General graph structure path planning method and system based on cortex-hippocampus biomimetic mechanism

Abstract

The application relates to a general graph structure path planning method and system based on a cortex-hippocampus biomimetic mechanism, which discretizes a continuous traffic network into a standard directed graph structure, simulates the use of a graph attention network by a posterior parietal cortex to extract global topological semantics, simulates a deep integration of multiple types of cells in an entorhinal cortex to generate unique multi-modal path node fingerprints for candidate nodes, simulates a three-stage pipeline of a hippocampus for hierarchical value evaluation, including mode separation through a dentate gyrus, value scoring through a CA3 recurrent attractor network and global calibration through a CA1, introduces a self-supervised prediction coding mechanism to replace traditional Hebb learning, realizes end-to-end derivable optimization of a memory mechanism, and finally outputs path decision by combining a simulated prefrontal cortex with mask and temperature sampling. The application realizes general planning capability for any graph topology, is highly compatible with a reinforcement learning gradient, significantly improves decision accuracy and training stability under a complex topology, and has good interpretability and scalability.

Description

Technical Field

[0001] This invention relates to the field of path planning technology, and in particular to a general graph structure path planning method and system based on a cortical-hippocampal biomimetic mechanism. Background Technology

[0002] In complex and dynamic real-time traffic scenarios, achieving robust global path planning with strong generalization capabilities is a key challenge in the field of autonomous driving. Current path planning research mainly falls into three technical paths: traditional heuristic search, deep learning-based end-to-end planning, and neuromorphic navigation models.

[0003] Traditional heuristic search algorithms, such as A* and Dijkstra's algorithm, exhibit high determinism and interpretability on discrete graph structures. However, these methods rely on predefined heuristic functions, face computational efficiency bottlenecks when dealing with large-scale complex road networks, and require frequent replanning when faced with dynamic obstacles or topological changes, resulting in insufficient adaptive capabilities.

[0004] In recent years, planning models based on graph neural networks (GNNs) have demonstrated strong feature extraction capabilities. For example, VectorNet uses graph convolutional networks to process vectorized high-precision maps, and PGP uses discrete strategies on directed graphs to traverse lane maps. However, these methods suffer from insufficient generalization performance when faced with unseen topological structures, lack the ability to model temporal decision consistency, and do not possess a general and scalable model framework to simulate the navigation function of biomimetic structures.

[0005] Although existing studies have introduced the concepts of position cells or grid cells into path planning, current schemes typically only simulate a single cell type, failing to achieve the synergistic integration of multiple spatial cells in the entorhinal cortex, and also failing to couple the predictive encoding memory mechanism of the hippocampus with deep reinforcement learning gradient optimization.

[0006] Based on the above analysis, the existing technology mainly has the following drawbacks: (1) Existing deep learning path planning models are trained for specific road networks. When faced with unknown graph topology, the planning performance drops sharply. They lack the planning ability for general graph structure input and have poor generalization.

[0007] (2) Traditional Hebbian Learning rules achieve memory by manually modifying synaptic weights. They are nondifferentiable in backpropagation of deep networks, which can easily interfere with gradient stability and lead to training failure.

[0008] (3) Existing brain-like models only simulate the function of a single spatial cell and cannot provide a multi-dimensional fusion spatial representation, resulting in insufficient environmental perception.

[0009] (4) Lacking a hierarchical bionic architecture from perception to memory to decision-making, the decision-making process cannot distinguish different cognitive stages such as spatial encoding, pattern separation, value assessment and final selection.

[0010] (5) Existing models are usually strongly coupled with specific simulation platforms or data formats, and cannot accept arbitrary graph structure data as input to complete path planning. At the same time, they cannot quickly integrate new feature extraction modules.

[0011] In summary, existing technologies face multiple bottlenecks in the field of autonomous driving path planning, including weak generalization ability, incompatibility between memory mechanisms and gradient optimization, single spatial encoding dimension, and lack of cognitive hierarchy in model architecture. Summary of the Invention

[0012] The purpose of this invention is to overcome the defects of the prior art by providing a general graph structure path planning method and system based on the cortical-hippocampal biomimetic mechanism. This invention has solved or partially solved the shortcomings of existing methods, such as poor generalization, non-differentiable memory mechanism, single spatial encoding, lack of hierarchical decision architecture, and limited model universality.

[0013] The objective of this invention can be achieved through the following technical solutions: One aspect of the present invention provides a general graph structure path planning method based on a cortical-hippocampal biomimetic mechanism, comprising the following steps: The traffic network is discretized into a directed graph structure, where the nodes and edges in the directed graph structure correspond to multi-dimensional attribute features. The posterior parietal cortex is simulated, and a graph attention network is used to extract features from the directed graph structure to generate contextual features of the vehicle nodes. Simulating the entorhinal cortex, a unique multimodal path node fingerprint is generated for each candidate successor node of the current node by deeply integrating multiple biomimetic spatial cell encoders. The hippocampus is simulated, and based on the contextual features of the vehicle nodes, a three-level pipeline hierarchical value evaluation is performed on the fingerprints of the path nodes, outputting the calibration value score of each candidate node. The three-level pipeline includes: simulating the dentate gyrus for pattern separation, simulating the CA3 cyclic attractor network for value scoring, and simulating CA1 for global calibration. The CA3 cyclic attractor network utilizes a self-supervised predictive coding mechanism to achieve end-to-end coupling optimization of memory consolidation and deep reinforcement learning gradients by minimizing the temporal transition prediction error of the CA3 hidden state. The prefrontal cortex is simulated, and the final path decision is performed based on the calibration value score, combined with the adjacency mask constraint and temperature scaling sampling strategy, to output the optimal action.

[0014] As a preferred technical solution, the process of generating the context features of the vehicle node is implemented using the following formula: ,

[0015] ,

[0016] ,

[0017] ,

[0018] ,

[0019] ,

[0020] in, For global topological semantic features; It is the set of nodes in a directed graph structure; For the first Node characteristics of each node; For activation functions; For nodes The set of neighboring nodes; For nodes , Attention coefficient between them; For the first The node feature projection matrix of the head; , For nodes , Multidimensional attribute features; For the first Under each attention head, the node For neighboring nodes Normalized attention weights; Indicates transpose; Use the LeakyReLU activation function; This represents the learnable feature extraction matrix of the self-attention mechanism; For directed edges Multidimensional attribute features; Indicates splicing; The contextual features of the vehicle node in the global graph include local and global spatial information observed by the vehicle. Feature information of the vehicle node extracted from GATv2; , This represents two different multilayer perceptron networks; This indicates that corresponding elements are stacked. Global edge features; It is the set of edges in a directed graph structure; For classification functions; This is a learnable weight matrix used to extract global edge feature information; For directed edges Features after reconstruction.

