Method and device for reconstructing thinking of unmanned vehicle, electronic equipment and storage medium

Abstract

The application provides an unmanned vehicle thought reconstruction method and device, electronic equipment and a storage medium. In the method, environment information is abstracted into a node and edge associated structure by constructing a dynamic memory lattice, and the lattice structure is reconstructed in real time according to a multi-dimensional risk assessment result, so that the decision logic is no longer fixed in preset rules or a static map, but dynamically evolves with the change of the environment. The structure reconstruction mechanism based on risk assessment enables the unmanned vehicle to actively adjust the decision framework according to the real-time environment state, rather than optimizing the path parameters in a fixed framework, thereby significantly improving the decision adaptability and real-time performance in a dynamic scene, and realizing online evolution and optimization of the decision logic.

Description

Technical Field

[0001] This invention relates to the field of autonomous driving and intelligent vehicle control technology, and in particular to an autonomous vehicle mind reconstruction method, device, electronic device and storage medium. Background Technology

[0002] Currently, the decision-making and control of autonomous vehicles mainly rely on a modular "perception-planning-control" architecture. In this architecture, the system first acquires environmental data through sensors such as LiDAR and cameras to generate point clouds or grid maps; then, path planning and behavior decisions are performed based on a pre-set static rule base or an offline-trained deep learning model (such as DRL or CNN). This traditional approach treats environmental modeling and decision generation as two independent processes, and the decision logic is usually fixed in algorithm parameters or rule trees, lacking a structured storage and dynamic retrieval mechanism for historical scene experience.

[0003] Specifically, existing technologies mostly employ graph search algorithms such as Algorithm A and DLite to find the optimal path on static maps, or utilize MPC (Model Predictive Control) for local trajectory optimization. However, when faced with dynamic environmental changes, these methods cannot dynamically adjust the decision-making framework itself based on real-time risk assessment results. Their "thinking" mode is preset and fixed, making it difficult to achieve online reconstruction and evolution of decision-making logic.

[0004] Traditional autonomous vehicle decision-making systems lack the ability to reconstruct decision structures online based on dynamic environmental assessments, resulting in poor adaptability in complex dynamic scenarios and an inability to achieve real-time evolution and optimization of decision logic. Summary of the Invention

[0005] In view of this, the purpose of the present invention is to provide a method, device, electronic device and storage medium for reconstructing the thinking of unmanned vehicles, so as to alleviate the technical problem that traditional unmanned vehicle decision-making systems lack the ability to reconstruct decision structures online based on dynamic environment assessment, resulting in poor adaptability in complex dynamic scenarios and the inability to realize the real-time evolution and optimization of decision logic.

[0006] In a first aspect, embodiments of the present invention provide a method for reconstructing the mind of an autonomous vehicle, comprising: Acquire multi-source environmental information around the unmanned vehicle, and construct an initial dynamic memory lattice based on the multi-source environmental information; The dynamic memory lattice is dynamically updated according to a preset update strategy, and a multi-dimensional risk assessment is performed on the updated dynamic memory lattice. The lattice structure is then reconstructed based on the results of the multi-dimensional risk assessment. Vehicle control signals are generated based on the reconstructed lattice structure and sent to the execution unit.

[0007] Furthermore, constructing an initial dynamic memory lattice based on the multi-source environmental information includes: The multi-source environmental information is spatiotemporally synchronized based on a timestamp alignment mechanism, and standardized environmental perception results are generated through noise reduction and semantic extraction algorithms. The standardized environmental perception results and vehicle historical behavior are abstracted into nodes of the dynamic memory lattice, and feature vectors of the nodes are constructed. The nodes include: environmental nodes representing environmental elements and behavioral nodes representing vehicle behavior. The feature vectors include: spatial location, motion state, temporal characteristics, node type, historical weight, and traffic rule constraint coefficients. Based on the spatial distance threshold and semantic association logic between the nodes, node connection edges are established, and edge weights are determined according to the initial association strength to form the initial dynamic memory lattice structure.

[0008] Furthermore, dynamically updating the dynamic memory lattice according to a preset update strategy includes: In response to updates to environmental sensing data or node feature deviations exceeding a preset threshold, the target node to be updated and its set of neighboring nodes in the dynamic memory lattice are determined. Attention weights are calculated based on the spatial distance decay factor and semantic similarity between neighboring nodes and the target node. Then, the feature vectors of the neighboring nodes are transformed into the influence increments on the target node using a feature mapping function. Finally, the feature vectors of the target node are iteratively updated according to the attention weights, the influence increments, and the current feature vectors of the target node. The edge weights of the node connection edges are dynamically adjusted based on the product of the feature deviation and semantic similarity between nodes, so as to realize the real-time evolution of the dynamic memory lattice.

[0009] Furthermore, a multidimensional risk assessment is performed on the updated dynamic memory lattice, and the lattice structure is reconstructed based on the results of the multidimensional risk assessment, including: Predict future environmental changes based on the updated node feature vectors, and generate multiple candidate paths by combining them with road structure nodes; For each candidate path, the distance risk based on minimum safe distance, the rule risk based on traffic rule conformity, and the historical risk based on the weight of historical behavior nodes are calculated respectively, and the three are weighted and summed to obtain the comprehensive risk value of the candidate path. Based on the comprehensive risk value, delete the behavior nodes corresponding to high-risk paths and add behavior nodes corresponding to low-risk paths, update the risk weight attributes of obstacle nodes, and adjust the edge weights of the node connection edges in low-risk and high-risk paths to achieve the structural reconstruction of the dynamic memory lattice.

[0010] Furthermore, vehicle control signals are generated based on the reconstructed lattice structure and sent to the execution unit, including: The low-risk candidate paths and node association weights in the reconstructed lattice structure are extracted and used as constraints or guidance information for the path planning algorithm to generate path parameters containing the desired vehicle speed and desired steering angle. The path parameters are converted into standardized vehicle control signals, wherein the vehicle control signals include: a pulse width modulation signal corresponding to the throttle opening and a controller area network bus command sent to the electric power steering system. The vehicle control signal is sent to the vehicle chassis execution unit to control the unmanned vehicle to drive according to the decision logic represented by the reconstructed lattice structure.

[0011] Furthermore, after achieving the structural reconstruction of the dynamic memory lattice, the method further includes: The reconstructed lattice structure is stored as long-term memory in the historical memory bank and classified and labeled according to scene type; When the current scene is detected to match the scene type in the historical memory bank, the reconstructed lattice structure corresponding to the matching scene is extracted and used as the initialization parameter of the current dynamic memory lattice.

[0012] Furthermore, after sending the vehicle control signal to the vehicle chassis actuator, the method further includes: The actual vehicle speed and actual steering angle fed back by the vehicle chassis execution unit are obtained, and the deviation between them and the desired vehicle speed and the desired steering angle is calculated. In response to the deviation exceeding a preset threshold, a secondary update of the dynamic memory lattice is triggered, and the prediction, evaluation, and reconstruction cycle is re-executed; In response to the detection of data loss or update anomalies in the dynamic memory lattice, the system switches to emergency decision mode and uses decision parameters from similar scenarios in the historical memory bank to control the vehicle to slow down and stop.

