A decision planning method based on scene oscillation synchronization and context reinforcement adaptation

Abstract

The application is suitable for the technical field of automatic driving decision planning, and provides a decision planning method based on scene oscillation synchronization and situation reinforcement adaptation, comprising the following steps: multi-dimensional scene embedding and adaptive fusion coding, cooperatively coding the input of multi-source sensors to generate high-dimensional scene embedding features; scene oscillation synchronization, based on the high-dimensional scene embedding features, constructing a dynamic synchronization mechanism among multiple agents to generate a synchronization feature vector; integrated scene planning decoding, based on the high-dimensional scene embedding features and the synchronization feature vector, generating a trajectory sequence conforming to the constraints of dynamics and smoothness; situation reinforcement adaptation, based on a situation-driven reinforcement learning mechanism, multi-objective optimizing the trajectory sequence to realize the dynamic balance of safety, efficiency and comfort. The method can realize real-time collaborative evolution of decision planning and scene dynamics, and significantly improve the safety, smoothness and environmental adaptability of the automatic driving system.

Description

Technical Field

[0001] This invention belongs to the field of autonomous driving decision planning technology, and particularly relates to a decision planning method based on scene oscillation synchronization and context reinforcement adaptation. Background Technology

[0002] Current autonomous driving systems typically consist of four main modules: perception, prediction, planning, and control. Traditionally, they employ a hierarchical architecture: the perception module extracts environmental information from multimodal sensors such as cameras and LiDAR; the prediction module infers the future state of surrounding traffic participants; the planning module generates the vehicle's driving path; and the control module executes specific commands. While this modular design facilitates engineering implementation, the fragmented interfaces between modules lead to information transmission delays and accuracy degradation. The independence of prediction and planning prevents the system from reflecting the dynamic feedback of the vehicle's decisions to the environment, creating a bottleneck in the sequential "perception-prediction-planning" process. Furthermore, the fixed-dimensional interface format lacks differentiable structures, limiting the possibility of global optimization, especially in complex scenarios where the system struggles to effectively capture the dynamic impact of the vehicle's behavior on the surrounding traffic.

[0003] To overcome the limitations of hierarchical architectures, research has gradually shifted towards an integrated decision-making and planning architecture. This architecture incorporates environmental understanding, motion prediction, and trajectory planning into a unified framework, using end-to-end deep neural networks to achieve overall optimization from sensor input to trajectory output, thereby reducing information loss and interface errors. However, this integrated structure still faces the high complexity of multimodal fusion. Differences in spatial alignment, temporal synchronization, and semantic scale of sensor data lead to high computational overhead and alignment errors in feature fusion. Simultaneously, existing models are insufficient in modeling the dynamic interactions between multiple agents, often relying on attention mechanisms or graph neural networks to describe the interactions of neighboring vehicles, but failing to capture the nonlinear coupling and dynamic cooperative characteristics of multi-agent systems. Furthermore, the static optimization design of the policy layer makes it difficult for the model to adaptively adjust under different traffic densities, weather conditions, and risks, often resulting in an imbalance between safety, efficiency, and comfort.

[0004] To address these issues, researchers have explored the combination of multi-objective reinforcement learning (MORL) and swarm oscillation synchronization theory. MORL achieves a dynamic balance between safety, efficiency, and comfort by finding Pareto optimal solutions among multiple optimization objectives; swarm dynamics, derived from nonlinear dynamics, describes the cooperative behavior of multiple agents through oscillatory phase coupling and frequency synchronization, and is widely used in neuroscience and swarm dynamics systems. However, in the field of autonomous driving, existing research mostly uses both independently, failing to achieve deep integration. MORL focuses on individual policy optimization and lacks swarm interaction modeling; while synchronization theory, although capable of characterizing complex coupling relationships, does not consider the bidirectional coupling between policy learning and environmental feedback, making it difficult to form a closed-loop optimization structure. In particular, existing models have not yet established a joint driving mechanism between synchronization features and contextual information, resulting in a gap between swarm dynamic modeling and policy evolution, making it difficult to achieve adaptive cooperative decision-making in real traffic scenarios. Summary of the Invention

[0005] The purpose of this invention is to provide a decision planning method based on scene oscillation synchronization and context reinforcement adaptation, which aims to solve the problems mentioned in the background art.

[0006] The present invention is implemented as follows: a decision planning method based on scene oscillation synchronization and context reinforcement adaptation includes the following steps:

[0007] Step 1: Multi-dimensional scene embedding and adaptive fusion coding: Co-encode multi-source sensor inputs to generate high-dimensional scene embedding features;

[0008] Step 2: Scene oscillation synchronization. Based on the high-dimensional scene embedding features, a dynamic synchronization mechanism among multiple agents is constructed to generate a synchronization feature vector.

[0009] Step 3: Integrated scene planning and decoding: Based on the high-dimensional scene embedding features and synchronization feature vectors, generate a trajectory sequence that meets the constraints of dynamics and smoothness.

[0010] Step 4: Context-driven reinforcement adaptation. Based on the context-driven reinforcement learning mechanism, the trajectory sequence is optimized for multiple objectives to achieve a dynamic balance between safety, efficiency and comfort.

