantum game decision module to redetermine the joint strategy probability distribution and the vehicle's game decision result. Additionally, it triggers the generative scenario deduction module to encode the redetermined joint strategy probability distribution and the vehicle's game decision result into an updated game condition vector, and performs driving scenario prediction again based on the updated game condition vector. The planned trajectory output module is data-connected to the generative scenario deduction module and is used to output the final vehicle planned trajectory generated after a predetermined number of iterations.

[0025] Preferably, the multimodal state encoding module includes a visual encoder, a text embedding layer, a visual-language projection layer, a multimodal large language model, an encoder of a variational autoencoder, and a traffic participant motion prediction head; the generative scene inference module includes a game-conditional projection layer, an inference decoding part of the multimodal large language model, a decoder of the variational autoencoder, and a decoding branch of the traffic participant motion prediction head.

[0026] Preferably, the planned trajectory output module is connected to the vehicle controller for sending the final planned trajectory of the vehicle to the vehicle controller, so that the vehicle controller can generate steering control commands, drive control commands and braking control commands based on the final planned trajectory of the vehicle.

[0027] The above technical solution has the following advantages: Compared with planning methods that directly generate vehicle trajectories based solely on multimodal inputs, this invention maps vehicles and traffic participants as game participants and uses joint strategy probability distributions to express the probabilistic relationships of strategy combinations among participants, enabling the joint influence of multiple traffic participants on vehicle decisions to be incorporated into the planning conditions.

[0028] The joint strategy probability distribution and the decision results of the autonomous vehicle game are encoded into game condition vectors, so that the generative scenario inference is subject to the condition constraints of the current interactive decision; the inferencer simultaneously generates the autonomous vehicle trajectory and the trajectory of traffic participants, so that the planning process can form multi-agent behavior information in the prediction time domain.

[0029] The dynamic payoff function is updated based on the deduced trajectory, the joint strategy probability distribution and the self-vehicle game decision results are obtained again, and the game condition vector is used to deduce again, so that the game decision and the scenario deduction form a feedback loop. This helps to reduce the inconsistency between the initial decision conditions and the future interaction trend, and helps to improve the interaction consistency and stability of the planned trajectory in complex interaction scenarios.

[0030] The feedback loop converts the collision frequency obtained from the previous round of scenario simulation into a payoff weight adjustment amount, so that the next round of quantum game and scenario simulation uses the updated decision conditions. This mechanism is used to improve the consistency between decision conditions and future interaction trends. Attached Figure Description

[0031] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings, wherein: Figure 1 The overall flowchart of the quantum game scenario deduction autonomous driving planning method provided in the embodiments of the present invention is shown. Detailed Implementation

[0032] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. The described embodiments are used to illustrate the technical concept and implementation process of the present invention, and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features of the various embodiments can be combined with each other.

[0033] Example 1 like Figure 1As shown, this embodiment provides a quantum game-theoretic scenario-based autonomous driving planning method. This method is applicable to roundabout entrances, traffic weaving within roundabouts, ramp merging, and other interactive driving scenarios involving competition for passage order. This embodiment takes a driving scenario with at least one surrounding traffic participant as the processing object, mapping the vehicle and selected surrounding traffic participants as game participants, and setting two candidate driving strategies—an acceleration strategy and a deceleration strategy—for each game participant.

[0034] The method in this embodiment is executed by a computing processing unit on the vehicle. This computing processing unit interacts with the onboard camera, vehicle state acquisition unit, and traffic participant state acquisition unit. Various input data are associated according to corresponding timestamps so that multi-view images, the vehicle's state, and the states of surrounding traffic participants correspond to the same processing time. The computing processing unit calls pre-stored model parameters and programs to execute four stages: multimodal state representation, quantum game decision-making, generative scenario deduction, and feedback iterative optimization.

[0035] Step S1: Multimodal state characterization At any moment Acquire a collection of multi-view images captured by the vehicle's onboard camera:

[0036] in, This represents the total number of camera viewpoints. For the first Each camera perspective at any time The captured images.

[0037] Get vehicle motion status The vehicle's motion state includes its three-dimensional position in the global coordinate system:

[0038] It also includes attitude angles consisting of roll, pitch, and yaw angles, and the vehicle's velocity vector:

[0039] And the vehicle's acceleration vector:

[0040] Get Time The set of motion states of surrounding traffic participants:

[0041] in, The number of surrounding traffic participants detected. The movement status of surrounding traffic participants includes categories 3D bounding box Historical movement trajectory Velocity vector and acceleration vector The three-dimensional bounding box includes at least width, height, length, and heading angle, and the historical motion trajectory is the trajectory of the traffic participant at any given time. A sequence of multiple trajectory points from the previous point.

[0042] Obtain scene semantic description And bicycle movement issues The scene semantic description includes the current scene category, key object information, and dynamic event descriptions; the vehicle action questions come from a pre-set question-and-answer set, used to guide the multimodal large language model to focus on the vehicle's passage decisions in the current scene.