[0013] Secondly, embodiments of the present invention also provide an unmanned vehicle mind reconstruction device, comprising: An acquisition and construction unit is used to acquire multi-source environmental information around the unmanned vehicle and construct an initial dynamic memory lattice based on the multi-source environmental information; The update and reconstruction unit is used to dynamically update the dynamic memory lattice according to a preset update strategy, perform a multi-dimensional risk assessment on the updated dynamic memory lattice, and reconstruct the lattice structure according to the multi-dimensional risk assessment results. The generation unit is used to generate vehicle control signals based on the reconstructed lattice structure and send them to the execution unit.

[0014] Thirdly, the present invention also provides an electronic device, including a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the method described in the first aspect.

[0015] Fourthly, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the method described in the first aspect.

[0016] This invention provides a method for reconstructing the thinking of an unmanned vehicle, comprising: acquiring multi-source environmental information around the unmanned vehicle and constructing an initial dynamic memory lattice based on the multi-source environmental information; dynamically updating the dynamic memory lattice according to a preset update strategy, and performing a multi-dimensional risk assessment on the updated dynamic memory lattice; reconstructing the lattice structure based on the multi-dimensional risk assessment results; generating vehicle control signals based on the reconstructed lattice structure and sending them to the execution unit. As described above, in this unmanned vehicle thinking reconstruction method, by constructing a dynamic memory lattice, environmental information is abstracted into a relational structure of nodes and edges, and the lattice structure is reconstructed in real time based on the multi-dimensional risk assessment results. This allows the decision-making logic to evolve dynamically with environmental changes, rather than being fixed to preset rules or static maps. This risk assessment-based structural reconstruction mechanism enables the unmanned vehicle to proactively adjust its decision-making framework according to the real-time environmental state, rather than simply optimizing path parameters within a fixed framework. This significantly improves the adaptability and real-time performance of decisions in dynamic scenarios, realizing the online evolution and optimization of decision-making logic, and solving the problems of rigid decision structures and inability to quickly respond to environmental changes in traditional technologies. Attached Figure Description

[0017] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0018] Figure 1 A flowchart of a method for reconstructing the mind of an unmanned vehicle provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of a driverless vehicle mind reconstruction device provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] Traditional autonomous vehicle decision-making systems lack the ability to reconstruct decision structures online based on dynamic environmental assessments, resulting in poor adaptability in complex dynamic scenarios and an inability to achieve real-time evolution and optimization of decision logic.

[0021] Based on this, the autonomous vehicle thinking reconstruction method of the present invention constructs a dynamic memory lattice, abstracts environmental information into a relational structure of nodes and edges, and reconstructs the lattice structure in real time according to the results of multi-dimensional risk assessment. This allows the decision-making logic to evolve dynamically with changes in the environment, rather than being fixed to preset rules or static maps. This risk assessment-based structural reconstruction mechanism enables the autonomous vehicle to actively adjust its decision-making framework according to the real-time environmental state, rather than simply optimizing path parameters within a fixed framework. This significantly improves the adaptability and real-time performance of decision-making in dynamic scenarios, and realizes the online evolution and optimization of decision-making logic.

[0022] To facilitate understanding of this embodiment, a method for reconstructing the thinking of an unmanned vehicle disclosed in this embodiment of the invention will first be described in detail.

[0023] Example 1: According to an embodiment of the present invention, an embodiment of a method for reconstructing the thinking of an autonomous vehicle is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0024] Figure 1 This is a flowchart of a method for reconstructing the mind of an autonomous vehicle according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps: Step S100: Obtain multi-source environmental information around the unmanned vehicle, and construct an initial dynamic memory lattice based on the multi-source environmental information.

[0025] This method constructs and dynamically maintains a "dynamic memory lattice" that simulates the human cognitive process, integrating the autonomous vehicle's perception, memory, and decision-making processes in a unified graph data structure, thereby achieving end-to-end cognitive reconstruction from environmental information to control commands.

[0026] The "dynamic memory lattice" (i.e., thinking) referred to in this embodiment is a topological data model used to characterize the environmental state and behavioral logic of an autonomous vehicle. This lattice consists of several "nodes" and "edges" connecting these nodes. Nodes are used to abstract various elements in the environment (such as obstacles, traffic signs, lane lines) and the vehicle's own historical or candidate behaviors (such as acceleration, steering, and parking); edges are used to characterize the spatiotemporal relationships, causal logic, or semantic dependencies between nodes. Unlike traditional static maps or instantaneous perception snapshots, the dynamic memory lattice possesses temporal evolution capabilities; its node features and edge weights can be iteratively updated in real time with the input of environmental information, thereby simulating the dynamic cognition and experience accumulation process of a human driver regarding the road environment.

[0027] In this step, the autonomous vehicle collects raw data about its surrounding environment using various onboard sensors. These multi-source environmental information includes, but is not limited to: point cloud data collected by LiDAR to obtain the 3D contours and distance information of obstacles; image data collected by cameras to identify semantic information such as traffic light colors, lane line types, and pedestrian postures; target trajectory data collected by millimeter-wave radar to supplement obstacle speed information in adverse weather conditions; and location and intent information of surrounding vehicles received by the vehicle-to-everything (V2X) communication module. After the system performs spatiotemporal synchronization and fusion processing on this heterogeneous data, it extracts key environmental elements and the vehicle's own state, abstracting them as initial nodes and connecting edges in a lattice, thereby constructing the initial dynamic memory lattice for the current moment. This process is equivalent to a human driver's "first glance" and "initial impression" of the road conditions.

[0028] Step S200: Dynamically update the dynamic memory lattice according to the preset update strategy, perform a multi-dimensional risk assessment on the updated dynamic memory lattice, and reconstruct the lattice structure based on the multi-dimensional risk assessment results.

[0029] As the vehicle moves, the surrounding environment is constantly changing. In this step, the system does not discard old memories but instead employs a dynamic update mechanism. When the sensors detect environmental changes (such as a vehicle ahead decelerating) or a change in the vehicle's own state, the system locates the corresponding node in the lattice, uses the new observation data to correct its feature vector (such as updating the position coordinates and velocity vector of obstacles), and adjusts the edge weights based on changes in the correlation strength between nodes. For example, when the distance between the vehicle and an obstacle ahead decreases, the edge weights between the node representing the obstacle and the node representing the "deceleration" behavior will be correspondingly increased.