[0011] A further technical solution, the specific steps of step 1 are as follows:

[0012] A unified representation framework is used to collaboratively encode multi-source sensor inputs, generating a high-dimensional, consistent scene representation; the input data includes visual features. Vehicle body condition semantic structure and environmental characteristics Formalization First, a weighted attention mechanism is used to allocate modal weights to generate multimodal fusion features. :

[0013] ;

[0014] in, For modal adaptive weights, This is the feature extraction function for the corresponding mode;

[0015] Introducing time embedding and spatial embedding A spatiotemporal joint mapping is established using a multilayer perceptron (MLP). :

[0016] ;

[0017] in, This indicates that the feature concatenation operation enables synchronous modeling in semantics and space, ensuring the temporal alignment of encoded features in dynamic environments; This represents the mapping function of a multilayer perceptron, which consists of a linear transformation layer and a nonlinear activation function, and is used to realize high-dimensional nonlinear mapping and interaction of input features;

[0018] Building upon this, by leveraging residual structures and nonlinear mappings, and through a layered approach of activation functions and prior transformations, the nonlinear coupling capability between modes is enhanced, generating preliminary high-dimensional features capable of capturing potential multimodal correlation patterns and dynamic coupling relationships. :

[0019] ;

[0020] in, It is the Sigmoid activation function. and It is a linear transformation matrix. and These are bias terms, all of which are learnable parameters; This indicates a modified linear unit activation function;

[0021] Finally, based on the spatiotemporal embedding features, an energy mapping and gating control mechanism is introduced. The information flow of different modalities is regarded as an energy channel, and the fusion intensity is controlled by the change of energy state to generate the final unified high-dimensional scene embedding features. :

[0022] ;

[0023] in, and These are the energy transformation matrix and the gate control matrix, respectively, used to allocate fusion weights among the feature channels; and For the corresponding bias term; This is the energy transformation branch, used to capture the potential energy flow relationships between modes; This is the gated modulation factor, used to adaptively control the contribution ratio of each mode according to the complexity of the scene; This is the scene energy mapping function, used to model the impact of traffic density, risk level, and environmental disturbances on the fusion strength; This indicates the Hadamard element-wise multiplication operation, which is the multiplication between corresponding elements and is used to achieve point-to-point modulation between the gated signal and the energy map.

[0024] In a further technical solution, step 2 includes the following specific steps:

[0025] Based on the high-dimensional scene embedding features generated in step 1 An improved multi-stage Kuramoto model is designed, and a dynamic synchronization mechanism is constructed to characterize the nonlinear interaction and cooperative behavior among multiple agents; the system is assumed to include There are 3 agents, each of which is considered a phase oscillator with a phase of 1 / 2. , indicating the th continuous time interval The phase variables of each agent, with natural frequencies of... The basic phase evolution equation is:

[0026] ;

[0027] in, This is an adjacency relation matrix. For global coupling strength, For the first An intelligent agent at time The phase variable;

[0028] To adapt to complex traffic scenarios, a multi-order synchronization model is adopted, introducing a second-order coupling term:

[0029] ;

[0030] in, It is a second-order coupling matrix. It is a second-order synchronous gain, used to enhance the ability to model complex nonlinear phase relationships;

[0031] Global synchronization stability is measured by synchronization metrics. Evaluate:

[0032] ;

[0033] in, The average phase under steady state. The closer the value is to 1, the more synchronized the system is. Indicates the first continuous time interval Phase variables of each agent;

[0034] To avoid excessive local synchronization, an adaptive coupling coefficient was designed:

[0035] ;

[0036] in, For intelligent agents The local synchronization degree is calculated by averaging the phases of the local neighborhood. To adjust the coefficients, the coupling strength in locally asynchronous regions is dynamically enhanced; The initial coupling strength represents the default coupling level of the system without synchronization adjustment. Represents intelligent agents The local coupling strength is used to characterize the agent's ability to synchronously interact with individuals in its domain;

[0037] Furthermore, a dynamic adjacency weight update mechanism is proposed to adjust the interaction intensity based on spatial distance and velocity difference:

[0038] ;

[0039] in, Indicates the current time step; Indicates the time step, which is the sampling period over continuous time and is used to control the update frequency of the adjacency weights; This represents the update rate coefficient, used to control the fusion ratio of historical adjacency relationships and current interaction features; Indicates the current time step Time-based intelligent agents and The adjacency weights reflect the interaction between the two elements, and are derived from the adjacency matrix. Expanded to a time-related form, Defined as a multivariable Gaussian kernel function:

[0040] ;

[0041] in, and For intelligent agents and Spatial distance and speed difference between them; For the current time step Traffic density; For reference density; These are road topological features used to quantify lane connectivity or geometric constraints. , , , The bandwidth parameters for each variable control the sensitivity of weight updates;

[0042] Meanwhile, the current time step is obtained after statistical averaging. Synchronous feature vector , as input guides for subsequent steps, where, For the current time step Global synchronization degree.

[0043] In a further technical solution, step 3 includes the following specific steps:

[0044] Based on high-dimensional scene embedding features and time context vector Generate trajectory sequences that conform to dynamic constraints and physical continuity. ,in Indicates the current time step The lateral position of the vehicle in the Cartesian coordinate system Indicates vertical position. Indicates the time length of the trajectory;

[0045] An integrated decoding structure is established within a unified feature space, and the trajectory generation process is realized through recursive relationships:

[0046] ;

[0047] in, Indicates the current time step The state vector at that time; This is a decoder function used to achieve differentiable prediction and continuous evolution of the trajectory in the time dimension; and These are the vehicle's speed and heading angle, respectively.

[0048] To ensure semantic consistency in the trajectory generation process across different modalities, a cross-modal collaborative constraint was designed to correct the differences between the visual and semantic modalities, preventing perceptual biases from accumulating and amplifying at the decoding layer. The goal of this constraint is to maintain decoding consistency of multi-source features within the high-dimensional embedding space, thereby achieving collaborative consistency between the perception and planning processes.