[0043] Visual encoding is performed on the images from each camera's viewpoint. All camera viewpoints share the same model parameters of the Visual Transformer. The visual features of each camera's viewpoint are represented as follows:

[0044] By stitching together the visual features from each camera's perspective along the spatial dimension, multi-view visual features are obtained:

[0045] Scene semantic description And bicycle movement issues After tokenization and text embedding, we get:

[0046]

[0047] Multi-view visual features are projected through a visual-linguistic layer. The projected visual features are then concatenated with the text embeddings in the language embedding space of the multimodal large language model, and input into the multimodal large language model to obtain the cross-modal fusion result.

[0048] in, Represents the visual-linguistic projection layer. This represents the forward processing of a multimodal large language model. It describes the process from the cross-modal fusion results. Extract the hidden state corresponding to the vehicle. And extract the hidden states corresponding to each surrounding traffic participant. Each entity's hidden state is associated with its corresponding motion state through traffic participant identification, detection results, or sequence position; for simplified notation, the following... and Both represent entity fusion representations after being associated with the corresponding motion state.

[0049] Among them, Visual Transformer represents a visual encoder based on a self-attention mechanism, LLM represents the forward processing of a multimodal large language model, Concat represents concatenation along a specified dimension, and Project represents the visual-language projection operation.

[0050] Integrating the physical representation of the vehicle The encoder is a variational autoencoder. The encoder outputs the mean of the latent variables. And the logarithmic variance, and obtain the standard deviation from the logarithmic variance. And use a reparameterization method to generate the vehicle's hidden space token:

[0051] in:

[0052] The first Entity fusion representation of surrounding traffic participants The input traffic participant motion prediction head's encoding branch is a three-layer multilayer perceptron with shared weights among surrounding traffic participants. The hidden space token for each traffic participant is:

[0053] in, , and This is the weight matrix. , and This is the bias vector.

[0054] Where I is the identity matrix, ReLU represents the linear rectified activation function; t is the processing time index, v is the camera view index, and i is the index of surrounding traffic participants.

[0055] Step S1 outputs the vehicle's hidden space token. And the collection of hidden space tokens for traffic participants:

[0056] Step S2: Quantum Game Decision Making When there is only one traffic participant in the surrounding area, the vehicle and the traffic participant are identified as game participants, forming a two-player game. When there are at least two traffic participants in the surrounding area, the two traffic participants closest to the vehicle are selected based on the spatial distance between the vehicle and each of the surrounding traffic participants, and the vehicle and the two selected traffic participants are identified as game participants, forming a three-player game.

[0057] Let the set of game participants be:

[0058] Wherein, index 0 represents the vehicle, and indices 1 to... This indicates the selected surrounding traffic participants.

[0059] For roundabout or merging scenarios, each player has two candidate driving strategies: an acceleration strategy and a deceleration strategy. The latent space tokens of each game participant are input into the policy classification head, which determines the feasible driving strategies for that player among acceleration and deceleration strategies. The driving strategy space of each game participant is represented as:

[0060] in, and These correspond to acceleration and deceleration strategies, respectively.

[0061] The joint strategy space is:

[0062] The joint strategy profile is as follows:

[0063] For relatively close speeds The The first game participant and the second There are 10 game participants, and the collision time is:

[0064] No. Individual game participants in the joint strategy profile The security benefits are as follows:

[0065] in, For indicator functions, This is the collision time threshold. In this embodiment, Take 3 seconds.

[0066] No. Individual game participants in the joint strategy profile The following comfort benefits are:

[0067] in, For the first Individual game participants in the joint strategy profile The acceleration below, This is the standard deviation parameter for acceleration.

[0068] No. Individual game participants in the joint strategy profile The efficiency gains are as follows:

[0069] in, For the first Individual game participants in the joint strategy profile The expected time to reach the target location. Ideal arrival time.

[0070] The dynamic return function is:

[0071] in:

[0072] All weights are within the range In the roundabout or merging scenario of this embodiment, the safety benefit weighting... Set to 0.6, and the remaining weights are determined according to the preset configuration, provided that the total weight sum is 1.

[0073] The acceleration standard deviation parameter is taken as a positive value; when the expected arrival time is less than or equal to 0, the efficiency gain is 0. Each gain weight can be used as a model configuration parameter or a trainable parameter, and its initial value is determined under the condition that each weight is located in the interval [0,1] and the sum is 1.

[0074] Three-dimensional position and distance are expressed in meters, velocity in meters per second, acceleration in meters per second², and time in seconds. Spatial distance can be expressed as the Euclidean distance between the center point of the three-dimensional bounding box and the road plane. The expected arrival time can be determined based on the predicted trajectory length and the velocity along the trajectory. When other units are used in the input data, they should be converted to the above units first.