[0030] After updating the nodes and edges, the system performs a multi-dimensional risk assessment based on the updated lattice structure. This assessment considers not only the physical distance to obstacles (distance risk) but also traffic rule constraints (rule risk) and historical behavior success rates (historical risk), calculating the risk value of each candidate behavioral path under the current lattice structure. Based on the assessment results, the system reconstructs the lattice structure: deleting high-risk behavioral nodes (such as illegal lane changes), adding or strengthening low-risk behavioral nodes (such as safe following), and adjusting the connections between nodes. This process simulates the "thinking and decision-making correction" of human drivers in complex road conditions, ensuring the lattice structure remains in an optimal decision-making state through a continuous "prediction-evaluation-reconstruction" cycle.

[0031] Step S300: Generate vehicle control signals based on the reconstructed lattice structure and send them to the execution unit.

[0032] After the aforementioned thought reconstruction process, the reconstructed lattice structure contains the optimal decision-making path information. In this step, the system extracts the desired driving path parameters from the lattice, such as the desired vehicle speed and desired steering angle, and converts them into control signals recognizable by the vehicle chassis actuators. Specific forms of these control signals include, but are not limited to, pulse width modulation (PWM) signals corresponding to throttle opening or braking pressure, and Controller Area Network (CAN) bus commands sent to the Electric Power Steering (EPS) system. Upon receiving the signals, the actuators drive the vehicle's throttle, brakes, or steering mechanism, thereby controlling the autonomous vehicle to drive safely according to the decision-making logic.

[0033] This embodiment constructs a thinking and decision-making framework with dynamic memory and real-time reconstruction capabilities through the above steps. This framework organically integrates environmental perception, historical experience, and decision-making planning, enabling autonomous vehicles to quickly recall memories, assess risks, and reconstruct decisions like humans when facing unexpected situations (such as pedestrians crossing the road), significantly improving the environmental adaptability and decision-making safety of the autonomous driving system.

[0034] In an optional embodiment of the present invention, step S100, which involves constructing an initial dynamic memory lattice based on multi-source environmental information, specifically includes the following steps: Step S101: Based on the timestamp alignment mechanism, multi-source environmental information is spatiotemporally synchronized, and standardized environmental perception results are generated through denoising processing and semantic extraction algorithms.

[0035] Because the sampling frequencies and transmission delays of sensors such as LiDAR, cameras, and millimeter-wave radar on autonomous vehicles vary, directly using raw data can lead to spatiotemporal misalignment in environmental modeling. Therefore, this embodiment employs high-precision timing signals based on the Global Navigation Satellite System (GNSS) to assign a unified timestamp to each frame of perception data, mapping all data to the same time reference, with synchronization accuracy controllable at the millisecond level. For example, the point cloud data of the LiDAR at time T is aligned with the image data of the camera at time T. After completing spatiotemporal synchronization, the system performs denoising processing to remove noise points and uses semantic extraction algorithms to identify environmental elements. For example, the RANSAC algorithm is used to remove ground point clouds, and the YOLOv8 target detection model is used to extract semantic information such as lane lines, traffic signs, and pedestrians from the images, ultimately forming a standardized environmental perception result containing attributes such as target ID, type, location, and speed. It should be understood that the above algorithm is only an example, and those skilled in the art can also use other mature denoising algorithms or semantic segmentation networks to implement it, as long as standardized environmental perception results can be generated.

[0036] Step S102: The standardized environmental perception results and vehicle historical behavior are abstracted into nodes of a dynamic memory lattice, and the feature vectors of the nodes are constructed.

[0037] This embodiment abstracts the complex traffic environment into two types of nodes: environment nodes and behavior nodes. Environment nodes represent environmental elements, such as obstacle nodes (e.g., vehicles ahead, pedestrians), traffic rule nodes (e.g., traffic light status, speed limit signs), and road structure nodes (e.g., lane lines, intersection areas). Behavior nodes represent vehicle behavior, such as historical path nodes (e.g., driving trajectory points over a past period) and decision candidate nodes (e.g., candidate actions such as "go straight," "turn left," and "slow down to avoid"). Each node is assigned a multi-dimensional feature vector to quantify its state attributes. This feature vector includes at least the following six dimensions: spatial location (e.g., three-dimensional coordinates x, y, z), motion state (e.g., velocity vector). , , The system incorporates various features, including time characteristics (such as timestamps of data generation), node types (such as codes to distinguish obstacles, rules, and behaviors), historical weights (node ​​weights, such as values ​​normalized based on the number of successful historical executions), and traffic rule constraint coefficients (such as higher constraint coefficients during red lights and lower ones during green lights). By constructing feature vectors containing rich attributes, the system can fully preserve the physical and logical attributes of environmental elements within the lattice structure, providing data support for subsequent graph computations.

[0038] Step S103: Establish node connection edges based on the spatial distance threshold and semantic association logic between nodes, and determine the edge weights according to the initial association strength to form the initial dynamic memory lattice structure.

[0039] After constructing the nodes, connections between them need to be established to form a topology. This embodiment employs a dual-logic approach to establish these connections: first, a spatial distance threshold logic, where a connection is established between two environmental nodes when their Euclidean distance is less than a preset threshold (e.g., 50 meters) to characterize the environmental correlation within the local space; second, a semantic association logic, where a connection is forcibly established when two nodes have a strong semantic association (e.g., a "red light" node and a "stop" behavior node, or a "obstacle ahead" node and a "avoid" behavior node) to characterize a causal or logical dependency. The initial weight of the edge is determined based on the strength of the association between the nodes. For example, the closer the spatial distance or the higher the semantic association, the larger the initial edge weight (e.g., set to 0.8), and vice versa (e.g., set to 0.2). Through these steps, the system successfully transforms discrete multi-source sensing data into a structured graph model, i.e., the initial dynamic memory lattice. This structured representation not only integrates multi-source data but also explicitly models the potential correlation between environmental elements and vehicle behavior, providing a solid topological foundation for subsequent dynamic updates and risk assessments.

[0040] In an optional embodiment of the present invention, dynamically updating the dynamic memory lattice according to a preset update strategy specifically includes the following steps: Step S201: In response to the update of environmental perception data or the node feature deviation exceeding a preset threshold, determine the target node to be updated and its set of neighboring nodes in the dynamic memory lattice.

[0041] The dynamic nature of the crystal lattice structure is reflected in its ability to evolve in real time with environmental changes. Update triggering conditions are divided into two categories: active updates, where the system automatically updates relevant nodes when sensors collect new data at a preset frequency (e.g., every 2 milliseconds); and passive updates, triggered when a node's characteristic deviation exceeds a preset threshold. For example, when an obstacle node's position deviation exceeds 0.5 meters, or its speed deviation exceeds 1 meter per second, or a traffic light status change occurs abruptly (e.g., from green to red), the system determines that node as the target node to be updated. After identifying the target node, the system searches the crystal lattice structure to find all nodes connected to the target node by edges, forming a neighbor node set. This mechanism ensures that lattice updates are both periodic and capable of rapid response to sudden events.