[0049] ;

[0050] in, and These are feature mapping functions for visual and semantic modalities, respectively; This represents cross-modal consistency loss, used to constrain the consistency between visual and semantic modalities in a high-dimensional feature space;

[0051] To ensure the geometric continuity and smoothness of the trajectory, curvature smoothing regularization constraints are introduced during the decoding process. By limiting the rate of change of curvature, sharp turns and non-physical oscillations are prevented from occurring in the generated path.

[0052] ;

[0053] in, For the trajectory at the current time step Curvature at time; This represents the curvature loss smoothing term, used to constrain the smoothness of trajectory geometric changes; Indicates that the vehicle is in The velocity component in the direction, Indicates that the vehicle is in The velocity component in the direction, and This represents the acceleration component in the corresponding direction;

[0054] At the dynamics level, to ensure that the trajectory output conforms to the vehicle's physical motion characteristics, a constraint model based on the vehicle's kinematic equations is established:

[0055] ;

[0056] Considering the above constraints, a trajectory optimization objective function is defined, and multimodal consistency, geometric smoothness, and dynamic feasibility are jointly optimized:

[0057] ;

[0058] in, For dynamic consistency loss, , , These are adjustable weighting coefficients used to balance various optimization objectives. This represents the overall optimization loss for trajectory generation;

[0059] Based on this, a synchronization-context guidance term was designed. By introducing scene synchronization features and context factors to jointly modulate the decoding process, trajectory generation is jointly driven by the scene group synchronization state and environmental risk factors.

[0060] ;

[0061] in, This is a synchronization-context guidance function used to map the group phase offset and context weight feedback to the trajectory update amount; For guiding strength coefficient, Represents contextual factors.

[0062] In a further technical solution, step 4 includes the following specific steps:

[0063] A dynamic balance between safety, efficiency, and comfort is achieved through a context-driven reinforcement learning optimization mechanism; defining states. ,action ,Strategy and reward function The optimization objective is:

[0064] ;

[0065] in, As a discount factor, This represents the expected cumulative reward function. Indicates policy-based Expectation operation, Indicates the current time step The reward discount factor;

[0066] Design Context-Enhanced Reward Model :

[0067] ;

[0068] in, Weight each objective accordingly, balancing safety, efficiency, comfort, and group synchronization; For synchronous feedback gain; This indicates a safety bonus, which measures the minimum distance between the vehicle and surrounding traffic elements, the safety margin, and the risk of collision. This indicates efficiency-related rewards, and assesses average speed, task completion rate, and travel time. This indicates a comfort-related bonus designed to improve ride smoothness.

[0069] The reward weights are updated through a dynamic adjustment mechanism:

[0070] ;

[0071] in, The learning rate ensures that the weights adjust adaptively.

[0072] Policy updates employ near-end policy optimization (PPO):

[0073] ;

[0074] in, This is the advantage estimate; To truncate parameters; This represents the pruning loss function, used to prevent large policy updates from causing training instability; This represents the expectation evaluation operation on empirical samples;

[0075] To achieve multi-objective balance, Pareto boundary constraints are introduced to ensure consistency of gradient directions across multiple objectives.

[0076] ;

[0077] in, This represents a multi-objective gradient consistency constraint term, used to limit conflicts between different objectives; Indicate target The gradient direction vector, Indicate target The gradient direction vector;

[0078] Context-driven state transition models embed synchronous feedback:

[0079] ;

[0080] in, and These are the disturbance coefficient and the synchronization feedback gain, respectively. and These are the gradient modulation coefficient and the synchronization deviation sensitivity coefficient, respectively. It represents the average synchronization degree within a certain time window, used to smooth synchronization feedback; The gradient term of the trajectory optimization loss function is used to guide the direction of state transition.

[0081] Final optimization goal:

[0082] ;

[0083] in, For regularization terms, , To lose weight, This represents the context-enhanced adaptive loss function;

[0084] Overall training loss integration trajectory generation and reinforcement optimization:

[0085] ;

[0086] in, , This is the overall loss adjustment coefficient. This represents the overall system loss function.

[0087] In a further technical solution, to ensure the stability of system training and execution in step 4, the following criterion is defined:

[0088] Global synchronization stability constraints ensure that the phase converges to the average phase:

[0089] ;

[0090] Convergence conditions for multi-objective optimization, ensuring gradient convergence:

[0091] ;

[0092] Trajectory dynamics smoothing conditions, limiting maximum acceleration to ensure trajectory smoothness:

[0093] .

[0094] The present invention provides a decision planning method based on scene oscillation synchronization and context reinforcement adaptation, the beneficial effects of which are as follows:

[0095] (1) By establishing a unified context awareness and collaborative behavior modeling framework, the synchronous decision-making of multiple agents in dynamic scenarios can be realized, which significantly enhances the security and stability of the system in sudden events and complex interaction scenarios.

[0096] (2) Introduce an integrated scene planning and decoding structure in the trajectory generation process, and realize the coordinated regulation of multimodal information and group dynamics through the synchronous-context guidance mechanism to ensure the global consistency of the trajectory in semantic, spatiotemporal and dynamic aspects.

[0097] (3) Adopting an adaptive optimization strategy based on context evolution, a dynamic balance is achieved among multiple objectives such as safety, efficiency and comfort, enabling the system to autonomously adjust its decision-making tendency according to traffic conditions and environmental risks.