[0075] Map the acceleration strategy to the computational ground state Map the deceleration strategy to the computational ground state . No. The quantum states of each game participant are:

[0076] in, and Let be a complex probability amplitude, and satisfy:

[0077] The quantum states of all game participants form an initial joint quantum state:

[0078] Entanglement processing is performed on the initial joint quantum state using entanglement operators:

[0079] in, For entanglement parameters, To act on the first Each game participant corresponds to a Pauli-X (Pauli X) operator for a qubit. It is the imaginary unit.

[0080] No. Each game participant performs a transformation on its quantum state using a single-parameter unitary operator:

[0081] in: Under the premise of satisfying the quantum state normalization constraint and the quantum operation unitary constraint, the complex probability amplitude, entanglement parameter and unitary operator parameter can be used as trainable parameters of the quantum game model. They can be optimized according to the overall loss function during the training phase and the trained parameters can be called during the inference phase; or they can be determined according to a pre-set parameter table.

[0082]

[0083] After each participant completes the unitary transformation, an anti-entanglement operator is applied to the joint quantum state to obtain the final joint quantum state:

[0084] Joint Strategy Profile The corresponding probabilities are:

[0085] Traverse each joint policy profile to obtain the joint policy probability distribution. .

[0086] After obtaining the joint strategy probability distribution, the decision outcome of the vehicle game is determined in the following progressive order.

[0087] When the vehicle strategy For any other autonomous vehicle strategy and strategy profiles of any other game participants All satisfy:

[0088] Then The result is determined to be the outcome of the game between the vehicles.

[0089] In the absence of a strictly dominant strategy for the vehicle, determine whether a unique pure strategy Nash equilibrium exists. Joint strategy profile. For a pure strategy Nash equilibrium, the following holds true for each player in the game:

[0090] When the pure strategy Nash equilibrium is unique, the vehicle strategy in the pure strategy Nash equilibrium is determined as the vehicle game decision outcome.

[0091] When there is no pure-policy Nash equilibrium or multiple pure-policy Nash equilibria exist, the quantum expected utility of each vehicle's policy is:

[0092] The decision-making outcome of the self-driving game is: Where A_i represents the driving strategy space of the i-th player, A_-i represents the joint strategy space of the other players (excluding the i-th player), A_0 represents the driving strategy space of the vehicle, and A_-0 represents the joint strategy space of the other players (excluding the vehicle). When multiple vehicle strategies have the same quantum expected utility, one of the vehicle strategies is determined according to a pre-fixed arrangement in the vehicle driving strategy space A_0.

[0093]

[0094] Step S3: Generative Scene Derivation Joint strategy probability distribution Flatten the joint strategy profile into a one-dimensional probability vector according to a predetermined order, and then combine the one-dimensional probability vector with the decision result of the autonomous vehicle game. The one-hot encoded vectors are concatenated and input into two multilayer perceptron projection layers to generate game condition vectors:

[0095] in:

[0096] and Let be the projection matrix. and This is the bias vector. In this embodiment, the hidden layer dimension of the projection layer is 512, and the output dimension is 2048. The game condition vector remains unchanged in each deduction step of the same round of scenario deduction.

[0097] The predetermined order is the fixed arrangement order of each joint policy profile in the joint policy space; flatten means flattening the joint policy probability data into a one-dimensional vector according to the fixed order; onehot means one-hot encoding; and Embed means text embedding operation.

[0098] The prediction time domain consists of K equally spaced positive integer deduction times. The deduction step interval is consistent with the timestamp interval of the input data, and k is the deduction step index.

[0099] For the For each derivation step, the input sequence of the generative deriver is:

[0100] Generative inferential generators utilize multimodal large language models to process input sequences. Predicting the first The simulation steps include the vehicle's hidden space token, the hidden space tokens of each traffic participant, the scene semantic description, and the vehicle's action response.

[0101] The predicted latent space tokens of the vehicle are input into the decoder of the variational autoencoder to obtain the trajectory points of the vehicle at the corresponding extrapolation time. The predicted latent space tokens of each traffic participant are input into the decoding branch of the traffic participant motion prediction head to obtain the trajectory points and 3D bounding boxes of the corresponding traffic participants. The predicted latent space tokens of the vehicle and the latent space tokens of the traffic participants are used as inputs for the next extrapolation step until all scene extrapolations within the predetermined prediction time domain are completed.

[0102] The predicted scene semantic description and the vehicle's action response serve as the text context for the next derivation step; wherein, the predicted scene semantic description serves as the scene semantic description for the next derivation step, and the predicted vehicle action response is associated with the question template in the preset question and answer set to form the vehicle action question for the next derivation step.