[0021] As a preferred technical solution, the plurality of biomimetic space cell encoders include: The position cell encoder is used to simulate the spatial encoding function of position cells, performing multi-scale periodic position encoding on the normalized coordinates of nodes. : ,

[0022] Among them, the Frequency components , The node coordinates are the location. Number of frequency components, frequency component index ; A grid cell encoder is used to calculate the orientation angle of a candidate node relative to the current node. With normalized distance The characteristics of simulated hexagonal grid discharge are generated by multi-scale periodic functions. : ,

[0023] in, For the first The period length of each scale, For phase shift; Boundary cell encoder is used to statistically analyze the out-degree and obstacle distribution characteristics of passable sectors in each direction around a candidate node; The temporal cell encoder is used to fuse three time dimensions of periodic encoding: real timestamp, training epoch, and local steps. The target vector cell encoder is used to encode the shortest path distance between candidate nodes and target nodes, the distance improvement rate, and the heading alignment. By concatenating the outputs of all encoders with the node features extracted by the graph attention network, and then performing fusion processing through a multilayer perceptron, a unique multimodal path node fingerprint is generated.

[0024] As a preferred technical solution, the simulated dentate gyrus pattern separation is achieved by the following formula: ,

[0025] in, Separate the output for the current node's pattern. For path node fingerprints, , , , These are linear dimensionality increase, nonlinear sparse activation, dropout operation, and linear dimensionality reduction operation, respectively.

[0026] As a preferred technical solution, the process of simulating the CA3 cyclic attractor network for value scoring includes the following steps: Constructing a gated loop unit It is used to extract step memory states (i.e., CA3 feature information). : ,

[0027] in, Candidate nodes The stride memory state, Separate the output for the current node's pattern. For time step The stride memory state, The vector is zero at the initial time step; Based on trainable synaptic plasticity weight matrix The stepped memory state is modified to obtain an associative memory state incorporating prior knowledge: ,

[0028] in, It is the hyperbolic tangent function. The mixing coefficient; Inputting prior associative memory states into the scoring network Candidate nodes are obtained. Original value score .

[0029] As a preferred technical solution, the global calibration of the simulated CA1 is achieved using the following formula: ,

[0030] in, Candidate nodes The calibration value score, It is a multilayer perceptron network. Candidate nodes The stride memory state, This refers to the contextual features of the vehicle node in the global graph.

[0031] As a preferred technical solution, the self-supervised predictive coding mechanism includes the following steps: By predicting the encoder according to CA3 hidden state prediction time step CA3 activation mode ; Calculate the predicted activation mode With actual encoding Received The cosine distance between the CA3 hidden states at each time step is used as the self-supervised predictive coding loss. ; The self-supervised prediction encoding loss is used Backpropagation optimization is performed as part of the total loss function to achieve end-to-end differentiable memory consolidation.

[0032] As a preferred technical solution, for an agent composed of the posterior parietal cortex, entorhinal cortex, hippocampus, and prefrontal cortex, the training process is optimized using a total loss function. for: ,

[0033] ,

[0034] ,

[0035] ,

[0036] ,

[0037] ,

[0038] ,

[0039] ,

[0040] ,

[0041] in, , , , These are the policy gradient loss, state value regression loss, self-supervised prediction encoding loss, and policy entropy regularization term, respectively. and The loss coefficient, It is the entropy regularization coefficient. This is the set of candidate actions for the current node. Here is the numerical stability constant. For time step At the node In the case of the first One action The conditional probability, Indicates the training samples at time step t The expected value of the following mathematical expression For a complete time step of a path planning episode from start to finish, This indicates finding the minimum value. Indicates the value loss cutoff threshold. For cross-entropy loss, For time steps status The selected node is output by CA1. The value estimate, For time steps The return target, The conditional probability output by the policy network represents the probability in state. Select action The probability, For time steps t The actions performed To calculate the advantage function using an exponentially weighted average, This is the normalized advantage estimate. This represents the mean of the dominance values ​​within the current training batch. This represents the standard deviation of the dominance values ​​within the current batch. Indicates the numerically stable term, by default. This is used to avoid the denominator being zero. As a discount factor, For time steps TD error, yes A multi-dimensional reward system is employed, integrating reward signals from various dimensions such as goal-driven, path optimization, safety constraints, behavioral constraints, and dynamic environment adaptation. This system provides comprehensive evaluation and feedback for agent decision-making, thereby improving the safety, efficiency, and environmental adaptability of path planning in dynamic environments. This approach is used in reinforcement learning training. This allows for increasing the probability of actions with higher advantage values ​​and decreasing the probability of actions with lower advantage values, thereby continuously optimizing the policy network. To sample all time steps in the batch Take the average of experience.

[0042] As a preferred technical solution, for nodes in the directed graph structure, the multidimensional attribute features include coordinate position, node type, speed, speed limit, yaw angle, lane change sign, obstacle sign, target sign, normalized target distance, and heading alignment; for edges in the directed graph structure, the multidimensional attribute features include edge length, edge type, and edge weight.

[0043] Another aspect of the present invention provides a general graph structure path planning system based on a cortical-hippocampal biomimetic mechanism, used to implement the aforementioned general graph structure path planning method, the system comprising: The road network graph structure discretization module is used to convert a continuous traffic road network into a standard directed graph structure; The apical cortex graph topology feature extraction module is used to extract global topological semantic features and vehicle node context features based on the GATv2 graph attention network; The entorhinal cortex multimodal spatial coding module integrates position cell, grid cell, boundary cell, temporal cell and target vector cell encoders to generate unique path node fingerprints for candidate nodes; The hippocampal memory and value assessment module is used to output a calibrated value score through a three-stage pipeline of simulated dentate gyrus pattern separation, CA3 cyclic attractor network value scoring, and CA1 value calibration. The predictive coding memory consolidation module is used to optimize the temporal transition of the CA3 hidden state through a self-supervised predictive coding mechanism, so as to achieve end-to-end differentiable memory learning. The prefrontal decision output module is used to combine adjacency masking and temperature scaling strategies to make the final path decision.

[0044] Compared with the prior art, the present invention has at least one of the following beneficial effects: (1) Strong cross-scenario generalization ability: Existing path planning methods are mostly designed for specific scenarios. When faced with new environments with changes in topology, they need to be remodeled or adjusted, resulting in weak generalization ability and inability to achieve universal planning across scenarios. However, this invention discretizes the continuous traffic network into a standard directed graph structure, whose nodes and edges contain multi-dimensional attribute features. It also extracts global topological semantic features through the GATv2 graph attention network that simulates the posterior parietal cortex, enabling the model to directly process graph data and achieve universal access to arbitrary graph topology data. The model can be applied to different road network scenarios without changing its architecture, thus completing the leap from specific scenario planning to universal graph structure path planning.