[0042] Step S202: Calculate attention weights based on the spatial distance decay factor and semantic similarity between neighboring nodes and the target node, and use the feature mapping function to transform the feature vectors of neighboring nodes into the influence increments on the target node. Then, iteratively update the feature vector of the target node based on the attention weights, influence increments, and the current feature vector of the target node.

[0043] In dynamic environments, the influence of different neighboring nodes on the target node varies. This embodiment introduces an attention mechanism to quantify this influence. The calculation of attention weights comprehensively considers two dimensions: spatial distance and semantic relevance. Specifically, neighboring nodes... With the target node attention weights The calculation formula can be expressed as: .in, For neighboring nodes With the target node The spatial Euclidean distance between them This is the distance attenuation coefficient (the preferred value can be set to 50 meters, indicating that the influence between nodes decreases exponentially with distance). The semantic similarity is calculated (ranging from 0 to 1; for example, the semantic similarity between the "pedestrian" node and the "avoid" behavior node can be set to 0.9, while the similarity with the "accelerate" behavior node is only 0.1). According to this formula, nodes that are closer in distance and have a stronger semantic connection receive a higher attention weight and have a greater impact on the target node.

[0044] After calculating the attention weights, the system uses the feature mapping function. neighbor nodes eigenvectors Transformed into target node The incremental impact. The function of feature mapping is to extract the changes in features among neighboring nodes that are relevant to the target node. For example, when a neighboring node (the vehicle in front) suddenly decelerates, its speed change is recorded. The feature mapping function transforms the data into speed adjustment suggestions for the target node (the vehicle's behavior node). Finally, the iterative update formula for the target node's feature vector is: .in, This represents the feature vector of the target node at the current time step. The updated feature vector is represented by the summation symbol, which indicates summation over all neighboring nodes. (Target node) The influence increment of the neighboring node set is weighted and accumulated. Through this iterative update, the nodes in the lattice can absorb the changing information of the surrounding environment in real time, realizing dynamic correction of features.

[0045] Step S203: Based on the product of the feature deviation and semantic similarity between nodes, the edge weights of the node connection edges are dynamically adjusted to achieve the real-time evolution of the dynamic memory lattice.

[0046] The strength of the association between nodes (i.e., edge weights) is not fixed but dynamically adjusted as the node states change. In this embodiment, the adjustment rule for edge weights is mainly based on the product of the feature deviation and semantic similarity between nodes, i.e. ,in, To adjust the coefficient, For the target node with neighboring nodes Characteristic deviation, For semantic similarity, The adjusted edge weights, The original edge weights are used. Specifically, when two nodes have small feature deviations (e.g., consistent movement trends) and high semantic similarity, it indicates a close connection, and the system will increase the weight of their connecting edge. Conversely, if the feature deviation increases or there is semantic conflict, the edge weights will decrease. For example, when a "red light" node and a "go straight" behavior node coexist, due to semantic conflict, the weight of their connecting edge will be significantly reduced, thus suppressing the propagation of illegal decisions. Through dynamic adjustment of edge weights, the lattice structure can adaptively strengthen reasonable connection paths and weaken conflicting or outdated connection paths, thereby achieving real-time evolution of the topology and providing a more accurate decision model for subsequent risk assessment.

[0047] In an optional embodiment of the present invention, a multidimensional risk assessment is performed on the updated dynamic memory lattice, and the lattice structure is reconstructed based on the multidimensional risk assessment results, specifically including the following steps: Step S204: Predict future environmental changes based on the updated feature vectors of the nodes, and generate multiple candidate paths by combining them with road structure nodes.

[0048] After dynamically updating the lattice nodes, the system possesses the latest environmental awareness at the current moment. To make decisions, the system needs to predict environmental change trends over a future period (e.g., 1 to 5 seconds). For example, for obstacle nodes, the system predicts their future position using a kinematic model based on their current motion state (velocity, acceleration); for traffic rule nodes (e.g., traffic lights), the system predicts the timing of their state changes. Based on this prediction information and road structure nodes (e.g., lane lines, intersection topology), the system searches and generates multiple candidate paths within the lattice topology. Each candidate path is essentially a continuous sequence of action nodes, such as "stay in the current lane and go straight," "change lanes to the left to avoid an obstacle," or "slow down and stop." It should be understood that the number of candidate paths can be dynamically adjusted based on the availability of computing resources; this embodiment does not limit this.

[0049] Step S205: For each candidate path, calculate the distance risk based on the minimum safe distance, the rule risk based on the traffic rule conformity, and the historical risk based on the weight of historical behavior nodes, and sum the three risks to obtain the comprehensive risk value of the candidate path.

[0050] This embodiment constructs a multi-dimensional risk assessment model to quantify the security of each candidate path. The model comprehensively considers three dimensions: physical security, regulatory constraints, and historical experience, ensuring the comprehensiveness and compliance of the decision-making process.

[0051] Specifically, distance from risk It is used to characterize the potential probability of a path colliding with an obstacle. Its calculation depends on the minimum safe distance between each point on the path and the obstacle node. Generally speaking, distance risk There is a negative correlation with the minimum safe distance; that is, the closer the distance, the higher the risk. For example, it can be calculated using an exponential function: ,in, This is the adjustment coefficient. When When it approaches infinity, Approaching 0; when When it approaches 0, It approaches 1. This reflects the rigid requirement for physical distance in safe driving.

[0052] Regulatory risks This is used to characterize the degree of conformity between a route and traffic rules. The system calculates this risk value by querying traffic rule nodes (such as speed limit signs and traffic light status) in the lattice. If the route violates traffic rules (such as running a red light or driving against traffic), the rule risk is calculated. A higher value (such as 0.9 or 1.0) is assigned if the path fully conforms to the rules; a lower value (such as 0.1) is assigned if the path fully conforms to the rules. The introduction of this dimension ensures the legality and social compliance of the autonomous vehicle's decision-making.

[0053] Historical risks This is used to characterize the reliability of a path based on historical experience. The system extracts historical weight attributes from the feature vectors of the behavioral nodes corresponding to the path. If the behavior has been successfully executed less frequently in similar historical scenarios or has experienced incidents, then historical risk is considered. Higher; conversely, if the historical execution record is good, then The score is relatively low. This dimension reflects the system's "memory" and "reuse" of historical experience, which helps avoid repeating mistakes.

[0054] Ultimately, the overall risk value The calculation formula is in the form of a weighted summation: ,in, , , These are the weighting coefficients for distance risk, rule risk, and historical risk, respectively, and they satisfy... + + =1. In a preferred embodiment of this invention, considering that driving safety is the primary objective, the weight of distance risk is... Set to the highest value, for example =0.5, =0.3, =0.2. It should be understood that these weighting coefficients are not fixed. The system can dynamically adjust the weight ratios based on different driving modes (such as aggressive mode and fuel-efficient mode) or road environments (such as highways and campus roads) to adapt to different decision-making preferences. For example, in highway scenarios, the weight of distance risk can be appropriately increased; in complex intersection scenarios, the weight of rule risk can be appropriately increased. Through this multi-dimensional weighted evaluation, the system can output a quantified comprehensive risk value. This provides a precise basis for decision-making in subsequent lattice reconstruction.