[0103] Step S4: Feedback Iterative Optimization After completing one round of future scenario simulation, the behavioral evolution information in the time domain is statistically predicted based on the generated vehicle trajectory points, traffic participant trajectory points, and 3D bounding boxes. This embodiment at least calculates the collision frequency in the simulation trajectory. The intersection of the vehicle-occupied area and the traffic participant-occupied area at each simulation time is determined. When the two occupied areas overlap, the corresponding simulation time is recorded as a collision event; the collision frequency is the proportion of the number of simulation times corresponding to collision events to the total number of simulation times evaluated.

[0104] Update the safety benefit weights and efficiency benefit weights based on collision frequency:

[0105]

[0106] in, To update the step size, To extrapolate the collision frequency in the trajectory, in this embodiment, the update step size is used. The weight of comfort benefits remains unchanged at 0.05.

[0107] After updating the weights, the weights for safety gains, comfort gains, and efficiency gains are normalized so that all three weights fall within the specified range. The sum of the three weights is 1. Based on the updated payoff weights, the dynamic payoff function is reconstructed, the payoffs of each joint strategy profile are recalculated, and quantum game processing is re-executed to obtain the updated joint strategy probability distribution and the vehicle game decision results. The updated joint strategy probability distribution and the vehicle game decision results are re-encoded into an updated game condition vector, and future scenario deduction is performed again using the updated game condition vector.

[0108] The driving scenario prediction, payoff weight update, quantum game re-solution, and game condition vector update constitute one iteration. This embodiment performs three iterations. After completing the three iterations, the sequence of vehicle trajectory points generated in the last iteration is output as the final planned trajectory of the vehicle.

[0109] When no surrounding traffic participants are detected, no interactive game steps are performed; when the relative approach speed is less than or equal to 0, the collision time of the corresponding participant pair is not calculated; when a certain simulation step does not obtain a valid model output, the current simulation is terminated and the most recently obtained valid vehicle trajectory is used.

[0110] Example 2 This embodiment illustrates the training process of multimodal latent space tokens and generative inferential generators. Training samples include time intervals. The data includes multi-view camera image sequences, vehicle motion status, motion status of surrounding traffic participants, scene semantic description, vehicle action issues, real vehicle future trajectory, real traffic participant future trajectory, traffic participant categories and 3D bounding box labels, with each data item aligned according to scene identifiers and timestamps.

[0111] The total pre-training loss is:

[0112] The planning loss is:

[0113] The motion prediction loss is:

[0114] The language prediction loss is:

[0115] An auxiliary detection head can also be set. The auxiliary detection head receives latent space tokens from traffic participants and outputs the 3D bounding box parameters at the current time. The predicted result is compared with the actual 3D bounding box to calculate the L1 loss.

[0116] After completing the latent space token pre-training, the generative inferrer undergoes multi-step unfolding training. The temporal consistency loss is:

[0117] in, and The first In the first deduction step, the predicted latent space distribution and the reference latent space distribution are in the [number]th step. The probability values ​​of each dimension.

[0118] The trajectory supervision loss is:

[0119] In training samples with joint policy labels, the policy classification head generates a joint policy profile based on the generated trajectory and the corresponding latent space token output. Its payoff vector is: u(a t+Δ ) = [u 0 (a t+Δ ), u 1 (a t+Δ ), …, u n 1 (a t+Δ )] T The expected payoff vector in quantum games is: u * = [EU 0 , EU 1 , …, EU n 1 ] T The profit deviation loss is:

[0120] The overall loss from multi-step simulation training is:

[0121] in, The weighting coefficients for the gain bias loss are used. After model training is complete, the model parameters of the visual encoder, text embedding layer, multimodal large language model, variational autoencoder, traffic participant motion prediction head, policy classification head, and game conditional projection layer are stored in the vehicle computing unit.

[0122] Where K is the number of future moments in the single-step trajectory prediction time domain, M is the text sequence length, T is the number of inference steps in multi-step expansion, Δ is the inference step index, and r is the latent space distribution dimension index; the capped symbol represents the predicted value, and the uncapped symbol represents the corresponding reference value; Coll represents the collision penalty function, Focal represents the focus loss used to handle class imbalance, and L1 loss represents the absolute error loss; λ1 to λ5 and λu are non-negative loss weights. The probability inputs in the cross-entropy loss and temporal consistency loss are taken as positive values, and numerical truncation commonly used in this field can be used during calculation to avoid the logarithmic input being 0.

[0123] Example 3 This embodiment illustrates how game participants use discrete quantum gates to perform quantum strategy transformations. The other steps in this embodiment are the same as in Embodiment 1.

[0124] The set of discrete quantum gates is:

[0125] in:

[0126]

[0127] Let the first The quantum gate determined by each game participant is The final joint quantum state is:

[0128] By measuring the final joint quantum state, the probabilities of each joint strategy profile are obtained:

[0129] The rules for determining discrete quantum gates are as follows: based on the candidate policy categories output by the policy classification head, the corresponding quantum gate is determined according to the fixed mapping relationship established in the discrete quantum gate set beforehand; the selected quantum gate satisfies the unitary condition.