[0045] (2) Achieving the combination of memory mechanism and deep reinforcement learning gradient optimization: Traditional biomimetic models rely on manual and non-differentiable weight update rules such as Heb learning for memory consolidation, which cuts off the end-to-end gradient propagation chain and cannot be effectively coupled with modern deep reinforcement learning frameworks, resulting in unstable training. To address this issue, this invention introduces a self-supervised predictive coding mechanism on the basis of the CA3 recurrent attractor network that simulates the hippocampus. The predictive encoder learns the temporal transition rules of the CA3 hidden state and takes minimizing the cosine distance loss between the predicted state and the real state as the optimization objective. A differentiable loss function replaces the traditional Heb learning rule, opens up the gradient propagation path of the memory module, and achieves end-to-end coupling optimization of memory consolidation and deep reinforcement learning gradient, which significantly improves the stability and convergence efficiency of the training process.

[0046] (3) High node recognition: Existing methods usually limit the encoding of spatial location to simple coordinate or distance information. In complex topologies such as overpasses and multi-level road networks, it is difficult to accurately distinguish nodes with similar structures, resulting in low decision accuracy. This invention simulates the entorhinal cortex and deeply integrates five bionic encoders: location cells, grid cells, boundary cells, time cells, and target vector cells. It deeply fuses multi-dimensional spatial and temporal information to generate a unique path node fingerprint for each candidate node, enriching the spatial semantic expression dimension of the node and giving each node a unique representation in space, time, and task objectives. This enhances the recognition of similar nodes and improves the decision accuracy and robustness of the model in complex topological structures.

[0047] (4) Strong interpretability: Traditional end-to-end black box models mix perception, memory and decision-making. The architecture design lacks hierarchical abstraction corresponding to biological cognitive processes, resulting in poor interpretability, difficulty in tracing errors, and difficulty in flexible expansion. This invention constructs a hierarchical biomimetic navigation architecture that simulates the parietal cortex, entorhinal cortex, hippocampus and prefrontal cortex. It constructs a cognitive link from graph topology perception, multimodal spatial encoding, three-level pipeline memory and value assessment to final decision output. This hierarchical architecture clearly simulates the complete cognitive process of the biological brain from perception, encoding, memory, assessment to decision-making. It not only improves the interpretability of the model, but also provides structured support for the independent optimization and functional expansion of each module, and provides support for building a more general intelligent navigation system.

[0048] (5) Improve training efficiency: Existing reinforcement learning models are trained from scratch in complex dynamic environments, which results in slow convergence and lack of effective course guidance, making it difficult to adapt to dynamic changes quickly and resulting in low training efficiency. This invention adopts a two-stage training strategy of behavior cloning pre-training and reinforcement learning fine-tuning. The difficulty of tasks is gradually increased in combination with course learning. Obstacles that satisfy global connectivity and local reachability constraints are dynamically injected during training. A composite loss function including policy loss, value loss, prediction encoding loss and entropy regularization term is designed. Pre-training provides the model with a good initial strategy, which accelerates the convergence of the reinforcement learning stage. The dynamic obstacle injection and composite loss function design enhance the model's adaptability and exploration ability to dynamic disturbances, and improve the robustness of the final strategy and training efficiency. Attached Figure Description

[0049] Figure 1 This is a flowchart of a general graph structure path planning method based on a cortico-hippocampal biomimetic mechanism in the embodiment. Figure 2 This is a diagram illustrating the overall architecture of the path planning agent in the embodiment. Figure 3 This is a schematic diagram of the overall process of reinforcement learning path planning based on the entorhinal cortex-hippocampus circuit in the embodiment. Figure 4 This is a flowchart of the model training process in the embodiment; Figure 5 The flowchart for the path planning method based on brain-inspired reinforcement learning in this embodiment is shown. Figure 6 This is a schematic diagram of a general graph structure path planning system based on a cortical-hippocampal biomimetic mechanism in the embodiment. Figure 7 This is a path length graph showing the output of the algorithm after 100 comparison runs in the example. Figure 8 The figure shows the comparison results of the benchmark algorithm in the example. Figure 9 This is a diagram showing the results of the dynamic obstacle avoidance experiment in the embodiment; Figure 10 This is a simulation trajectory diagram of the complete steps of the model from the starting point to the ending point in the road network in the example. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0051] Example 1 To address the problems existing in the prior art, this embodiment provides a general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism, the main contents of which include: (1) Provide a general path planning framework that can accept arbitrary graph structure data as input to realize cross-scene migration.

[0052] (2) A multimodal coding method is proposed to integrate multiple spatial cells into a unique path node fingerprint.

[0053] (3) Design a differentiable self-supervised predictive encoding memory mechanism to replace the non-differentiable manual weight update in traditional Heb learning, and realize the high coupling between the memory mechanism and the gradient optimization of deep reinforcement learning.

[0054] (4) Construct a hierarchical biomimetic navigation architecture that simulates the parietal cortex-entorhinal cortex-hippocampus-prefrontal cortex to realize a complete cognitive link from perception encoding to memory integration to value assessment and final decision-making.

[0055] See Figure 2This diagram illustrates the overall architecture of a path planning agent, depicting its structure based on brain-inspired neural mechanisms. The agent takes road network data as input and simulates the information processing mechanisms between the parietal cortex (PPC), entorhinal cortex (EC), hippocampus (HPC), and prefrontal cortex (PFC) to achieve path perception, spatial memory prediction, and action decision-making in complex traffic environments, thereby completing the dynamic path planning task for intelligent vehicles. Figure 2 The explanations of each element are as follows: 1. Road Network Map: A road network map describes the road topology in a real-world traffic scenario, including road nodes, road connections, lane information, and traffic constraints. As the initial input to the entire path planning system, the road network map provides the fundamental spatial data for subsequent graph structure modeling.

[0056] 2. Graph Structure Data: The road network map is further converted into a graph structure data format that can be processed by neural networks. This graph structure data includes node sets, edge connections, and node attribute features, and uses a graph data structure to uniformly represent the road network, providing input for topology learning of graph neural networks.

[0057] 3. Environment: The environment module is used to construct the interaction system between the agent and the traffic scenario, responsible for simulating vehicle operating status, road changes, and dynamic traffic behavior. The environment module receives action decisions from the agent and feeds back new state and reward information, realizing a closed-loop interaction process in reinforcement learning.

[0058] 4. Intelligent Agent: The intelligent agent is used to make path planning decisions based on the current environmental state. It integrates a brain-inspired spatial cognitive structure, including functional modules such as topological perception, path memory, and action decision-making. The intelligent agent continuously optimizes its path planning strategy by constantly interacting with the environment.

[0059] 4. Starting and ending point path trajectory: The environment generates the corresponding vehicle trajectory based on the path planning actions output by the agent, forming a dynamic path result from the starting point to the ending point. This trajectory is used to characterize the path planning effect of the agent at the current stage and also serves as an important basis for subsequent reinforcement learning reward evaluation.