[0098] (4) Construct a synchronous-context feedback closed-loop reinforcement optimization mechanism, combine the training of supervision signals and reinforcement signals to achieve bidirectional synergy between policy optimization and trajectory optimization, so as to maintain excellent generalization ability and robustness in unseen scenarios. Attached Figure Description

[0099] Figure 1 A flowchart illustrating a decision planning method based on scene oscillation synchronization and context reinforcement adaptation, provided as an embodiment of the present invention;

[0100] Figure 2 This is a diagram of the SAPNet framework. Detailed Implementation

[0101] To make the objectives, technical solutions, and advantages of this invention clearer, the 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 and not intended to limit the invention.

[0102] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0103] like Figure 1 and Figure 2 As shown, an embodiment of the present invention provides a decision-planning method (SAPNet) based on scene oscillation synchronization and context reinforcement adaptation. By constructing scene extraction, synchronization oscillation, and context reinforcement adaptation, an efficient integrated decision-planning autonomous driving framework is established, which specifically includes the following steps:

[0104] Step 1: Multi-dimensional scene embedding and adaptive fusion coding;

[0105] By collaboratively encoding multi-source sensor inputs using a unified representation framework, a high-dimensional, consistent scene representation is generated, providing a foundation for subsequent stages. Input data includes visual features. Vehicle body condition semantic structure and environmental characteristics Formalization To achieve a unified representation across different modalities, a weighted attention mechanism is first used to allocate modal weights and generate multimodal fusion features. :

[0106] ;

[0107] in, For modal adaptive weights, This is the feature extraction function for the corresponding mode.

[0108] To enhance the spatiotemporal consistency and alignment of fused features, temporal embedding is introduced. and spatial embedding Establishing a spatiotemporal joint mapping through MLP :

[0109] ;

[0110] in, This indicates that the feature concatenation operation enables synchronous modeling in semantics and space, ensuring the temporal alignment of encoded features in dynamic environments; This represents the mapping function of a multilayer perceptron, which consists of a linear transformation layer and a nonlinear activation function, and is used to realize high-dimensional nonlinear mapping and interaction of input features.

[0111] Building upon this, by leveraging residual structures and nonlinear mappings, and through a layered approach of activation functions and prior transformations, the nonlinear coupling capability between modes is enhanced, generating preliminary high-dimensional features capable of capturing potential multimodal correlation patterns and dynamic coupling relationships. :

[0112] ;

[0113] in, It is the Sigmoid activation function. and It is a linear transformation matrix. and These are bias terms, all of which are learnable parameters; This represents the modified linear unit activation function, used to introduce nonlinear feature transformations while maintaining computational efficiency, thereby improving the model's expressive power and convergence stability.

[0114] Finally, to achieve dynamic adjustment and energy-driven fusion of modal features, a dynamic energy-gated fusion model was designed. Based on spatiotemporal embedding features, an energy mapping and gating control mechanism was introduced. The information flow of different modalities was regarded as an energy channel, and the fusion intensity was controlled by changes in energy state, thereby generating a final unified high-dimensional scene embedding feature. :

[0115] ;

[0116] in, and These are the energy transformation matrix and the gate control matrix, respectively, used to allocate fusion weights among the feature channels; and For the corresponding bias term; This is the energy transformation branch, used to capture the potential energy flow relationships between modes; This is the gated modulation factor, used to adaptively control the contribution ratio of each mode according to the complexity of the scene; This is the scene energy mapping function, used to model the impact of traffic density, risk level, and environmental disturbances on the fusion strength; This indicates the Hadamard element-wise multiplication operation, which is the multiplication between corresponding elements and is used to achieve point-to-point modulation between the gated signal and the energy map.

[0117] Step 2: Synchronize scene oscillations;

[0118] Based on the high-dimensional scene embedding features generated in step 1 An improved multi-stage Kuramoto model was designed, and a dynamic synchronization mechanism was constructed to characterize the nonlinear interaction and cooperative behavior among multiple agents. The system is assumed to contain... There are 3 agents, each of which is considered a phase oscillator with a phase of 1 / 2. , indicating the th continuous time interval The phase variables of each agent, with natural frequencies of... The basic phase evolution equation is:

[0119] ;

[0120] in, This is an adjacency relation matrix. For global coupling strength, For the first An intelligent agent at time The phase variable.

[0121] To adapt to complex traffic scenarios, a multi-order synchronization model is adopted, introducing a second-order coupling term:

[0122] ;

[0123] in, It is a second-order coupling matrix. It is a second-order synchronous gain, which enhances the ability to model complex nonlinear phase relationships.

[0124] Global synchronization stability is measured by synchronization metrics. Evaluate:

[0125] ;

[0126] in, The average phase under steady state. The closer the value is to 1, the more synchronized the system is. Indicates the first continuous time interval Phase variables of each agent.

[0127] To avoid excessive local synchronization, an adaptive coupling coefficient was designed:

[0128] ;

[0129] in, For intelligent agents The local synchronization degree is calculated by averaging the phases of the local neighborhood. To adjust the coefficients, the coupling strength in locally asynchronous regions is dynamically enhanced; This represents the initial coupling strength (baseline value), indicating the default coupling level of the system without synchronization adjustment; Represents intelligent agents The local coupling strength is used to characterize the synchronous interaction capability of the agent with its domain individuals.