[0130] Example 4 When a nearby vehicle with a direct interactive effect on the player's vehicle is detected in the merging area, the player's vehicle is identified as Player 0, and the nearby vehicle is identified as Player 1. Both players' driving strategy spaces include acceleration and deceleration strategies.

[0131] Where strategy 0 represents acceleration and strategy 1 represents deceleration. The joint policy space is:

[0132] After entanglement, player unitary transformation, and deentanglement, the final joint quantum state is:

[0133] The probability distribution of the joint strategy is as follows:

[0134] A one-dimensional probability vector of length 4 is concatenated with a one-hot encoded vector of length 2 for the vehicle's strategy, and then a game condition vector is generated through a game condition projection layer. The generative inferrer progressively outputs the vehicle's trajectory and the trajectories of surrounding vehicles; when the inference results indicate that the occupied areas of the two vehicles overlap in the prediction time domain, the safety benefit weight is increased and the efficiency benefit weight is decreased based on the collision frequency, the benefit data is reconstructed, the quantum game is solved again, and the future trajectory is generated again.

[0135] Example 5 When a vehicle approaches the roundabout entrance or is driving inside the roundabout, select the two traffic participants closest to the vehicle from the detected surrounding traffic participants. The vehicle is identified as Player 0, and the two traffic participants are identified as Player 1 and Player 2, respectively.

[0136] Each player has an acceleration strategy and a deceleration strategy. The joint strategy space of the three players contains eight joint strategy profiles. 0 represents an acceleration strategy, and 1 represents a deceleration strategy. The joint strategy probability data is represented in a predetermined order as follows:

[0137] The initial joint quantum state is:

[0138] The final joint quantum state is:

[0139] Self-driving strategy The quantum expected utility is:

[0140] The game condition vector is generated by concatenating the joint policy probability vector of length 8 with the one-hot encoded vector of the vehicle's policy. The generative inferrer simultaneously predicts the future trajectories of the vehicle and the two traffic participants based on the game condition vector, updates the payoff weights according to the collision event, resolves the three-player quantum game, and executes the next round of inference.

[0141] Example 6 This embodiment provides a quantum game scenario simulation autonomous driving planning system. The system includes a multimodal data acquisition module, a multimodal state encoding module, a quantum game decision-making module, a generative scenario simulation module, a feedback update module, and a planned trajectory output module.

[0142] The multimodal data acquisition module interacts with the vehicle-mounted camera, vehicle status acquisition unit, and traffic participant status acquisition unit to acquire multi-view camera image sequences, scene semantic descriptions, vehicle action questions, vehicle motion status, and the motion status of surrounding traffic participants, and associates various types of input data according to timestamps.

[0143] The multimodal state encoding module includes a visual encoder, a text embedding layer, a visual-linguistic projection layer, a multimodal large language model, an encoder of a variational autoencoder, and a traffic participant motion prediction head. The visual encoder encodes images from each camera's viewpoint; the text embedding layer encodes scene semantic descriptions and vehicle motion questions; the visual-linguistic projection layer projects visual features into the linguistic embedding space; the multimodal large language model fuses visual and linguistic features to form entity representations corresponding to the vehicle and traffic participants; the encoder of the variational autoencoder generates latent space tokens for the vehicle; and the encoding branch of the traffic participant motion prediction head generates latent space tokens for each traffic participant.

[0144] The quantum game decision-making module receives latent space tokens from both the vehicle and other traffic participants. These tokens are used to determine the game participants, driving strategy space, and dynamic payoff function, forming a joint quantum state. Entanglement processing, unitary transformation of the game participants, and deentanglement processing are then performed sequentially. Finally, the joint quantum state is measured to obtain the joint strategy probability distribution. The quantum game decision-making module is also used to determine the vehicle's game decision outcome based on a strictly dominant strategy, a unique pure strategy Nash equilibrium, or quantum expected utility.

[0145] The generative scene extrapolation module includes a game-conditional projection layer, an extrapolation decoding part of a multimodal large language model, a decoder of a variational autoencoder, and a decoding branch of a traffic participant motion prediction head. The game-conditional projection layer converts the joint policy probability distribution and the vehicle's game decision results into game-conditional vectors; the multimodal large language model predicts the latent space tokens and scene semantics of the next extrapolation step based on the game-conditional vectors and each latent space token; the decoder of the variational autoencoder outputs the vehicle's trajectory points; and the decoding branch of the traffic participant motion prediction head outputs the traffic participant's trajectory points and 3D bounding boxes.

[0146] The feedback update module receives the vehicle trajectory and traffic participant trajectories output by the generative scenario simulation module. It uses these to statistically predict the collision frequency in the time domain and updates the payoff weights of the dynamic payoff function based on the collision frequency. The feedback update module then sends the updated dynamic payoff function to the quantum game decision module to trigger a re-solution of the quantum game. The joint strategy probability distribution and the vehicle game decision results obtained from the re-solution are sent to the generative scenario simulation module to generate an updated game condition vector and execute the driving scenario simulation again.