[0060] for Figure 2 The path planning agent in the diagram has the following internal structure: 1. Parietal PPC Topology Perception and Graph Representation Layer: Used to simulate the spatial topology perception ability of the parietal cortex of the biological brain. A graph neural network is used to learn the topological relationships of the input graph structure data, extracting spatial connection features between road nodes and neighborhood association information, thereby generating high-dimensional graph representation features. It is mainly used to complete the following functions: (1) modeling the spatial relationships between road nodes; (2) extracting graph structure topology features; (3) perceiving candidate path nodes; (4) global scene spatial encoding; and (5) encoding vehicle node information. Its output is simultaneously transmitted to the entorhinal cortex EC module and the prefrontal cortex PFC decision layer.

[0061] 2. Entorhinal Cortex EC Path Node Fingerprint: This module simulates the spatial location encoding function of the entorhinal cortex in the biological brain. Candidate path nodes undergo multi-dimensional feature fusion encoding to form path node fingerprint information with spatial semantic features. The node fingerprint not only includes the node's own topological features but also incorporates information such as target orientation, boundary constraints, positional state, and temporal dynamics to enhance the spatial identification ability of path nodes. The path node fingerprint output by this module is further input into the hippocampus module for spatial memory modeling.

[0062] 3. Hippocampal Predictive Memory Integration: This module simulates the spatial memory and path prediction functions of the hippocampus. The system achieves prediction of future path trends and state integration through joint modeling of historical trajectory memory and current path node states. Specifically, the hippocampal module can: (1) store historical path sequence information; (2) learn dynamic change patterns in traffic scenarios; (3) establish temporal correlations between candidate paths; and (4) predict future path benefits. The predictive memory information output by this module is transmitted to the prefrontal cortex (PFC) decision layer to assist in action selection.

[0063] 4. Prefrontal cortex (PFC) decision-making layer: This layer simulates the advanced decision-making functions of the prefrontal cortex in biological brains. It combines the spatial topological features output from the PPC layer with the predictive memory information output from the HPC layer to comprehensively evaluate the value of each candidate path node and output the optimal action decision at the current moment. It is mainly used to perform the following functions: (1) candidate action screening; (2) path value evaluation; (3) dynamic action selection; and (4) multi-objective decision balancing. Finally, it generates the next path planning action for the vehicle.

[0064] See Figure 3 This is a schematic diagram of the overall process of reinforcement learning path planning based on the entorhinal cortex-hippocampus circuit. The process is used to construct a path planning reinforcement learning model that integrates graph neural networks and brain-inspired hippocampal memory mechanisms, enabling intelligent evaluation and dynamic decision-making of candidate path nodes in complex traffic scenarios. See also... Figure 3 It includes the following processes: 1. Real-time dynamic input: Convert the real-time road scene in the SUMO traffic simulation environment into NetworkX graph structure data; take the vehicle Ego as the center, extract the surrounding discrete nodes and candidate path nodes to form the initial input graph for path planning.

[0065] 2. Top Leaf PPC Topology Awareness and Graph Representation Layer: The GATv2 graph attention network is used to extract features of the traffic topology, learn the spatial relationships between nodes, and output a high-dimensional graph structure feature representation of each node.

[0066] 3. Graph Feature Output: Outputs global node feature information after GATv2 encoding, forming a 512-dimensional feature vector of all nodes, providing input for subsequent brain-inspired path memory encoding.

[0067] 4. Entorhinal cortex-hippocampal circuit spatial information representation layer: Construct a biomimetic hippocampal spatial memory circuit, perform spatial semantic encoding and path memory modeling on candidate nodes, and realize multi-source spatial information fusion.

[0068] 5. Entorhinal cortex EC path node fingerprint: The candidate nodes are fused and encoded to map the node features, boundary information, positional relationship, time information and target state into a unified path node representation vector.

[0069] 6. Hippocampal Predictive Memory Integration: The C-shaped cell bands of the hippocampus are often divided into four regions: CA1-CA4. Morphologically, the segment starting dorsally is generally called CA1, the part extending ventrally is CA3, the transition area between CA1 and CA3 is CA2, and the segment after CA3 turns and inserts into the small C-shape of the dentate gyrus is CA4. Here, the dentate gyrus (DG), CA3, and CA1 structures are used to complete path sequence memory and state prediction; CA3, combined with a gated recurrent unit (GRU), encodes historical trajectories, while CA1 outputs the candidate node value evaluation results.

[0070] 7. Candidate Node Feature Output: Output the candidate node feature information after being encoded by the CA3 memory in the hippocampus, forming a high-dimensional state representation for each candidate node.

[0071] 8. CA1 Value Assessment: Based on the CA1 layer, each candidate node is scored to generate a corresponding value function. It is used to measure the future revenue of different path nodes.

[0072] 9. Vehicle State Feature Output: Outputs high-dimensional state features of the vehicle's Ego node, used to fuse the current vehicle motion state with candidate path information.

[0073] 10. Prefrontal cortex (PFC) decision-making layer: Combining candidate node value, vehicle state, and action mask mechanism, it completes the next action decision and outputs the optimal path node selection result.

[0074] 11. Path execution and node migration: The vehicle moves to the next target node according to the decision result, and updates the node connection relationship and vehicle position status in the current traffic scenario.

[0075] 12. Reward Feedback: Calculate reinforcement learning reward and penalty values ​​based on vehicle performance, including indicators such as target approach, safety, traffic efficiency, and collision risk.

[0076] 13. Reward-based memory update GRU: Environmental reward feedback is input into the GRU to dynamically update the historical state memory of the CA3 in the hippocampus, thereby improving the ability to predict and make decisions on subsequent paths.

[0077] See Figure 1 The flowchart of a general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism includes the following steps: Step S1: Discretization and representation of the road network diagram structure.

[0078] Discretize the continuous road network into a directed graph. Each node Contains 11-dimensional information , ,in, , , , , , , , , , , These represent the x-coordinate of position, y-coordinate of position, node type, speed, maximum speed, yaw angle, lane change sign, obstacle sign, target sign, normalized target distance, and heading alignment, respectively. Each directed edge... Includes 3D feature vectors , , , These represent the edge length, edge type, and edge weight, respectively.

[0079] Step S2: Extract graph topological features from the simulated apical cortex.

[0080] The GATv2 graph attention network is used to process the discretized graph structure, and a multi-head attention mechanism is employed to capture the global topological semantics and local interaction dependencies of the road network. For nodes... and its neighbor node set Nodes in The attention coefficient is calculated as follows: ,

[0081] Node feature update: , The initial value is Global graph features (i.e., global topological semantic features) are obtained through mean pooling: Edge representation reconstruction uses the Post-GAT mechanism: Subsequently, global edge features are obtained through gated attention weighted pooling: Finally, the features of the vehicle nodes are fused to generate the ego-conditioned context (i.e., the vehicle node context features): , The feature information of the vehicle node extracted from GATv2.