[0130] Furthermore, to adapt to the time-varying characteristics of interaction relationships in complex traffic scenarios, a dynamic adjacency weight update mechanism is proposed to adjust the interaction intensity based on spatial distance and speed difference:

[0131] ;

[0132] in, Indicates the current time step; Indicates the time step, which is the sampling period over continuous time and is used to control the update frequency of the adjacency weights; This represents the update rate coefficient, used to control the fusion ratio of historical adjacency relationships and current interaction features; Indicates the current time step Time-based intelligent agents and The adjacency weights reflect the interaction between the two elements, and are derived from the adjacency matrix. Expanded to a time-related form, Defined as a multivariable Gaussian kernel function:

[0133] ;

[0134] in, and For intelligent agents and Spatial distance and speed difference between them; For the current time step Traffic density; For reference density (e.g., normal flow rate); These are road topological features used to quantify lane connectivity or geometric constraints (such as...). Indicates the same lane. (Indicates different lanes or intersections) , , , The bandwidth parameters for each variable control the sensitivity of weight updates.

[0135] Meanwhile, the current time step is obtained after statistical averaging. Synchronous feature vector , as input guides for subsequent steps, where, For the current time step global synchronization degree (i.e. (Time sequence form).

[0136] Step 3: Integrated Scene Planning and Decoding;

[0137] This stage is based on high-dimensional scene embedding features. and time context vector Generate trajectory sequences that conform to dynamic constraints and physical continuity. ,in Indicates the current time step The lateral position of the vehicle in the Cartesian coordinate system Indicates vertical position. Indicates the time length of the trajectory (number of prediction steps).

[0138] An integrated decoding structure is established within a unified feature space, and the trajectory generation process is realized through recursive relationships:

[0139] ;

[0140] in, Indicates the current time step The state vector at that time; This is a decoder function used to achieve differentiable prediction and continuous evolution of the trajectory in the time dimension; and These are the vehicle's speed and heading angle, respectively.

[0141] To ensure semantic consistency in the trajectory generation process across different modalities, a cross-modal collaborative constraint was designed to correct the differences between the visual and semantic modalities, preventing perceptual biases from accumulating and amplifying at the decoding layer. The goal of this constraint is to maintain the decoding consistency of multi-source features within the high-dimensional embedding space, thereby achieving collaborative consistency between the perception and planning processes.

[0142] ;

[0143] in, and These are feature mapping functions for visual and semantic modalities, respectively; This represents the cross-modal consistency loss, used to constrain the consistency between visual and semantic modalities in a high-dimensional feature space.

[0144] To ensure the geometric continuity and smoothness of the trajectory, curvature smoothing regularization constraints are introduced during the decoding process. By limiting the rate of change of curvature, sharp turns and non-physical oscillations are prevented from occurring in the generated path.

[0145] ;

[0146] in, For the trajectory at the current time step Curvature at time; This represents the curvature loss smoothing term, used to constrain the smoothness of trajectory geometric changes; Indicates that the vehicle is in The velocity component in the direction, Indicates that the vehicle is in The velocity component in the direction, and This represents the acceleration component in the corresponding direction.

[0147] At the dynamics level, to ensure that the trajectory output conforms to the vehicle's physical motion characteristics, a constraint model based on the vehicle's kinematic equations is established:

[0148] ;

[0149] Considering the above constraints, a trajectory optimization objective function is defined, and multimodal consistency, geometric smoothness, and dynamic feasibility are jointly optimized:

[0150] ;

[0151] in, For dynamic consistency loss, , , These are adjustable weighting coefficients used to balance various optimization objectives. This represents the overall optimization loss for trajectory generation.

[0152] Based on this, a synchronization-context guidance term was designed. By introducing scene synchronization features and context factors to jointly modulate the decoding process, trajectory generation is jointly driven by the scene group synchronization state and environmental risk factors.

[0153] ;

[0154] in, This is a synchronization-context guidance function used to map the group phase offset and context weight feedback to the trajectory update amount; For guiding strength coefficient, Represents contextual factors (including risk level and environmental density indicators).

[0155] Step 4: Contextual reinforcement adaptation;

[0156] This stage achieves a dynamic balance between safety, efficiency, and comfort through a context-driven reinforcement learning optimization mechanism. (Define state) ,action ,Strategy and reward function The optimization objective is:

[0157] ;

[0158] in, As a discount factor, This represents the expected cumulative reward function. Indicates policy-based Expectation operation, Indicates the current time step The reward discount coefficient.

[0159] Design Context-Enhanced Reward Model :

[0160] ;

[0161] in, Weight each objective accordingly, balancing safety, efficiency, comfort, and group synchronization; For synchronous feedback gain; This indicates a safety bonus, which measures the minimum distance between the vehicle and surrounding traffic elements, the safety margin, and the risk of collision. This indicates efficiency-related rewards, and assesses average speed, task completion rate, and travel time. This indicates a comfort-related bonus, used to constrain factors such as the rate of change of acceleration and the rate of change of direction, thereby improving ride smoothness.

[0162] The reward weights are updated through a dynamic adjustment mechanism:

[0163] ;

[0164] in, The learning rate ensures that the weights adjust adaptively.

[0165] Policy updates use PPO:

[0166] ;

[0167] in, This is the advantage estimate; To truncate parameters; This represents the pruning loss function, used to prevent large policy updates from causing training instability; This represents the expected evaluation operation on empirical samples.

[0168] To achieve multi-objective balance, Pareto boundary constraints are introduced to ensure consistency of gradient directions across multiple objectives.

[0169] ;

[0170] in, This represents a multi-objective gradient consistency constraint term, used to limit conflicts between different objectives. Indicate target The gradient direction vector, Indicate target The gradient direction vector.

[0171] Context-driven state transition models embed synchronous feedback:

[0172] ;

[0173] in, and These are the disturbance coefficient and the synchronization feedback gain, respectively. and These are the gradient modulation coefficient and the synchronization deviation sensitivity coefficient, respectively. It represents the average synchronization degree within a certain time window, used to smooth synchronization feedback; This represents the gradient term of the trajectory optimization loss function, used to guide the direction of state transition.

