Heavy truck platoon control method and system based on latent communication graph world model

Abstract

The present application relates to the technical field of intelligent networked vehicle formation control, and particularly relates to a heavy truck formation control method and system based on a potential communication graph world model, comprising: acquiring a target measurement of a vehicle, V2V messages, vehicle capabilities and road information, constructing physical perception relationships and V2V communication relationships, and outputting observation features, physical states and target effective markers; determining the timeliness of each edge information, maintaining edge memory in combination with the target effective markers, screening effective edges and aggregating graph contexts; inputting the graph contexts and the actual execution actions of the previous period into the graph world model, and fusing the potential features of the vehicle observation output; generating original actions through a shared policy network; performing safety projection in combination with the physical perception relationships, the physical states and the vehicle capabilities, and outputting safety execution instructions; and collecting actual execution actions after execution, which are used for online and offline updating of the graph world model. The present application enables the formation to still follow safely under communication delay and external cut-in, and avoids action semantic mismatch.

Description

Technical Field

[0001] This invention relates to the field of intelligent connected vehicle platooning control technology, and in particular to a method and system for heavy truck platooning control based on a potential communication graph world model. Background Technology

[0002] Heavy trucks are characterized by their large mass, long braking distance, and limited actuator response. Multi-vehicle platooning utilizes inter-vehicle information exchange to coordinate speed and spacing; however, in real-world mixed traffic, neighboring vehicles cutting in, traffic fluctuations ahead, and communication delays can occur simultaneously. Relying solely on speed error feedback under a fixed topology is insufficient to simultaneously represent changes in both the actual preceding vehicle and communication adjacency relationships.

[0003] Existing methods include: Cooperative Adaptive Cruise Control (CACC) calculates longitudinal commands based on the preceding vehicle's status and the vehicle's measurements; Model Predictive Control (MMCC) optimizes finite-time-domain actions using explicit models; and Reinforcement Learning learns control strategies through interactive data. Each method has its own applicable conditions. When the predictor, communication compensator, policy network, and safety corrector are set up independently, the following problems can easily arise: 1. Confusion between the actual preceding vehicle and the communication lead-in vehicle. After an external vehicle enters the formation, the actual collision risk object changes, but the formation number may remain the same; if safety constraints are still constructed based on the formation lead-in vehicle distance, the entering vehicle may be missed. 2. Outdated communication information is not included in state estimation. The receiver may mistakenly identify historical messages as the current state, or forcibly map messages exceeding the cache duration to the maximum cache index, easily underestimating the true information age. Packet loss is not equivalent to a fixed delay. 3. Inconsistency between generated and executed actions. The original actions output by the policy network change after passing through the safety layer, rate of change constraints, and actuator responses. If the model still pairs the original actions with the actual next state during training, incorrect action-state transition relationships will be formed. Fourth, safety, comfort, and power capability lack clear priorities. Simply using fixed, small spacing, or continuing to apply conventional comfort limitations during emergency braking, cannot adequately cover low-adhesion, high-load, and delayed operating conditions. A single collision-free simulation cannot be equated with safety proof for all operating conditions.

[0004] Therefore, it is necessary to incorporate traffic interaction diagrams of different types, communication information, executable actions, and vehicle physical boundaries into the same control closed loop, clarify the processing objects and inputs / outputs of each stage, so that the technical solution can be implemented and verified according to the given parameters. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a method and system for heavy truck platooning control based on a potential communication graph world model.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a heavy truck platooning control method based on a latent communication graph world model. The method includes the following steps: acquiring vehicle and target measurement data, V2V communication messages, vehicle capability information, and road information; constructing mutually independent physical perception relationships and V2V communication relationships; outputting vehicle observation characteristics, original physical state quantities, and target validity flags; determining the information timeliness of each edge in the physical perception relationship and the V2V communication relationship; maintaining the edge memory state of each edge by combining the information timeliness with the target validity flags; filtering out valid edges; aggregating the edge memory states of the valid edges to obtain a graph context; and combining the graph context with... The actual actions executed in the previous cycle are input into the graph world model, and observation fusion is performed by combining the observed features of the vehicle itself to output latent features. Based on the latent features and the graph context, original actions are generated through a shared cooperative strategy network. Combining the physical perception relationship, the original physical state variables, and the vehicle capability information, safety constraint projection is performed on the original actions to output safety execution instructions. The safety execution instructions are executed, and actual actions are collected. These actual actions are used for online updates of the graph world model. After storage, the actual actions are paired with the observed features of the vehicle collected in the next cycle for offline updates of the graph world model and the shared cooperative strategy network. This invention achieves safe cooperative control of heavy truck platoons under communication delays and traffic disturbances by separating physical perception and V2V communication relationships, utilizing information timeliness to maintain edge memory, and combining the graph world model with safety projection.

[0007] Optionally, the step of acquiring vehicle and target measurement data, V2V communication messages, vehicle capability information, and road information, constructing independent physical perception relationships and V2V communication relationships, and outputting vehicle observation features, raw physical state quantities, and target validity indicators includes: performing five-branch encoding and normalized aggregation based on the vehicle and target measurement data and the road information to output the vehicle observation features; outputting the raw physical state quantities based on the vehicle and target measurement data, and simultaneously generating the target validity indicator; constructing the physical perception relationship based on the vehicle and target measurement data, wherein the physical perception relationship includes physical perception edges, each corresponding to the motion state of the perceived target, and the validity of the edges is determined by the target validity indicator; and constructing the V2V communication relationship based on the V2V communication messages, wherein the V2V communication relationship includes V2V communication edges, each corresponding to the semantic interaction state of vehicles within the platoon. This invention improves the platoon's ability to perceive and express relationships in complex traffic environments through five-branch encoding and normalized aggregation, as well as the independent construction of physical perception and V2V communication relationships.