[0082] Step S3: Simulate the multimodal spatial coding of the entorhinal cortex.

[0083] For each candidate successor node of the current node, a unique path node fingerprint is generated using various biomimetic spatial cell models: 1. Location cell coding .

[0084] Assign a unique embedding vector to each node and perform sine / cosine periodic encoding on the normalized coordinates: ,in This is a frequency division term.

[0085] 2. Grid cell coding .

[0086] Calculate the orientation angle of the candidate node relative to the current node. and normalized distance Hexagonal mesh features are generated using multi-scale periodic functions: ,in For the first The period length of each scale, This is the phase shift.

[0087] 3. Boundary cell coding .

[0088] The physical constraints around the candidate node are perceived, and the out-degree and obstacle distribution characteristics of the passable sectors in each direction are statistically analyzed. Preferably, the sector can be divided into 8 sectors by azimuth angle, and the ratio of the number of out-edges to the number of obstacle nodes in each sector is statistically analyzed and normalized to [0,1].

[0089] 4. Time-based cell coding .

[0090] By integrating three time dimensions—real timestamp, training rounds, and local steps—periodic encoding provides temporal tags for episodic memory.

[0091] 5. Target Vector Cell Encoding .

[0092] Encode target guidance information such as the shortest path distance between candidate nodes and the target, distance improvement rate, and heading alignment.

[0093] 6. Fingerprint fusion of path nodes.

[0094] The above-mentioned position cell encoding, grid cell encoding, boundary cell encoding, temporal cell encoding, and target vector cell encoding are concatenated and then fused using a multilayer perceptron. The fusion information for each candidate node is as follows: .

[0095] Step S4: Simulate memory and value assessment in the hippocampus.

[0096] Specifically, it includes the following sub-steps: Step S401, dentate gyrus pattern separation.

[0097] The nonlinear transformation of 512*4 dimensionality increase-sparse activation-512 / 4 dimensionality reduction is used to differentiate similar fingerprints. High-dimensional projection simulates the unfolding of pattern separation in neuroscience by increasing the intrinsic dimension of the representation space. In the feature compression stage, layer normalization and dimensionality compression are used to restore the feature to its original dimension in order to extract key topological invariants. .

[0098] Step S402, CA3 cyclic attractor network value scoring.

[0099] The cyclic connection dynamics of CA3 are simulated using gated cyclic units to maintain the step memory state. : To incorporate associative memories stored in prior experience, a trainable weight matrix is ​​used. Introducing synaptic plasticity components ,in This is the mixing coefficient. The final output is a value score. .

[0100] Step S403, CA1 value calibration.

[0101] Integrating direct pathways in the entorhinal cortex (autonomous node context features) (This is related to) the CA3 associative memory pathway, which calibrates the original value score. .

[0102] Step S5: Self-supervised predictive encoding memory consolidation.

[0103] By learning the temporal transition pattern of the CA3 state through the predictive encoder, the prediction error of the cosine distance between adjacent time steps is minimized. ,

[0104] in This predicts the CA3 activation pattern for the next time step. This mechanism replaces traditional Heb learning, achieving end-to-end differentiable optimization.

[0105] Step S6: Action selection in the prefrontal decision layer.

[0106] An adjacency mask is applied to the value scores of candidate nodes after CA1 calibration to filter physically disconnected nodes. An action probability distribution is generated by temperature-scaled Softmax. During training, the node is sampled according to probability, and during inference, the node with the highest score is selected.

[0107] See Figure 4 The model training flowchart includes the following steps: Step A1: System initialization.

[0108] Parse the SUMO road network file and perform lane discretization to construct a directed graph structure and instance the intelligent agent model.

[0109] Step A2: Initiate reinforcement learning training iteration.

[0110] Select the starting node and the target node, insert the vehicle into the graph structure as a token node, and establish bidirectional connection edges with the surrounding lane nodes.

[0111] Step A3: Dynamically update the graph structure.

[0112] The attribute feature vectors of each node are updated in real time according to the current simulation environment state to ensure that the graph structure is consistent with the environment state.

[0113] Step A4: The single-round path planning task is initiated.

[0114] Reset the CA3 cyclic hidden state of the hippocampus, initialize the trajectory cache, and set the termination condition for the current training round.

[0115] Step A5: Dynamic environmental disturbance injection.

[0116] A random obstacle configuration strategy is adopted, which, under the premise of satisfying global connectivity constraints and local reachability constraints, randomly injects obstacle nodes in real time and updates the feature flag bits of the corresponding nodes in the graph structure synchronously.

[0117] Step A6, Model forward reasoning and decision-making.

[0118] The parietal cortex module extracts real-time topological features, the entorhinal cortex module generates path node fingerprints for each candidate successor node, the hippocampus module performs value assessment, and the prefrontal decision layer samples and outputs the next action based on the probability distribution after masking.

[0119] Step A7, gradient optimization and parameter update.

[0120] The weighted sum of the policy gradient loss, state value regression loss, and self-supervised prediction encoding loss is calculated and then used to update all trainable parameters of the model via backpropagation.

[0121] To ensure the agent possesses efficient learning capabilities and representation stability in complex graph topologies, this embodiment constructs a multi-task collaborative optimization framework that integrates reinforcement learning and self-supervised predictive coding. This framework achieves end-to-end training through the following four complementary optimization methods: (1) Using the Generalized Advantage Estimation (GAE) algorithm. In order to balance the trade-off between bias and variance in policy gradient updates, this embodiment uses the Generalized Advantage Estimation (GAE) algorithm to calculate the advantage value. To balance the trade-off between bias and variance in the policy gradient, we first define the temporal difference (TD) error. : ,

[0122] in, It is a time step Compound rewards; This is the discount factor, defaulting to 0.99; Indicates time step The value estimate of the selected node is output by CA1. For intelligent agents at any time The environmental state in which it is located, including state features such as current position, candidate nodes, obstacle information, and target direction. The state value function is used to select the state at the current moment. Based on this, GAE calculates the advantage function using an exponentially weighted average: ,

[0123] in, This represents a complete time step of a single origin-endpoint path planning episode. Subsequently, the policy loss function... By maximizing the weighted log-likelihood, the action probability is guided to evolve towards a higher dominance direction; this is the method used in reinforcement learning training. This can increase the probability of actions with higher advantage values ​​and decrease the probability of actions with lower advantage values, thereby continuously optimizing the policy network. ,

[0124] in, The conditional probability output by the policy network represents the probability in state. Select action The probability, For time step The actions performed The normalized dominance estimate: ,

[0125] in, This represents the mean of the dominance values ​​within the current training batch. This represents the standard deviation of the dominance values ​​within the current batch. Indicates the numerically stable term, by default. This is used to avoid the denominator being zero.