[0055] Step S206: Based on the comprehensive risk value, delete the behavior nodes corresponding to high-risk paths and add behavior nodes corresponding to low-risk paths, update the risk weight attributes of obstacle nodes, and adjust the edge weights of the node connection edges in low-risk and high-risk paths to achieve the structural reconstruction of the dynamic memory lattice.

[0056] The ultimate goal of risk assessment is to guide the optimization of crystal lattice structures. This step reconstructs the nodes and edges in the crystal lattice through a "survival of the fittest" mechanism, causing the crystal structure to evolve towards a lower-risk direction.

[0057] Specifically, the refactoring operation includes three levels: First, there's the addition and deletion of nodes. The system sets risk thresholds (such as...). =0.6), when the comprehensive risk value of a candidate path is... When the risk level exceeds this threshold, the system determines the path as a high-risk path and deletes the corresponding behavior node (such as the "forced lane change" node), thus removing dangerous options from the decision space. Conversely, if the overall risk value of a path exceeds this threshold... Below another threshold (e.g.) =0.3), the system determines that the path is a low-risk path, and adds or strengthens the corresponding behavior node (such as the "smooth following" node) in the lattice, and uses it as the preferred decision.

[0058] Second, there are attribute-level updates. For environmental nodes, especially obstacle nodes, the system updates the risk weight attribute in their feature vectors based on their association risk with candidate paths. For example, if an obstacle node is located on a high-risk path, its risk weight is increased, giving it higher priority in subsequent perception and updates, simulating a human driver's "focus on" hazards.

[0059] Thirdly, adjustments are made at the edge weight level. The system strengthens the edge weights connecting nodes on low-risk paths, making information transmission between these nodes smoother; simultaneously, it weakens the edge weights connecting nodes on high-risk paths, suppressing the propagation of dangerous decision-making logic. For example, the edge weight between the "vehicle ahead" node and the "decelerate" behavior node on a low-risk path is increased from 0.5 to 0.8, while the edge weight between the "vehicle ahead" node and the "accelerate overtake" behavior node on a high-risk path is decreased from 0.4 to 0.1. Through these multi-dimensional reconstruction operations, the dynamic memory lattice can eliminate high-risk decisions in real time, strengthen safety strategies, and achieve self-evolution and optimization of decision-making thinking.

[0060] In an optional embodiment of the present invention, generating vehicle control signals based on the reconstructed lattice structure and sending them to the execution unit specifically includes the following steps: Step S301: Extract low-risk candidate paths and node association weights from the reconstructed lattice structure, and use them as constraints or guidance information for the path planning algorithm to generate path parameters containing the desired vehicle speed and desired steering angle.

[0061] After risk assessment and reconstruction, high-risk behavioral nodes have been removed from the dynamic memory lattice, retaining the overall risk value. Low-risk candidate paths. This step first extracts the topological sequence of these low-risk paths from the lattice topology. This sequence contains not only the geometric information of the paths (such as a series of coordinate points) but also the association weights between nodes. These weights reflect the strength of the association between behavioral nodes and environmental nodes and can serve as important inputs to the path planning algorithm. For example, the low-risk paths output by the lattice can be used as state constraints for a Model Predictive Control (MPC) algorithm or as sampling guidance information for a Rapid Exploratory Random Tree (RRT) algorithm. Based on these constraints and guidance, the path planning algorithm calculates the specific driving trajectory and further solves for the expected vehicle speed and expected steering angle required for the autonomous vehicle to follow the trajectory. It should be understood that the specific choice of path planning algorithm does not constitute a limitation of the present invention, and any algorithm that can output control parameters based on the lattice structure is within the protection scope of the present invention.

[0062] Step S302: Convert the path parameters into standardized vehicle control signals, wherein the vehicle control signals include: a pulse width modulation signal corresponding to the throttle opening and a controller area network bus command sent to the electric power steering system.

[0063] Since the desired vehicle speed and desired steering angle are physical control targets, and vehicle chassis actuators typically require specific electrical signals to operate, signal conversion is necessary. This embodiment employs a standardized conversion protocol to convert the desired vehicle speed into a pulse width modulation (PWM) signal corresponding to the throttle opening or brake pressure. The PWM signal controls the average voltage of the actuator by changing the pulse width, thereby precisely controlling the throttle opening or brake fluid pressure. For example, when the desired vehicle speed increases, the system increases the duty cycle of the PWM signal to increase the throttle opening; when the desired deceleration occurs, the duty cycle is decreased or a reverse PWM signal is output to control the brakes. Simultaneously, the desired steering angle is converted into a Controller Area Network (CAN) bus command. The CAN bus is a standard protocol for communication between electronic control units (ECUs) within a vehicle, characterized by high real-time performance and high reliability. The system encapsulates the steering angle data into a CAN data frame and sends it to the Electric Power Steering (EPS) system. Upon receiving the command, the EPS controls the motor to output the corresponding torque, driving the steering mechanism. Through this conversion, a precise mapping from digital decision-making to physical execution is achieved.

[0064] Step S303: The vehicle control signal is sent to the vehicle chassis execution unit to control the unmanned vehicle to drive according to the decision logic represented by the reconstructed lattice structure.

[0065] After generating standardized control signals, the system sends the signals to the corresponding execution units via the vehicle's drive-by-wire chassis interface. Throttle / brake control signals are sent to the engine electronic control unit (ECU) or electronic stability control system (ESC), while steering control signals are sent to the EPS system. Upon receiving the commands, the chassis execution units drive the corresponding motors, hydraulic valves, or relays to change the vehicle's driving state. This process completes the closed loop from "thought reconstruction" to "physical execution," ensuring that the autonomous vehicle can drive strictly according to the decision logic planned by the dynamic memory lattice. For example, when the lattice decision is "avoid pedestrians," the generated control signal will drive the vehicle to decelerate and turn left until the risk value is reduced below the safety threshold. Through the above steps, this embodiment achieves a complete transformation from abstract graph structure data to underlying hardware control commands, ensuring the effective execution of the decision-making scheme.

[0066] In an optional embodiment of the present invention, after realizing the structural reconstruction of the dynamic memory lattice, the method further includes: Step S401: The reconstructed lattice structure is stored as long-term memory in the historical memory bank and classified and labeled according to the scene type.