[0147] The trajectory planning output module receives the sequence of vehicle trajectory points generated after a predetermined number of iterations and outputs it as the final planned trajectory. The final planned trajectory is sent to the vehicle controller, which generates steering control commands, drive control commands, and braking control commands based on the planned trajectory.

[0148] Each module can be implemented by different program functional units within the same computing processing unit, or by multiple computing processing units capable of data interaction. Modules exchange images, motion states, latent space tokens, joint policy probability distributions, game condition vectors, and planned trajectories through data transmission interfaces.

Claims

1. A quantum game scenario deduction method for autonomous driving planning, characterized in that, include: Acquire multi-view camera image sequences, scene semantic descriptions, vehicle action questions, vehicle motion states, and the motion states of surrounding traffic participants. Extract visual features from the multi-view camera image sequences and linguistic features from the scene semantic descriptions and vehicle action questions. Fuse the visual features, linguistic features, vehicle motion states, and the motion states of surrounding traffic participants using a multimodal large language model to generate latent space tokens that represent the motion states of the vehicle and each surrounding traffic participant. At least one traffic participant is identified from the surrounding traffic participants, and the vehicle and the identified traffic participant are taken as game participants. A driving strategy space including two candidate driving strategies is determined for each game participant. An interactive game model with a dynamic payoff function is constructed based on the latent space token and the driving strategy space. Each player's two candidate driving strategies are mapped to two computational ground states of a qubit. The quantum states of each player form a joint quantum state. Entanglement operators, unitary operators of each player, and anti-entanglement operators are sequentially applied to the joint quantum state. The final joint quantum state after applying the anti-entanglement operator is measured to obtain the joint strategy probability distribution. The decision result of the self-vehicle game is determined based on the joint strategy probability distribution and the dynamic payoff function. The joint strategy probability distribution and the vehicle game decision result are encoded into a game condition vector. The game condition vector and the latent space token are input into a generative inference engine to predict the driving scenario at multiple future moments through autoregression, and generate the vehicle trajectory points and the trajectory points of each traffic participant at the corresponding moments. Based on the generated vehicle trajectory points and the trajectory points of each traffic participant, the behavioral evolution information in the prediction time domain is determined. The dynamic reward function is updated using the behavioral evolution information. Based on the updated dynamic reward function, quantum game processing is re-executed to obtain the updated joint strategy probability distribution and vehicle game decision results. The updated joint strategy probability distribution and vehicle game decision results are encoded into an updated game condition vector. Based on the updated game condition vector, driving scenario prediction is performed again. After iteration, the final vehicle planned trajectory is output.

2. The quantum game scenario deduction autonomous driving planning method according to claim 1, characterized in that, The vehicle's motion state includes its three-dimensional position, attitude angle, velocity vector, and acceleration vector in the global coordinate system; the motion state of each surrounding traffic participant includes the traffic participant category, three-dimensional bounding box, historical motion trajectory, velocity vector, and acceleration vector; the scene semantic description includes scene category, key object information, and dynamic event description; the vehicle action question is used to guide the multimodal large language model to generate the vehicle driving decision.

3. The quantum game scenario deduction autonomous driving planning method according to claim 2, characterized in that, The first Images from the perspective of a camera Input a visual Transformer with shared weights for each camera viewpoint to obtain the corresponding visual features. and in accordance with Multi-perspective visual features are formed along the spatial dimension, among which, The total number of camera viewpoints; the scene semantic description and the vehicle's motion problem are encoded into text embeddings. and in accordance with Cross-modal fusion is performed, in which, This indicates that the multi-view visual features are projected onto the visual-linguistic projection layer of the language embedding space. This indicates that the multimodal large language model, from the cross-modal fusion results Extract the initial representation of the vehicle token and the initial representation of the tokens of each surrounding traffic participant; where t is the current processing time, v is the camera view index, and v takes values ​​from 1 to V.

4. The quantum game scenario deduction autonomous driving planning method according to claim 3, characterized in that, The initial representation of the vehicle token is input into the encoder of the variational autoencoder to obtain the mean value. And the logarithmic variance, and obtain the standard deviation based on the logarithmic variance. ,according to Generate a hidden space token for the vehicle; Initial representation of tokens from surrounding traffic participants The input traffic participant motion prediction head's encoding branch is a three-layer multilayer perceptron with shared weights among surrounding traffic participants, and is processed according to... Generate corresponding hidden space tokens for traffic participants, where... , , This is the weight matrix of the three-layer multilayer perceptron. , , Let t be the corresponding bias vector; where t is the current processing time, i is the index of the surrounding traffic participants, I is the identity matrix, and ReLU represents the linear rectified activation function.