[0126] (2) Cut-off value loss. To suppress overfitting caused by extreme reward fluctuations in the Critic network, such as the reward peak at the target time step, this embodiment uses the cut-off constraint idea used in the Proximal Policy Optimization (PPO) algorithm as inspiration, and introduces it into the policy to enhance training stability. By using this loss cut-off mechanism, numerical stability is optimized, and the training robustness and environmental adaptability of the model in complex road network scenarios are improved: ,

[0127] ,

[0128] in, Represents cross-entropy loss, For time step Step-by-step reward target, defined as ; This represents the value loss cutoff threshold, which defaults to 10. It is used to suppress gradient explosion and improve robustness to complex road network environments.

[0129] (3) Generating policy entropy. To encourage agents to fully explore the action space and alleviate the problem of policies converging prematurely to local optima, a reinforcement learning entropy regularization term is introduced to constrain the policy distribution: ,

[0130] in, For policy entropy, For time step At the node In the case of the first The policy entropy of an action. This is the set of candidate actions (nodes) for the current node, where the set of candidate actions (nodes) represents the successor nodes of the node where the current vehicle is located in the graph structure. This embodiment uses a directed graph. is the numerical stability constant. The action space is the set of all out-degree neighbors of the current node in the adjacency matrix, and the policy network output is the discrete probability distribution on this set.

[0131] In summary, model training is performed using the following multi-task loss function. Perform end-to-end constraints: ,

[0132] in, and These are the loss coefficients, all defaulting to 1.0. It is the entropy regularization coefficient. The self-supervised prediction encoding loss value obtained in step S5 is used. This collaborative optimization framework ensures the agent's robust navigation capability in unstructured real-time scenarios through joint training of policy execution, value calibration, and memory prediction.

[0133] Step A8: Determine the termination condition for a single round.

[0134] The system determines whether the agent has reached the target node, exceeded the maximum allowed number of steps, or entered a dead end state with no valid candidate nodes. If any of the termination conditions are met, the hippocampal reverse playback mechanism is triggered to consolidate memory.

[0135] Step A9: Persist the model checkpoints.

[0136] It periodically saves the current model parameters, optimizer status, and training statistics.

[0137] Step A10: Global training termination determination.

[0138] Determine whether the number of completed training rounds has reached the preset maximum number of episodes. If not, return to step A2 to start the next training iteration.

[0139] See Figure 5 This is a flowchart illustrating a path planning and prediction process based on brain-inspired reinforcement learning. It describes a path planning method based on a collaborative mechanism involving the parietal cortex, entorhinal cortex, hippocampus, and prefrontal cortex. Through graph structure environment perception, candidate node value evaluation, and reinforcement learning decision-making, a dynamic path search process from the starting point to the target point is achieved. It includes the following steps: Step B1: Load the model weights.

[0140] Load the parameters of the trained neural network model and initialize the path planning agent.

[0141] Step B2: Input the starting node and the target node and create the graph structure data in real time.

[0142] Input the vehicle's current location and target location, and construct road map structure data in real time based on the current traffic environment.

[0143] Step B3: Extract topological features from the parietal cortex.

[0144] The topological relationships and spatial structure features between road nodes are extracted using the graph representation network corresponding to the apical cortex.

[0145] Step B4: Obtain all candidate nodes for vehicle Ego.

[0146] Based on the current node where the vehicle is located, search for all passable successor candidate nodes to form a set of possible actions for the next step.

[0147] In step B5, the entorhinal cortex generates path node fingerprints for each candidate node.

[0148] For each candidate node, spatial location and state features are encoded to generate corresponding path node fingerprint information.

[0149] Step B6: Evaluate the value scores of candidate nodes in the hippocampus.

[0150] By combining historical path memory with the current environmental status, value prediction and benefit assessment are performed for each candidate node.

[0151] Step B7: The prefrontal cortex selects the optimal next node.

[0152] Based on the value results of the candidate nodes, select the optimal next path node for the current time step.

[0153] Step B8: Move to the selected node.

[0154] Simulate and control the vehicle to move to the selected target node, and update the current environmental state.

[0155] Step B9: Determine whether the target has been reached or the maximum number of steps has been exceeded.

[0156] Determine whether the vehicle has reached the target node, or whether the current search steps have exceeded the preset maximum number of iterations.

[0157] Step B10: The termination condition is met. If the target node is reached or the maximum number of steps is exceeded, the current path planning process ends.

[0158] Step B11: Termination condition not met. If the termination condition is not met, return to the candidate node search step and continue to the next round of path planning decision-making.

[0159] To verify the effectiveness of this method, a dataset simulating an automated container port lane network was constructed using the SUMO simulation platform. The task objective is to input the starting and ending points of the AGVs and have the model plan the globally optimal path within the complex road network topology. Road network discretization and graph construction: Through lane structure and graph transformation modules, we convert the continuous lane topology into a directed graph. The discretization step size was set to 50.0 meters. To simulate realistic dynamic interaction, a vehicle encoding mechanism was introduced. Both the vehicle and dynamic obstacle graph structures were independent nodes, linked to neighboring lane nodes via bidirectional edges to ensure real-time information transmission between vehicle information and real-time dynamic lane information. High-order node features: Each node includes its x-coordinate, y-coordinate, node type, speed, maximum speed, yaw angle, lane change sign, obstacle sign, target sign, normalized target distance, and heading alignment. Hyperparameter configuration: The training device used an NVIDIA GPU, and the software framework was based on PyTorch. The hidden layer dimension was fixed at 512, and the learning rate was set to 1×10⁻⁶. -4 The total training duration is 500 episodes, supplemented by an exponential temperature decay mechanism. To balance exploration and utilization.

[0160] To ensure the fairness and interpretability of the experimental comparisons, the network architecture parameters of all benchmark algorithms were systematically configured, ensuring that the number of parameters in the neural network modules of each comparison model was on the same order of magnitude as the proposed model. Furthermore, by strictly controlling the model capacity, independent evaluation of the advantages of the proposed biomimetic design was achieved, effectively eliminating the impact of parameter differences on performance gains.

[0161] In the benchmark algorithm comparison experiment, we used the classic graph algorithms A* and Dijkstra's algorithm on static graphs. The main comparisons were made by comparing the accuracy and generated path length. To study stability, 100 comparison experiments were conducted. The experimental results are as follows: Figure 7 The comparative experimental results show that the model of this invention achieves 100% accuracy with traditional graph algorithms A* and Dijkstra's algorithm. However, our algorithm achieves 94% accuracy with traditional graph algorithms in terms of the number of generated path nodes, i.e., the optimal path in the static graph. This demonstrates that our algorithm also has good performance in the optimal path.