[0067] During the autonomous vehicle's operation, after each successful lattice reconstruction and execution of avoidance or passage tasks, the current lattice structure actually contains the optimal decision-making logic and environmental feature association for that specific scenario. To enable the autonomous vehicle to possess "experience accumulation" capabilities similar to humans, the system persistently stores these reconstructed lattice structures. During storage, the system extracts key feature tags of the lattice as scenario types for classification. The dimensions for classifying scenario types include, but are not limited to: road topology (such as intersections, T-junctions, ramp entrances), traffic participant types (such as pedestrian-intensive areas, congested road sections), weather and lighting conditions (such as rainy days, nighttime), and specific traffic events (such as ambulance avoidance, construction zone detours). For example, when an autonomous vehicle successfully avoids a crossing electric bicycle at an intersection without traffic lights, the system will mark the lattice structure at this time as "no signalized intersection - crossing vehicle avoidance" and store it in the historical memory database. This categorized storage mechanism allows the historical memory bank to gradually form an "experience knowledge graph" covering a variety of complex working conditions, providing data support for rapid decision-making in the future.

[0068] Step S402: When the current scene is detected to match the scene type in the historical memory bank, the reconstructed lattice structure corresponding to the matching scene is extracted and used as the initialization parameter of the current dynamic memory lattice.

[0069] When the autonomous vehicle enters a similar environment again, it doesn't need to build a lattice from scratch and iterate step by step; instead, it can directly call upon historical experience. The system extracts the feature vectors of the current environment in real time and calculates the matching degree with the feature labels of each scene type in the historical memory bank. The matching algorithm can use cosine similarity or Euclidean distance as a metric. When the similarity exceeds a preset threshold (e.g., 0.85), the scene is considered successfully matched. At this point, the system directly extracts the lattice structure corresponding to the matched scene, loads its node attributes and edge weight parameters into the current dynamic memory lattice, and uses it as the initialization state.

[0070] This mechanism delivers significant technical benefits. Traditional initialization methods typically use default parameters or a blank state, requiring multiple rounds of perception and updates to converge to a stable state. This embodiment, however, achieves "hot start" by reusing long-term memory. For example, when the autonomous vehicle encounters an uncontrolled intersection again, the system directly loads historical lattice parameters. At this point, the weights of the behavioral nodes representing "deceleration observation" and "lateral detection" have been pre-strengthened, and the connection edges between nodes have also been pre-set with reasonable weights. This allows the autonomous vehicle to skip the initial tentative perception phase and directly enter a high-precision decision-making state. According to measured data, this long-term memory reuse mechanism can shorten the decision response time in similar scenarios from the traditional 50 milliseconds to approximately 30 milliseconds, significantly improving the reaction speed and decision accuracy of the autonomous vehicle in repetitive or similar complex scenarios. This effectively solves the problem in existing technologies where autonomous vehicles "remember mistakes but not lessons" and cannot utilize historical experience.

[0071] In an optional embodiment of the present invention, after sending the vehicle control signal to the vehicle chassis actuator, the method further includes: Step S501: Obtain the actual vehicle speed and actual steering angle fed back by the vehicle chassis execution unit, and calculate the deviation between them and the desired vehicle speed and desired steering angle.

[0072] This process details the execution feedback mechanism and emergency decision-making mode after the vehicle control signal is sent. In the actual operation of autonomous vehicles, simply issuing control commands is insufficient to guarantee that the vehicle will drive exactly as intended. Factors such as changes in road friction, actuator response delays, or sensor noise can all lead to deviations between the actual execution results and the intended decision. Therefore, this embodiment introduces closed-loop feedback and safety redundancy strategies to ensure the robustness of the system.

[0073] While executing control commands, the chassis execution units (such as the engine ECU and EPS system) provide real-time feedback on the current operating status via the vehicle's CAN bus. The system reads this feedback data to obtain the actual vehicle speed and actual steering angle. Subsequently, the system calculates the deviation between the actual and expected values. For example, if the expected speed is 20 m / s and the actual feedback speed is 19 m / s, the speed deviation is 1 m / s; if the expected steering angle is 5 degrees and the actual steering angle is 4.5 degrees, the steering angle deviation is 0.5 degrees. This process enables real-time monitoring of the execution effect, giving the autonomous vehicle the ability to "perceive the consequences of its own actions."

[0074] In step S502, in response to the deviation exceeding the preset threshold, a second update of the dynamic memory lattice is triggered, and the prediction, evaluation and reconstruction cycle is re-executed.

[0075] The system has a preset deviation threshold, which can be dynamically adjusted based on the vehicle's current speed or road conditions. For example, at high speeds, the speed deviation threshold can be set to 2 m / s, and the steering angle deviation threshold to 1 degree; while in low-speed parking scenarios, the threshold can be tightened accordingly. When the calculated deviation exceeds the preset threshold, it indicates a disconnect between the current lattice decision logic and the actual execution in the physical world. At this point, the system determines that the current decision may no longer be applicable and immediately triggers a secondary update of the dynamic memory lattice. The system will then re-execute a multi-dimensional risk assessment of the updated dynamic memory lattice, reconstruct the lattice structure based on the multi-dimensional risk assessment results, and replan the path or adjust the control parameters based on the latest vehicle status. This closed-loop mechanism effectively eliminates the accumulation of control errors, ensuring that the autonomous vehicle can promptly self-correct when faced with execution deviations, and avoiding continuous errors caused by a single erroneous decision.

[0076] In step S503, in response to the detection of data loss or update anomaly in the dynamic memory lattice, the system switches to emergency decision mode and calls decision parameters of similar scenarios in the historical memory bank to control the vehicle to decelerate and stop.

[0077] As the system's last line of defense, when a serious fault is detected in the dynamic memory lattice, such as prolonged loss of sensor data, non-numerical (NaN) anomalies in lattice node updates, or computation timeouts, the system will proactively disconnect from the regular decision-making process and switch to emergency decision-making mode. In this mode, the system no longer performs complex lattice evolution but directly calls historical decision parameters from the historical memory bank that are most similar to the characteristics of the current scene (such as road type and weather conditions). For example, if the current scene is a "rainy urban road," the system will extract the safe deceleration curve and parking parameters for this type of scene and control the vehicle to decelerate smoothly until it stops. Compared to unknown real-time calculations, calling verified historical parameters has higher determinism and security. This safety redundancy strategy ensures that even in the extreme case of core decision-making module failure, the autonomous vehicle can still execute the minimum-risk strategy, avoiding serious traffic accidents and significantly improving the system's functional safety level.

[0078] The present invention has the following effects: Improved decision-making accuracy: Through dynamic memory lattices, vehicles can retain historical environmental and behavioral patterns for extended periods, enhancing the foresight and stability of decision-making. Enhanced environmental adaptability: The system can reconstruct thought structures in real time, respond quickly to dynamic obstacles and emergencies, and improve the success rate of path planning; High efficiency in utilizing multi-source data: GCN combined with attention mechanisms enables deep information fusion and makes full use of sensor and communication information; Excellent interpretability: The state and weights of lattice nodes and edges are visualized, providing traceability of the vehicle decision-making process.