[0174] Final optimization goal:

[0175] ;

[0176] in, For regularization terms, , To lose weight, This represents the context-enhanced adaptive loss function.

[0177] Overall training loss integration trajectory generation and reinforcement optimization:

[0178] ;

[0179] in, , This is the overall loss adjustment coefficient. This represents the overall system loss function.

[0180] To ensure the stability of system training and execution, the following criteria are defined:

[0181] (a) Global synchronization stability constraints ensure that the phase converges to the average phase:

[0182] ;

[0183] (b) Convergence conditions for multi-objective optimization to ensure gradient convergence:

[0184] ;

[0185] (c) Trajectory dynamics smoothing conditions, limiting maximum acceleration to ensure trajectory smoothness:

[0186] .

[0187] The effectiveness of this method is verified through simulation examples below.

[0188] SAPNet training employs a distributed computing framework based on PyTorch's Distributed Data Parallel module, utilizing four NVIDIA RTX3090 GPUs for parallel computation to accelerate large-scale data processing and model optimization. The training batch size is set to 64, optimized for the diverse scenarios of the nuPlan dataset, ensuring efficient use of GPU resources while avoiding memory overflow. An adaptive learning rate adjustment strategy is employed, with an initial value of 1e-4. A cosine annealing (LR) scheduler is used to dynamically reduce the learning rate during training, with a minimum value of 1e-6 to balance model convergence speed and stability. The entire training process iterates for 50 epochs, lasting approximately 100 hours of computation, ensuring the model's generalization ability across diverse scenarios. Gradient clipping (maximum norm 1.0) is applied during training to prevent gradient explosion and ensure the stability of the optimization process. Furthermore, an early stopping mechanism is employed (patiently waiting 4 epochs and monitoring the combined loss on the validation set). Training stops when the validation loss fails to decrease for 10 consecutive epochs to prevent overfitting.

[0189] The evaluation framework employs the industry-standard nuPlan benchmark to comprehensively evaluate the performance of autonomous driving planning algorithms. The nuPlan benchmark provides massive amounts of real-world driving data, totaling over 1282 hours, covering diverse scenarios in cities such as Boston, Pittsburgh, Las Vegas, and Singapore, including busy city streets, highways, complex intersections, and pedestrian-intensive areas. nuPlan supports closed-loop simulation, allowing the planner to interact in real-time with reactive agents based on Intelligent Driving Models (IDMs) and non-reactive agents replaying data from the dataset, thus realistically reproducing the dynamics and complexity of traffic scenarios. This closed-loop evaluation captures the long-term cascading effects of planning decisions, ensuring the reliability and robustness of the algorithm in real-world deployment scenarios.

[0190] Table 1. Comparison of methods on the nuPlan Val14 closed-loop benchmark.

[0191]

[0192] Table 1 presents a comparative analysis of SAPNet and other state-of-the-art methods on the nuPlan Val14 benchmark. The results show that, compared to methods such as PlanTF-H and PLUTO, SAPNet more accurately captures the dynamic interactions between multiple agents through a scene oscillation synchronization mechanism and achieves global consistency in trajectory generation through a synchronization-context-guided decoding structure. Its improved minimum collision time reflects a longer reaction time in unexpected events, while the significant improvements in driving progress and average speed scores are attributed to the ability of context-enhanced adaptive optimization to balance smoothness and dynamic risk. SAPNet performs excellently on all key metrics, especially approaching the optimum in collision score (98.62) and average speed (97.87), demonstrating its superior capabilities in safety and regulatory compliance. Compared to GameFormer (collision score 93.57, average speed 93.88), SAPNet significantly reduces collision risk and improves speed control accuracy through multimodal fusion and group collaboration mechanisms. SAPNet's minimum collision time score of 94.77 significantly surpasses PlanTF-H (92.37) and PLUTO (93.09), reflecting its longer safety reaction time in dynamic interactive scenarios. This is attributed to the accurate modeling of multi-agent behavior by its improved Kuramoto model. The driving progress index (94.51) shows that SAPNet outperforms other methods in navigation efficiency, thanks to cross-modal constraints that optimize trajectory smoothness and task completion. Furthermore, SAPNet effectively reduces vehicle acceleration fluctuations in the passenger comfort dimension through trajectory smoothing constraints and synchronous feedback mechanisms, resulting in an overall score improvement of approximately 7%-12% compared to baseline methods. To further demonstrate SAPNet's performance, Table 2 shows the performance changes of SAPNet under different module removal conditions.

[0193] Table 2 Ablation Experiment

[0194]

[0195] The results show that each core module significantly contributes to the overall system performance: removing Contextual Reinforcement Adaptation (CRA) weakens the model's self-adjustment ability in dynamic scenarios and reduces the policy's flexibility in responding to environmental changes; removing Contextual Oscillation Synchronization (SOS) disrupts the cooperative stability among multiple agents, causing group behavior predictions to deviate from a consistent rhythm; removing Dynamic Energy Gated Fusion (Gate) leads to a decrease in the alignment accuracy between multimodal features, thus affecting the unified representation of perceptual information; and disabling the Sync-Context guidance term weakens the spatiotemporal consistency and global coordination in the trajectory generation process. Overall, these four modules work synergistically and complement each other, jointly ensuring the system's comprehensive advantages in safety, smoothness, and global consistency, providing a solid guarantee for stable decision-making and efficient planning in complex traffic scenarios.