[0162] Furthermore, algorithms using neural networks were compared: Deep Q-Network (DQN), Double DQN (DDQN), Proximal Policy Optimization (PPO), Prioritized Experience Replay (PER), and Hindsight Experience Replay (HER). This experiment performed 60 path planning iterations for the same start and end point, comparing the stability of the model outputs. Table 1 presents the experimental results, mainly including the maximum number of steps, minimum number of steps, average number of steps, variance, and number of successes across the 60 trials. Figure 8 It is a visualization of the benchmark algorithm comparison experiment.

[0163] Table 1. Comparison of benchmark algorithms and experimental results.

[0164] from Figure 8 As can be seen from the comparison with the path lengths in Table 1, although the accuracy rate is 100% in the path planning task with static graph information, the model of this invention has good stability and basically maintains consistent output.

[0165] Simultaneously, a set of dynamic obstacle avoidance experiments were conducted. In this invention, the model generates decisions based on real-time graph information every second during the planning process. Therefore, obstacle information was randomly added two seconds before the decision in a complete path planning operation, and six sets of experiments were conducted. The experimental results are as follows: Figure 9 As shown, the six sets of experiments correspond to (a)-(f). The orange dashed line in the figure represents the optimal path of the model of the present invention when no obstacles are added, and the solid line represents the final path of the model when obstacles are added in real time. It can be seen that the model successfully avoids the obstacle nodes and successfully reaches the destination.

[0166] The prototype was broken down into the following variations to verify the functionality of each module: 1. Full model: A complete model containing all components.

[0167] 2. Remove the graph attention network model and directly use the original multidimensional features as input to verify the necessity of topology awareness.

[0168] 3. Models with self-supervised prediction loss removed: Models with self-supervised prediction loss removed We switched to traditional associative learning to verify memory stability.

[0169] 4. Remove spatial cell encoding model: Remove spatial cell encoding such as grid cells and position cells, and retain only GATv2 features to verify the effect of biomimetic fingerprint.

[0170] To quantify the independent contribution of each structural component, this study increased the number of nodes and edges (763 nodes and 1023 edges) during the ablation experiment, and conducted systematic ablation experiments on five model variants through 30 sets of randomized trials on the discretized lane map. The results are summarized in Table 2.

[0171] Table 2 Comparison of ablation test results of the core components of this invention

[0172] Systematic ablation experiments validated the neuroscience-inspired functional decomposition structure of the model in this invention. The hippocampus (HPC) and entorhinal cortex (EC) modules constitute the cognitive core. Replacing the HPC with cosine similarity based on graph attention networks resulted in severe path planning failure, with the success rate dropping to 40.0% and path length increasing by 109.6%, indicating that the model degenerated into goalless exploration. Similarly, removing the EC module reduced performance to the baseline level of the greedy policy, with a success rate of 66.7% and a path length increase of 110%, confirming that the EC is a key semantic interface. In contrast, the parietal cortex module, acting as a path optimizer, maintained a 100% success rate after directly replacing the PPC with multidimensional raw feature inputs, but the path length increased by 21.5% with greater fluctuations, while inference latency decreased by 97%.

[0173] The results show that HPC and EC are indispensable for cognitive map construction, while the PPC module improves the efficiency of global path planning without destroying connectivity.

[0174] To verify the physical feasibility of the model in this invention under real traffic constraints, a closed-loop simulation test process based on the SUMO Traffic Control Interface (TraCI) was designed. Unlike the traditional discrete graph node jump mode, the model in this invention achieves real-time bidirectional interaction with the SUMO simulation environment through the Python API, directly controlling the continuous motion state of the vehicle.

[0175] (1) Using a simulation-driven mechanism, within each simulation step, environmental observation information is first processed and path node selection is completed; subsequently, the global coordinates of the selected nodes are... The system accurately maps vehicle positions to Edge IDs and lane indices in the SUMO road network; ultimately, it uses TraCI-related commands to update vehicle positions in real time, completing a single-step planning loop. This mechanism ensures immediate feedback and consistency verification between neural decision-making and the physical simulation state.

[0176] (2) Decision visualization analysis: To intuitively display the decision logic of the model, a set of start-end paths is randomly selected for full-process score visualization. For example... Figure 10 As shown in the figure, after discretizing the continuous lanes, the model extracts features and scores them to select the optimal solution for the next step. Above the continuous lane graph, a score can be seen for each step. In Step 2 and Step 5, there are three candidate nodes, where yellow represents nodes with low scores and green represents the finally selected node. In Step 6, the vehicle reaches its destination. In each decision step, the model of this invention selects the graph node with the optimal cumulative reward in the current state, ultimately generating the globally optimal path solution. This result verifies the rationality of the proposed method in complex road network environments.

[0177] Example 2 Building upon Example 1, this example provides a general graph structure path planning system based on a cortical-hippocampal biomimetic mechanism to implement the general graph structure path planning method of Example 1. (See also...) Figure 6 The system includes: 1. Road network graph structure discretization module, used to convert continuous traffic road networks into standard directed graph structures.

[0178] 2. The apical cortex topology feature extraction module is used to extract global topological semantic features and vehicle node context features based on the GATv2 graph attention network.

[0179] 3. The entorhinal cortex multimodal spatial coding module integrates position cell, grid cell, boundary cell, time cell and target vector cell encoders to generate unique path node fingerprints for candidate nodes.

[0180] 4. The hippocampal memory and value assessment module is used to output a calibrated value score through a three-stage pipeline of simulated dentate gyrus pattern separation, CA3 cyclic attractor network value scoring, and CA1 value calibration.

[0181] 5. Predictive coding memory consolidation module, used to optimize the temporal transition of CA3 hidden states through a self-supervised predictive coding mechanism, to achieve end-to-end differentiable memory learning.

[0182] 6. Prefrontal decision output module, used to combine adjacency mask and temperature scaling strategy to make final path decision.

[0183] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0184] This invention can be described in the general context of computer-executable instructions, such as program modules, that are executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This invention can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0185] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0186] The above description is merely an embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principle of the present invention should be included within the scope of the claims of the present invention.

[0187] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in the present invention, and these modifications or substitutions should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A general graph-based path planning method based on a cortical-hippocampal biomimetic mechanism, characterized in that, Includes the following steps: The traffic network is discretized into a directed graph structure, where the nodes and edges in the directed graph structure correspond to multi-dimensional attribute features. The posterior parietal cortex is simulated, and a graph attention network is used to extract features from the directed graph structure to generate contextual features of the vehicle nodes. Simulating the entorhinal cortex, a unique multimodal path node fingerprint is generated for each candidate successor node of the current node by deeply integrating multiple biomimetic spatial cell encoders. The hippocampus is simulated, and based on the contextual features of the vehicle nodes, a three-level pipeline hierarchical value evaluation is performed on the fingerprints of the path nodes, outputting the calibration value score of each candidate node. The three-level pipeline includes: simulating the dentate gyrus for pattern separation, simulating the CA3 cyclic attractor network for value scoring, and simulating CA1 for global calibration. The CA3 cyclic attractor network utilizes a self-supervised predictive coding mechanism to achieve end-to-end coupling optimization of memory consolidation and deep reinforcement learning gradients by minimizing the temporal transition prediction error of the CA3 hidden state. The prefrontal cortex is simulated, and the final path decision is performed based on the calibration value score, combined with the adjacency mask constraint and temperature scaling sampling strategy, to output the optimal action.

2. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, The process of generating the context features of the vehicle node is implemented using the following formula: ， ， ， ， ， ， in, For global topological semantic features; It is the set of nodes in a directed graph structure; For the first Node characteristics of each node; For activation functions; For nodes The set of neighboring nodes; For nodes , Attention coefficient between them; For the first The node feature projection matrix of the head; , For nodes , Multidimensional attribute features; For the first Under each attention head, the node For neighboring nodes Normalized attention weights; Indicates transpose; Use the LeakyReLU activation function; This represents the learnable feature extraction matrix of the self-attention mechanism; For directed edges Multidimensional attribute features; Indicates splicing; The contextual features of the vehicle node in the global graph include local and global spatial information observed by the vehicle. Feature information of the vehicle node extracted from GATv2; , This represents two different multilayer perceptron networks; This indicates that corresponding elements are stacked. Global edge features; It is the set of edges in a directed graph structure; For classification functions; This is a learnable weight matrix used to extract global edge feature information; For directed edges Reconstructed features.

3. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, The plurality of biomimetic space cell encoders include: The position cell encoder is used to simulate the spatial encoding function of position cells, performing multi-scale periodic position encoding on the normalized coordinates of nodes. : Among them, the Frequency components , The node coordinates are the location. Number of frequency components, frequency component index ; A grid cell encoder is used to calculate the orientation angle of a candidate node relative to the current node. With normalized distance The characteristics of simulated hexagonal grid discharge are generated by multi-scale periodic functions. : ， in, For the first The period length of each scale, For phase shift; Boundary cell encoder is used to statistically analyze the out-degree and obstacle distribution characteristics of passable sectors in each direction around a candidate node; The temporal cell encoder is used to fuse three time dimensions of periodic encoding: real timestamp, training epoch, and local steps. The target vector cell encoder is used to encode the shortest path distance between candidate nodes and target nodes, the distance improvement rate, and the heading alignment. By concatenating the outputs of all encoders with the node features extracted by the graph attention network, and then performing fusion processing through a multilayer perceptron, a unique multimodal path node fingerprint is generated.

4. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, The simulated dentate gyrus is used for pattern separation by the following formula: ， in, Separate the output for the current node's pattern. For path node fingerprints, , , , These are linear dimensionality increase, nonlinear sparse activation, dropout operation, and linear dimensionality reduction operation, respectively.

5. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, The process of value scoring using the simulated CA3 cyclic attractor network includes the following steps: Constructing a gated loop unit Used to extract step memory states : ， in, Candidate nodes The stride memory state, Separate the output for the current node's pattern. For time steps The stride memory state, The vector is zero at the initial time step; Based on trainable synaptic plasticity weight matrix The stepped memory state is modified to obtain an associative memory state incorporating prior knowledge: ， in, It is the hyperbolic tangent function. The mixing coefficient; Inputting prior associative memory states into the scoring network Candidate nodes are obtained. Original value score .

6. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 5, characterized in that, The global calibration of the simulated CA1 is achieved using the following formula: ， in, Candidate nodes The calibration value score, It is a multilayer perceptron network. Candidate nodes The stride memory state, This refers to the contextual features of the vehicle node in the global graph.

7. A general graph structure path planning method based on a cortical-hippocampal biomimetic mechanism according to claim 5, characterized in that, The self-supervised predictive coding mechanism includes the following steps: By predicting the encoder According to time step CA3 hidden state prediction time step CA3 activation mode ; Calculate the predicted activation mode With actual encoding The obtained time step The cosine distance between the CA3 hidden states is used as the self-supervised predictive coding loss. ; The self-supervised prediction encoding loss Backpropagation optimization is performed as part of the total loss function to achieve end-to-end differentiable memory consolidation.

8. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, For an agent composed of the posterior parietal cortex, entorhinal cortex, hippocampus, and prefrontal cortex, the training process is optimized using a total loss function. for: ， ， ， ， ， ， ， ， ， in, , , , These are the policy gradient loss, state value regression loss, self-supervised prediction encoding loss, and policy entropy regularization term, respectively. and The loss coefficient, It is the entropy regularization coefficient. This is the set of candidate actions for the current node. It is the numerical stability constant. For time steps At the node In the case of the first One action The conditional probability, For a complete time step of a path planning from start to end, This indicates finding the minimum value. Indicates the value loss cutoff threshold. For cross-entropy loss, For time steps status The selected node is output by CA1. The value estimate, For time steps The return target, The conditional probability output by the policy network represents the probability in state. Select action The probability, For time steps t The actions performed To calculate the advantage function using an exponentially weighted average, This is the normalized advantage estimate. This represents the mean of the dominance values ​​within the current training batch. This represents the standard deviation of the dominance values ​​within the current batch. This represents a numerically stable term, used to avoid the denominator being zero. As a discount factor, For time steps TD error, yes Momentary compound rewards, To sample all time steps in the batch Take the average of experience.

9. The general graph structure path planning method based on the cortex-hippocampus biomimetic mechanism according to claim 1, characterized in that, For nodes in the directed graph structure, the multidimensional attribute features include coordinate position, node type, speed, speed limit, yaw angle, lane change sign, obstacle sign, target sign, normalized target distance, and heading alignment; for edges in the directed graph structure, the multidimensional attribute features include edge length, edge type, and edge weight.

10. A general graph-structured path planning system based on a cortical-hippocampal biomimetic mechanism, characterized in that, The system for implementing the general graph structure path planning method as described in any one of claims 1 to 9 includes: The road network graph structure discretization module is used to convert a continuous traffic road network into a standard directed graph structure; The apical cortex graph topology feature extraction module is used to extract contextual features of vehicle nodes based on the GATv2 graph attention network. The entorhinal cortex multimodal spatial coding module integrates position cell, grid cell, boundary cell, temporal cell and target vector cell encoders to generate unique path node fingerprints for candidate nodes; The hippocampal memory and value assessment module is used to output a calibrated value score through a three-stage pipeline of simulated dentate gyrus pattern separation, CA3 cyclic attractor network value scoring, and CA1 value calibration. The predictive coding memory consolidation module is used to optimize the temporal transition of the CA3 hidden state through a self-supervised predictive coding mechanism, so as to achieve end-to-end differentiable memory learning. The prefrontal decision output module is used to combine adjacency masking and temperature scaling strategies to make the final path decision.

Images