[0079] Example 2: This invention also provides an unmanned vehicle mind reconstruction device, which is mainly used to execute the unmanned vehicle mind reconstruction method provided in Embodiment 1 of this invention. The unmanned vehicle mind reconstruction device provided in this invention will be described in detail below.

[0080] Figure 2 This is a schematic diagram of an unmanned vehicle mind reconstruction device according to an embodiment of the present invention, such as... Figure 2 As shown, the device mainly includes: an acquisition and construction unit 10, an update and reconstruction unit 20, and a generation unit 30, wherein: Acquisition and construction unit 10 is used to acquire multi-source environmental information around the unmanned vehicle and construct an initial dynamic memory lattice based on the multi-source environmental information; The update and reconstruction unit 20 is used to dynamically update the dynamic memory lattice according to a preset update strategy, perform a multi-dimensional risk assessment on the updated dynamic memory lattice, and reconstruct the lattice structure according to the multi-dimensional risk assessment results. The generation unit 30 is used to generate vehicle control signals based on the reconstructed lattice structure and send them to the execution unit.

[0081] This invention provides an autonomous vehicle mind reconstruction device, comprising: acquiring multi-source environmental information around the autonomous vehicle and constructing an initial dynamic memory lattice based on the multi-source environmental information; dynamically updating the dynamic memory lattice according to a preset update strategy, and performing a multi-dimensional risk assessment on the updated dynamic memory lattice; reconstructing the lattice structure based on the multi-dimensional risk assessment results; generating vehicle control signals based on the reconstructed lattice structure and sending them to the execution unit. As described above, in the autonomous vehicle mind reconstruction device of this invention, by constructing a dynamic memory lattice, environmental information is abstracted into a relational structure of nodes and edges, and the lattice structure is reconstructed in real time based on the multi-dimensional risk assessment results. This allows the decision-making logic to evolve dynamically with environmental changes, rather than being fixed to preset rules or static maps. This risk assessment-based structural reconstruction mechanism enables the autonomous vehicle to proactively adjust its decision-making framework according to the real-time environmental state, rather than simply optimizing path parameters within a fixed framework. This significantly improves the adaptability and real-time performance of decision-making in dynamic scenarios, realizing the online evolution and optimization of decision-making logic, and solving the problems of rigid decision-making structures and inability to quickly respond to environmental changes in traditional technologies.

[0082] Optionally, the acquisition and construction unit is also used to: perform spatiotemporal synchronization of multi-source environmental information based on a timestamp alignment mechanism, and generate standardized environmental perception results through denoising processing and semantic extraction algorithms; abstract the standardized environmental perception results and vehicle historical behavior into nodes of a dynamic memory lattice, and construct feature vectors for the nodes, wherein the nodes include: environmental nodes representing environmental elements and behavioral nodes representing vehicle behavior, and the feature vectors include: spatial location, motion state, temporal features, node type, historical weight, and traffic rule constraint coefficients; establish node connection edges based on the spatial distance threshold and semantic association logic between nodes, and determine the edge weights according to the initial association strength to form an initial dynamic memory lattice structure.

[0083] Optionally, the update and reconstruction unit is further configured to: in response to updates to environmentally perceived data or node feature deviations exceeding a preset threshold, determine the target node to be updated and its set of neighboring nodes in the dynamic memory lattice; calculate attention weights based on the spatial distance decay factor and semantic similarity between neighboring nodes and the target node, and use a feature mapping function to transform the feature vectors of neighboring nodes into an influence increment on the target node, and then iteratively update the feature vector of the target node according to the attention weights, influence increments, and the current feature vector of the target node; and dynamically adjust the edge weights of the node connection edges according to the product of feature deviations and semantic similarity between nodes, so as to realize the real-time evolution of the dynamic memory lattice.

[0084] Optionally, the updating and reconstructing unit is also used to: predict future environmental changes based on the feature vectors of the updated nodes, and generate multiple candidate paths in combination with road structure nodes; for each candidate path, calculate the distance risk based on the minimum safe distance, the rule risk based on the traffic rule conformity, and the historical risk based on the weight of historical behavior nodes, and sum the three to obtain the comprehensive risk value of the candidate path; based on the comprehensive risk value, delete the behavior nodes corresponding to high-risk paths and add the behavior nodes corresponding to low-risk paths, update the risk weight attributes of obstacle nodes, and adjust the edge weights of the node connection edges in low-risk paths and high-risk paths to achieve the structural reconstruction of the dynamic memory lattice.

[0085] Optionally, the generation unit is also used to: extract low-risk candidate paths and node association weights from the reconstructed lattice structure, use them as constraints or guidance information for the path planning algorithm, and generate path parameters containing the desired vehicle speed and desired steering angle; convert the path parameters into standardized vehicle control signals, wherein the vehicle control signals include: pulse width modulation signals corresponding to the throttle opening and controller area network bus commands sent to the electric power steering system; and send the vehicle control signals to the vehicle chassis execution unit to control the unmanned vehicle to drive according to the decision logic represented by the reconstructed lattice structure.

[0086] Optionally, the device is also used to: store the reconstructed lattice structure as a long-term memory in a historical memory bank and classify and label it according to scene type; when the current scene is detected to match the scene type in the historical memory bank, extract the reconstructed lattice structure corresponding to the matching scene and use it as the initialization parameter of the current dynamic memory lattice.

[0087] Optionally, the device is also used to: acquire the actual vehicle speed and actual steering angle fed back by the vehicle chassis execution unit, and calculate the deviation between them and the desired vehicle speed and desired steering angle; in response to the deviation exceeding a preset threshold, trigger a secondary update of the dynamic memory lattice and re-execute the prediction, evaluation and reconstruction cycle; in response to the detection of data loss or update anomaly in the dynamic memory lattice, switch to emergency decision mode and call decision parameters of similar scenarios in the historical memory bank to control the vehicle to decelerate and stop.

[0088] The device provided in this embodiment of the invention has the same implementation principle and technical effect as the aforementioned method embodiment. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the aforementioned method embodiment.

[0089] like Figure 3 As shown in the embodiment of this application, an electronic device 600 includes a processor 601, a memory 602, and a bus. The memory 602 stores machine-readable instructions that can be executed by the processor 601. When the electronic device is running, the processor 601 communicates with the memory 602 via the bus, and the processor 601 executes the machine-readable instructions to perform the steps of the above-described autonomous vehicle mind reconstruction method.

[0090] Specifically, the aforementioned memory 602 and processor 601 can be general-purpose memory and processor, without any specific limitations. When the processor 601 runs the computer program stored in the memory 602, it can execute the aforementioned autonomous vehicle mind reconstruction method.

[0091] The processor 601 may be an integrated circuit chip with signal processing capabilities. In implementation, each step of the above method can be completed by the integrated logic circuitry in the hardware of the processor 601 or by instructions in software form. The processor 601 may be a general-purpose processor, including a Central Processing Unit (CPU), a Network Processor (NP), etc.; it may also be a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor may be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly manifested as execution by a hardware decoding processor, or execution by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory 602, and processor 601 reads the information from memory 602 and, in conjunction with its hardware, completes the steps of the above method.