5. The quantum game scenario deduction autonomous driving planning method according to claim 1, characterized in that, When there are at least two surrounding traffic participants, the two closest surrounding traffic participants to the vehicle are selected to form a three-player game with the vehicle; when there is only one surrounding traffic participant, the surrounding traffic participant and the vehicle form a two-player game. The method is applied to roundabout scenarios or merging scenarios. Each player has two candidate driving strategies, namely an acceleration strategy and a deceleration strategy. The latent space tokens of each player are input into the strategy classification head, and the strategy classification head determines the feasible driving strategy of the corresponding player from the acceleration strategy and deceleration strategy.

6. The quantum game scenario deduction autonomous driving planning method according to claim 5, characterized in that, For relatively close speeds The The first game participant and the second Each participant in the game, according to Determine the collision time, where, For the two game participants in terms of strategy and strategy The distance below; for the joint strategy profile , No. The security benefit for each game participant is The comfort benefits are Efficiency gains are ,in, For indicator functions, The collision time threshold, For the first The acceleration of each participant in the game. For the acceleration standard deviation parameter, For the expected arrival time, The ideal arrival time; the dynamic revenue function is ,in, , and All are located in the interval And satisfy Where i and j are the game participant indices, a is the joint strategy profile, and the acceleration standard deviation parameter is greater than 0; the efficiency gain is calculated according to the above formula only when the expected arrival time is greater than 0, and the efficiency gain is 0 when the expected arrival time is less than or equal to 0.

7. The quantum game scenario deduction autonomous driving planning method according to claim 6, characterized in that, For the first Each game participant is assigned a qubit, and the acceleration and deceleration strategies are mapped to the computational ground state, respectively. and and with Indicates the first The quantum states of each participant in the game, where... and Let be a complex probability amplitude, and satisfy . ;according to form The initial joint quantum state of each game participant, and through entanglement operators Entanglement processing is performed on the initial joint quantum state, wherein, For entanglement parameters, To act on the first Each game participant corresponds to a Pauli-X operator for a qubit. is the imaginary unit; n is an integer not less than 2, and i is the index of the game participant, with a value from 0 to n-1.

8. The quantum game scenario deduction autonomous driving planning method according to claim 7, characterized in that, No. The unitary operators of each game participant are ,in, ;according to Obtain the final joint quantum state and cross-reference the joint policy in the final joint quantum state. Modulus square of corresponding component The probability of being identified as the joint strategy profile.

9. The quantum game scenario deduction autonomous driving planning method according to claim 8, characterized in that, Determining the decision outcome of the self-vehicle game includes: determining whether there exists a strictly dominant strategy for the self-vehicle, when the self-vehicle strategy... For any other autonomous vehicle strategy and strategy profiles of any other game participants All meet At that time, the aforementioned autonomous vehicle strategy The outcome of the game is determined as the outcome of the self-vehicle game; when there is no strictly dominant strategy for the self-vehicle, determine whether there exists a unique pure strategy Nash equilibrium, when the joint strategy profile... For each player in the game, the following conditions are met: When the pure strategy Nash equilibrium is unique, the vehicle's strategy in that pure strategy Nash equilibrium is determined as the vehicle's game decision outcome; when no pure strategy Nash equilibrium exists or multiple pure strategy Nash equilibria exist, the strategy is determined according to... Calculate the quantum expected utility of each vehicle's strategy, and according to... Determine the decision outcome of the self-vehicle game; where A_i is the driving strategy space of the i-th player, A_-i is the joint strategy space of the other players besides the i-th player, A_0 is the driving strategy space of the self-vehicle, and A_-0 is the joint strategy space of the other players besides the self-vehicle; when the quantum expected utility of multiple self-vehicle strategies is the same, determine the decision outcome of the self-vehicle game according to the pre-fixed arrangement order in the self-vehicle driving strategy space A_0.

10. The quantum game scenario deduction autonomous driving planning method according to claim 9, characterized in that, The joint policy probability data representing the probabilities of each joint policy profile. Flattened into a one-dimensional probability vector, the one-dimensional probability vector is then compared with the game decision result of the autonomous vehicle. The one-hot encoded vectors are concatenated and then... Generate the game condition vector, where, , and For the projection matrix, and The corresponding bias vector is used, and the game condition vector remains unchanged in each inference step of the same round of driving scenario prediction; wherein, the joint strategy probability data is flattened according to the fixed arrangement order of each joint strategy profile in the joint strategy space.