[0196] Experimental results on the nuPlan Val14 benchmark comprehensively validate SAPNet's significant advantages in integrated decision-making and planning, as well as trajectory generation quality. The improved consistency across metrics and multiple scenarios validates the core innovations of this method: A dynamic energy-gated fusion model is proposed, achieving adaptive fusion and spatiotemporal alignment of multimodal features through energy-driven and gating modulation mechanisms, significantly enhancing the consistency of multi-source information expression and dynamic response capabilities; an improved multi-order Kuramoto scenario oscillation synchronization model is designed to accurately characterize the nonlinear phase coupling and cooperative behavior among multiple agents, effectively improving the stability and global coordination of group dynamic modeling; a synchronization-context-guided decoding mechanism is constructed, introducing joint modulation of synchronization states and context factors during trajectory generation to achieve spatiotemporal consistency and global self-consistency in the planning process; and a context-reinforced adaptive optimization framework is proposed, combining context-driven reward modeling and a multi-objective policy update mechanism to achieve a dynamic balance between safety, efficiency, and comfort.

[0197] In summary, SAPNet, through a collaborative design of multimodal fusion, group synchronization, context guidance, and reinforcement optimization, constructs a closed-loop integrated decision-making system from scene understanding to trajectory planning and policy learning. This significantly improves the system's safety, smoothness, and environmental adaptability under complex traffic conditions, providing a new high-performance technology paradigm for autonomous driving decision-making and planning.

[0198] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation, characterized in that, Includes the following steps: Step 1: Multi-dimensional scene embedding and adaptive fusion coding: Co-encode multi-source sensor inputs to generate high-dimensional scene embedding features; Step 2: Scene oscillation synchronization. Based on the high-dimensional scene embedding features, a dynamic synchronization mechanism among multiple agents is constructed to generate a synchronization feature vector. Step 3: Integrated scene planning and decoding: Based on the high-dimensional scene embedding features and synchronization feature vectors, generate a trajectory sequence that meets the constraints of dynamics and smoothness. Step 4: Context-driven reinforcement adaptation. Based on the context-driven reinforcement learning mechanism, the trajectory sequence is optimized for multiple objectives to achieve a dynamic balance between safety, efficiency and comfort.

2. The decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation according to claim 1, characterized in that, The specific steps of step 1 are as follows: A unified representation framework is used to collaboratively encode multi-source sensor inputs, generating a high-dimensional, consistent scene representation; the input data includes visual features. Vehicle body condition semantic structure and environmental characteristics Formalization First, a weighted attention mechanism is used to allocate modal weights to generate multimodal fusion features. : ； in, For modal adaptive weights, This is the feature extraction function for the corresponding mode; Introducing time embedding and spatial embedding Establishing a spatiotemporal joint mapping through MLP : ； in, This represents a feature concatenation operation, used to achieve simultaneous semantic and spatial modeling; This represents the mapping function of a multilayer perceptron, which consists of a linear transformation layer and a nonlinear activation function, and is used to realize high-dimensional nonlinear mapping and interaction of input features; Building upon this, by utilizing residual structures and nonlinear mappings, and through a layered approach of activation functions and prior transformations, preliminary high-dimensional features capable of capturing potential multimodal correlation patterns and dynamic coupling relationships are generated. : ； in, It is the Sigmoid activation function. and It is a linear transformation matrix. and These are bias terms, all of which are learnable parameters; This indicates a modified linear unit activation function; Finally, based on the spatiotemporal embedding features, an energy mapping and gating control mechanism is introduced. The information flow of different modalities is regarded as an energy channel, and the fusion intensity is controlled by the change of energy state to generate the final unified high-dimensional scene embedding features. : ； in, and These are the energy transformation matrix and the gate control matrix, respectively, used to allocate fusion weights among the feature channels; and For the corresponding bias term; This is the energy transformation branch, used to capture the potential energy flow relationships between modes; This is the gated modulation factor, used to adaptively control the contribution ratio of each mode according to the complexity of the scene; This is the scene energy mapping function, used to model the impact of traffic density, risk level, and environmental disturbances on the fusion strength; This indicates the Hadamard element-wise multiplication operation, which is the multiplication between corresponding elements and is used to achieve point-to-point modulation between the gated signal and the energy map.

3. The decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation according to claim 2, characterized in that, Step 2 includes the following specific steps: Based on the high-dimensional scene embedding features generated in step 1 An improved multi-stage Kuramoto model is designed, and a dynamic synchronization mechanism is constructed to characterize the nonlinear interaction and cooperative behavior among multiple agents; the system is assumed to include There are 3 agents, each of which is considered a phase oscillator with a phase of 1 / 2. , indicating the th continuous time interval The phase variables of each agent, with natural frequencies of... The basic phase evolution equation is: ； in, This is an adjacency relation matrix. For global coupling strength, For the first An intelligent agent at time The phase variable; To adapt to complex traffic scenarios, a multi-order synchronization model is adopted, introducing a second-order coupling term: ； in, It is a second-order coupling matrix. It is a second-order synchronous gain, used to enhance the ability to model complex nonlinear phase relationships; Global synchronization stability is measured by synchronization metrics. Evaluate: ； in, The average phase under steady state. The closer the value is to 1, the more synchronized the system is. Indicates the first continuous time interval Phase variables of each agent; To avoid excessive local synchronization, an adaptive coupling coefficient was designed: ； in, For intelligent agents The local synchronization degree is calculated by averaging the phases of the local neighborhood. To adjust the coefficients, the coupling strength in locally asynchronous regions is dynamically enhanced; The initial coupling strength represents the default coupling level of the system without synchronization adjustment. Represents intelligent agents The local coupling strength is used to characterize the agent's ability to synchronously interact with individuals in its domain; Furthermore, a dynamic adjacency weight update mechanism is proposed to adjust the interaction intensity based on spatial distance and velocity difference: ； in, Indicates the current time step; Indicates the time step, which is the sampling period over continuous time and is used to control the update frequency of the adjacency weights; This represents the update rate coefficient, used to control the fusion ratio of historical adjacency relationships and current interaction features; Indicates the current time step Time-based intelligent agents and The adjacency weights reflect the interaction between the two elements, and are derived from the adjacency matrix. Expanded to a time-related form, Defined as a multivariable Gaussian kernel function: ； in, and For intelligent agents and Spatial distance and speed difference between them; For the current time step Traffic density; For reference density; These are road topological features used to quantify lane connectivity or geometric constraints. , , , The bandwidth parameters for each variable control the sensitivity of weight updates; Meanwhile, the current time step is obtained after statistical averaging. Synchronous feature vector , as input guides for subsequent steps, where, For the current time step Global synchronization degree.