[0092] Corresponding to the above-described autonomous vehicle mind reconstruction method, this application embodiment also provides a computer-readable storage medium storing machine-executable instructions. When the machine-executable instructions are invoked and executed by a processor, the machine-executable instructions cause the processor to perform the steps of the above-described autonomous vehicle mind reconstruction method.

[0093] The autonomous vehicle mind reconstruction device provided in this application embodiment can be specific hardware on the device or software or firmware installed on the device. The implementation principle and technical effects of the device provided in this application embodiment are the same as those in the foregoing method embodiments. For the sake of brevity, any parts not mentioned in the device embodiment can be referred to the corresponding content in the foregoing method embodiments. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can all be referred to the corresponding processes in the above method embodiments, and will not be repeated here.

[0094] In the embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Additionally, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some communication interfaces; indirect couplings or communication connections between devices or units may be electrical, mechanical, or other forms.

[0095] For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0096] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0097] In addition, the functional units in the embodiments provided in this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0098] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause an electronic device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the unmanned vehicle mind reconstruction method described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0099] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In addition, the terms "first", "second", "third", etc. are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0100] Finally, it should be noted that the above-described embodiments are merely specific implementations of this application, used to illustrate the technical solutions of this application, and not to limit them. The protection scope of this application is not limited thereto. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this application; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application. All should be covered within the protection scope of this application. Therefore, the protection scope of this application should be determined by the protection scope of the claims.

Claims

1. A method for reconstructing the mindset of an autonomous vehicle, characterized in that, include: Acquire multi-source environmental information around the unmanned vehicle, and construct an initial dynamic memory lattice based on the multi-source environmental information; The dynamic memory lattice is dynamically updated according to a preset update strategy, and a multi-dimensional risk assessment is performed on the updated dynamic memory lattice. The lattice structure is then reconstructed based on the results of the multi-dimensional risk assessment. Vehicle control signals are generated based on the reconstructed lattice structure and sent to the execution unit.

2. The method according to claim 1, characterized in that, Constructing an initial dynamic memory lattice based on the multi-source environmental information includes: The multi-source environmental information is spatiotemporally synchronized based on a timestamp alignment mechanism, and standardized environmental perception results are generated through noise reduction and semantic extraction algorithms. The standardized environmental perception results and vehicle historical behavior are abstracted into nodes of the dynamic memory lattice, and feature vectors of the nodes are constructed. The nodes include: environmental nodes representing environmental elements and behavioral nodes representing vehicle behavior. The feature vectors include: spatial location, motion state, temporal characteristics, node type, historical weight, and traffic rule constraint coefficients. Based on the spatial distance threshold and semantic association logic between the nodes, node connection edges are established, and edge weights are determined according to the initial association strength to form the initial dynamic memory lattice structure.

3. The method according to claim 1, characterized in that, The dynamic memory lattice is dynamically updated according to a preset update strategy, including: In response to updates to environmental sensing data or node feature deviations exceeding a preset threshold, the target node to be updated and its set of neighboring nodes in the dynamic memory lattice are determined. Attention weights are calculated based on the spatial distance decay factor and semantic similarity between neighboring nodes and the target node. Then, the feature vectors of the neighboring nodes are transformed into the influence increments on the target node using a feature mapping function. Finally, the feature vectors of the target node are iteratively updated according to the attention weights, the influence increments, and the current feature vectors of the target node. The edge weights of the node connection edges are dynamically adjusted based on the product of the feature deviation and semantic similarity between nodes, so as to realize the real-time evolution of the dynamic memory lattice.

4. The method according to claim 1, characterized in that, A multidimensional risk assessment is performed on the updated dynamic memory lattice, and the lattice structure is reconstructed based on the results of the multidimensional risk assessment, including: Predict future environmental changes based on the updated node feature vectors, and generate multiple candidate paths by combining them with road structure nodes; For each candidate path, the distance risk based on minimum safe distance, the rule risk based on traffic rule conformity, and the historical risk based on the weight of historical behavior nodes are calculated respectively, and the three are weighted and summed to obtain the comprehensive risk value of the candidate path. Based on the comprehensive risk value, delete the behavior nodes corresponding to high-risk paths and add behavior nodes corresponding to low-risk paths, update the risk weight attributes of obstacle nodes, and adjust the edge weights of the node connection edges in low-risk and high-risk paths to achieve the structural reconstruction of the dynamic memory lattice.

5. The method according to claim 1, characterized in that, The vehicle control signals are generated based on the reconstructed lattice structure and sent to the execution unit, including: The low-risk candidate paths and node association weights in the reconstructed lattice structure are extracted and used as constraints or guidance information for the path planning algorithm to generate path parameters containing the desired vehicle speed and desired steering angle. The path parameters are converted into standardized vehicle control signals, wherein the vehicle control signals include: a pulse width modulation signal corresponding to the throttle opening and a controller area network bus command sent to the electric power steering system. The vehicle control signal is sent to the vehicle chassis execution unit to control the unmanned vehicle to drive according to the decision logic represented by the reconstructed lattice structure.

6. The method according to claim 4, characterized in that, After achieving the structural reconstruction of the dynamic memory lattice, the method further includes: The reconstructed lattice structure is stored as long-term memory in the historical memory bank and classified and labeled according to scene type; When the current scene is detected to match the scene type in the historical memory bank, the reconstructed lattice structure corresponding to the matching scene is extracted and used as the initialization parameter of the current dynamic memory lattice.

7. The method according to claim 5, characterized in that, After sending the vehicle control signal to the vehicle chassis actuator, the system further includes: The actual vehicle speed and actual steering angle fed back by the vehicle chassis execution unit are obtained, and the deviation between them and the desired vehicle speed and the desired steering angle is calculated. In response to the deviation exceeding a preset threshold, a secondary update of the dynamic memory lattice is triggered, and the prediction, evaluation, and reconstruction cycle is re-executed; In response to the detection of data loss or update anomalies in the dynamic memory lattice, the system switches to emergency decision mode and uses decision parameters from similar scenarios in the historical memory bank to control the vehicle to slow down and stop.

8. A mind reconstruction device for unmanned vehicles, characterized in that, include: An acquisition and construction unit is used to acquire multi-source environmental information around the unmanned vehicle and construct an initial dynamic memory lattice based on the multi-source environmental information; The update and reconstruction unit is used to dynamically update the dynamic memory lattice according to a preset update strategy, perform a multi-dimensional risk assessment on the updated dynamic memory lattice, and reconstruct the lattice structure according to the multi-dimensional risk assessment results. The generation unit is used to generate vehicle control signals based on the reconstructed lattice structure and send them to the execution unit.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program executable on the processor, characterized in that, When the processor executes the computer program, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the method of any one of claims 1 to 7.

Images