11. The quantum game scenario deduction autonomous driving planning method according to claim 10, characterized in that, For the Each step of the deduction process, according to Construct the input sequence of the generative inferrer, wherein, For the vehicle's hidden space token, A hidden space token for all traffic participants. Let be the game condition vector. For scene semantic description, For the problem of vehicle motion; the multimodal large language model predicts the first... The latent space tokens of the vehicle, each traffic participant, the scene semantic description, and the vehicle's action response for each inference step are used. The decoder of the variational autoencoder decodes the predicted latent space tokens of the vehicle into vehicle trajectory points. The decoding branch of the traffic participant motion prediction head decodes the predicted latent space tokens of the traffic participants into trajectory points and 3D bounding boxes for the corresponding traffic participants. The predicted latent space tokens of the vehicle and the latent space tokens of the traffic participants are then used as the first inference step. The input for each inference step is used, and the predicted scene semantic description and the vehicle's action response are used as the text context for the next inference step to construct the scene semantic description and vehicle action question for the next inference step; where k is a non-negative integer and Embed represents the text embedding operation.

12. The quantum game scenario deduction autonomous driving planning method according to claim 11, characterized in that, The collision frequency in the prediction time domain is determined based on the generated vehicle trajectory points and the trajectory points of each traffic participant. ,according to and The safety and efficiency benefit weights are updated, while the comfort benefit weight remains unchanged in the update formula. Then, the updated safety, comfort, and efficiency benefit weights are normalized so that the sum of the three benefit weights is 1. To update the step size, the update step size is set such that both the updated safety gain weight and efficiency gain weight are within the interval [range]. The dynamic payoff function is reconstructed using the updated payoff weights, the autonomous vehicle game decision result and game condition vector are redefined, and driving scenario prediction is performed again based on the redefined game condition vector. The driving scenario prediction, payoff weight update, and redefined autonomous vehicle game decision result and game condition vector constitute one iteration, and three iterations are performed. The trajectory set {τ} consists of the autonomous vehicle trajectory and the trajectory of each traffic participant generated in the prediction time domain, and CollisionRate({τ}) is the proportion of the number of simulation times corresponding to the collision event to the total number of simulation times evaluated.

13. A quantum game scenario deduction autonomous driving planning system, characterized in that, The system includes a multimodal data acquisition module, a multimodal state encoding module, a quantum game decision-making module, a generative scene deduction module, a feedback update module, and a planned trajectory output module. The multimodal data acquisition module is connected to the vehicle-mounted camera, the vehicle state acquisition unit, and the traffic participant state acquisition unit, respectively, and is used to acquire multi-view camera image sequences, scene semantic descriptions, vehicle action problems, vehicle motion states, and the motion states of surrounding traffic participants, and send the data associated with timestamps to the multimodal state encoding module. The multimodal state encoding module is used to extract visual features and linguistic features, fuse the visual features, linguistic features, vehicle motion states, and the motion states of surrounding traffic participants through a multimodal large language model, and generate latent space tokens that respectively represent the motion states of the vehicle and each surrounding traffic participant. The quantum game decision-making module is data-connected to the multimodal state encoding module. It determines the game participants, driving strategy space, and dynamic payoff function based on the latent space token, forming a joint quantum state. Entanglement processing, unitary transformation of each game participant, and anti-entanglement processing are then performed sequentially. The final joint quantum state is measured to obtain the joint strategy probability distribution. The vehicle's game decision result is determined based on the joint strategy probability distribution and the dynamic payoff function. The generative scenario deduction module is data-connected to both the multimodal state encoding module and the quantum game decision-making module. It encodes the joint strategy probability distribution and the vehicle's game decision result into a game condition vector, and uses the game condition vector and the latent space token to autoregressively predict future scenarios. The system generates vehicle trajectory points and trajectory points of each traffic participant for driving scenarios at multiple time points. The feedback update module is connected to the generative scenario inference module and the quantum game decision module. It is used to determine the behavioral evolution information in the prediction time domain based on the generated vehicle trajectory points and trajectory points of each traffic participant. It updates the dynamic payoff function using the behavioral evolution information and triggers the quantum game decision module to redetermine the joint strategy probability distribution and the vehicle game decision result. It also triggers the generative scenario inference module to encode the redetermined joint strategy probability distribution and the vehicle game decision result into an updated game condition vector, and performs driving scenario prediction again based on the updated game condition vector. The planned trajectory output module is connected to the generative scene deduction module and is used to output the final planned trajectory of the vehicle after completing a predetermined number of iterations.

14. The quantum game scenario deduction autonomous driving planning system according to claim 13, characterized in that, The multimodal state encoding module includes a visual encoder, a text embedding layer, a visual-language projection layer, a multimodal large language model, an encoder of a variational autoencoder, and a traffic participant motion prediction head; the generative scene inference module includes a game-conditional projection layer, an inference decoding part of the multimodal large language model, a decoder of the variational autoencoder, and a decoding branch of the traffic participant motion prediction head.

15. The quantum game scenario deduction autonomous driving planning system according to claim 13, characterized in that, The planned trajectory output module is connected to the vehicle controller and is used to send the final planned trajectory of the vehicle to the vehicle controller, so that the vehicle controller can generate steering control commands, drive control commands and braking control commands based on the final planned trajectory of the vehicle.

Images