4. The decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation according to claim 3, characterized in that, Step 3 includes the following specific steps: Based on high-dimensional scene embedding features and time context vector Generate trajectory sequences that conform to dynamic constraints and physical continuity. ,in Indicates the current time step The lateral position of the vehicle in the Cartesian coordinate system Indicates vertical position. Indicates the time length of the trajectory; An integrated decoding structure is established within a unified feature space, and the trajectory generation process is realized through recursive relationships: ； in, Indicates the current time step The state vector at that time; This is a decoder function used to achieve differentiable prediction and continuous evolution of the trajectory in the time dimension; and These are the vehicle's speed and heading angle, respectively. Design cross-modal collaborative constraints to correct for discrepancies between the visual and semantic modalities: ； in, and These are feature mapping functions for visual and semantic modalities, respectively; This represents cross-modal consistency loss, used to constrain the consistency between visual and semantic modalities in a high-dimensional feature space; During the decoding process, curvature smoothing regularization constraints are introduced to limit the rate of curvature change and prevent abrupt changes and non-physical oscillations in the generated path. ； in, For the trajectory at the current time step Curvature at time; This represents the curvature loss smoothing term, used to constrain the smoothness of trajectory geometric changes; Indicates that the vehicle is in The velocity component in the direction, Indicates that the vehicle is in The velocity component in the direction, and Indicates the acceleration component in the corresponding direction; At the dynamics level, a constraint model based on the vehicle's kinematic equations is established: ； Considering the above constraints, a trajectory optimization objective function is defined, and multimodal consistency, geometric smoothness, and dynamic feasibility are jointly optimized: ； in, For dynamic consistency loss, , , These are adjustable weighting coefficients used to balance various optimization objectives. This represents the overall optimization loss for trajectory generation; Based on this, a synchronization-context guidance term was designed. By introducing scene synchronization features and context factors to jointly modulate the decoding process, trajectory generation is jointly driven by the scene group synchronization state and environmental risk factors. ； in, This is a synchronization-context guidance function used to map the group phase offset and context weight feedback to the trajectory update amount; For guiding strength coefficient, Represents contextual factors.

5. The decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation according to claim 4, characterized in that, Step 4 includes the following specific steps: A dynamic balance between safety, efficiency, and comfort is achieved through a context-driven reinforcement learning optimization mechanism. Define state ,action ,Strategy and reward function The optimization objective is: ； in, As a discount factor, This represents the expected cumulative reward function. Indicates policy-based Expectation operation, Indicates the current time step The reward discount factor; Design Context-Enhanced Reward Model : ； in, Weight each objective accordingly, balancing safety, efficiency, comfort, and group synchronization; For synchronous feedback gain; This indicates a safety bonus, which measures the minimum distance between the vehicle and surrounding traffic elements, the safety margin, and the risk of collision. This indicates efficiency-related rewards, and assesses average speed, task completion rate, and travel time. This indicates a comfort-related bonus designed to improve ride smoothness. The reward weights are updated through a dynamic adjustment mechanism: ； in, The learning rate ensures that the weights adjust adaptively. Policy updates use PPO: ； in, This is the advantage estimate; To truncate parameters; This represents the pruning loss function, used to prevent large policy updates from causing training instability; This represents the expectation evaluation operation on empirical samples; To achieve multi-objective balance, Pareto boundary constraints are introduced to ensure consistency of gradient directions across multiple objectives. ； in, This represents a multi-objective gradient consistency constraint term, used to limit conflicts between different objectives; Indicate target The gradient direction vector, Indicate target The gradient direction vector; Context-driven state transition models embed synchronous feedback: ； in, and These are the disturbance coefficient and the synchronization feedback gain, respectively. and These are the gradient modulation coefficient and the synchronization deviation sensitivity coefficient, respectively. It represents the average synchronization degree within a certain time window, used to smooth synchronization feedback; The gradient term of the trajectory optimization loss function is used to guide the direction of state transition. Final optimization goal: ； in, For regularization terms, , To lose weight, This represents the context-enhanced adaptive loss function; Overall training loss integration trajectory generation and reinforcement optimization: ； in, , This is the overall loss adjustment coefficient. This represents the overall system loss function.

6. The decision-making and planning method based on scene oscillation synchronization and context reinforcement adaptation according to claim 5, characterized in that, In step 4, to ensure the stability of system training and execution, the following criteria are defined: Global synchronization stability constraints ensure that the phase converges to the average phase: ； Convergence conditions for multi-objective optimization, ensuring gradient convergence: ； Trajectory dynamics smoothing conditions, limiting maximum acceleration to ensure trajectory smoothness: 。

Images