[0008] Optionally, determining the information timeliness of each side in the physical sensing relationship and the V2V communication relationship includes: determining the information timeliness of the V2V communication side based on the difference between the generation time of the V2V communication message and the current time, and generating a communication information age; determining the information timeliness of the physical sensing side based on the difference between the sampling time of the vehicle and target measurement data and the current time, and generating a sensing information age; the communication information age and the sensing information age are calculated separately and used independently. This invention, by calculating the communication information age and the sensing information age separately and using them independently, accurately reflects the timeliness of communication and sensing data, and improves the accuracy of state estimation.

[0009] Optionally, the step of maintaining the edge memory state of each edge by combining the information timeliness and the target validity flag, filtering out valid edges, and aggregating the edge memory states of the valid edges to obtain the graph context includes: maintaining the recursive edge memory state of the V2V communication edge by combining the communication information age, and maintaining the recursive edge memory state of the physical perception edge by combining the perception information age; determining the validity of the V2V communication edge according to the communication information age, and when the communication information age exceeds a preset failure threshold or the model version is incompatible, determining the V2V communication edge as an invalid edge and removing it from the aggregation range; determining the validity of the physical perception edge according to the target validity flag, and when the target validity flag indicates that the corresponding perception target is invalid, determining the physical perception edge as an invalid edge; obtaining the valid edges after removing the V2V communication edges and physical perception edges determined to be invalid, summing and normalizing the edge memory states of all valid edges, and obtaining the graph context through aggregation transformation; when the V2V communication edge fails, the physical perception edge still independently participates in the aggregation calculation. This invention enhances robustness by separately maintaining the recursive edge memory of the V2V communication edge and the physical sensing edge, and independently determining their validity, ensuring that the physical sensing edge still participates in aggregation when communication fails.

[0010] Optionally, the step of inputting the graph context and the actual actions performed in the previous cycle into the graph world model, and combining the observation features of the current vehicle to perform observation fusion and output latent features includes: inputting the graph context, the actual actions performed in the previous cycle, and the deterministic and random states of the previous cycle into the graph world model to recursively obtain the deterministic state of the current cycle; obtaining the prior distribution of the random state based on the deterministic state of the current cycle; fusing the observation features of the current vehicle with the deterministic state of the current cycle to obtain the posterior distribution of the random state, wherein the prior distribution is only used when there is no imagined development of future real observations; and using the random state corresponding to the posterior distribution, concatenating the deterministic state of the current cycle and the random state corresponding to the posterior distribution to construct the latent features. This invention employs a hybrid representation of deterministic and random states, combining prior / posterior distributions, and using the posterior distribution online to improve the fusion capability of latent features for current observations and the accuracy of state estimation.

[0011] Optionally, generating the original action through a shared cooperative strategy network based on the latent features and the graph context includes: inputting the latent features, the graph context, the role vector, and the local physics vector into the shared cooperative strategy network; the shared cooperative strategy network outputting the mean and standard deviation of the original action; generating action variables based on the mean and standard deviation; and mapping the action variables to an asymmetric action range between the upper and lower limits of normal operating acceleration to obtain the original action. This invention outputs the mean and standard deviation through a shared strategy network and maps them to an asymmetric action range, enabling the strategy to cover the asymmetric acceleration range of the vehicle and improving the adaptability of action generation.

[0012] Optionally, the step of combining the physical perception relationship, the original physical state quantity, and the vehicle capability information to project the original action with safety constraints and output a safe execution command includes: calculating parking reserve and time interval safety margin based on the physical perception relationship, the original physical state quantity, and the vehicle capability information; calculating the upper bound of actual acceleration based on the time interval safety margin and kinematic recursion; transforming the upper bound of actual acceleration into an upper bound of command based on a first-order discrete response relationship, where the first-order discrete response relationship characterizes the actuator dynamic response with an equivalent time constant; determining a feasible interval by combining the upper bound of command and the parking reserve; projecting the original action onto the feasible interval; and outputting the safe execution command. This invention, by calculating the parking reserve, time interval safety margin, upper bound of actual acceleration, and upper bound of command, determines a feasible interval and projects the original action, ensuring that the safe execution command meets vehicle dynamic constraints.

[0013] Optionally, the execution of the safety execution command and the collection of actual execution actions; the actual execution actions are used for online updates of the graph world model, including: verifying the validity of the safety execution command and executing it; collecting and storing the actual execution actions of the current period; using the actual execution actions of the current period as state transitions to input the graph world model of the next period for online updates; online inference only updates the hidden states and the edge memory states, without modifying the weights in real time. This invention ensures consistency between the model and the actual vehicle response by verifying the validity of the safety command and executing it, collecting actual execution actions, and inputting them into the graph world model of the next period for online updates.

[0014] Optionally, the actual executed action, after being stored, is paired with the vehicle's observation features collected in the next cycle for offline updating of the graph world model and the shared cooperative policy network. This includes: pairing the actual executed action with the vehicle's observation features collected in the next cycle to construct a multi-vehicle transfer record; using the multi-vehicle transfer record, with the actual executed action and the vehicle's observation features collected in the next cycle as real state transition samples, offline updating of the graph world model; using the multi-vehicle transfer record, starting from the random state corresponding to the posterior distribution, unfolding an imagined trajectory in the graph world model, and obtaining a predicted actual executed action after processing by the safety constraint projection and actuator response model at each step, using the predicted actual executed action to advance the next state transition, and offline updating of the shared cooperative policy network based on the reward of the imagined trajectory. This invention achieves safe and efficient offline training of the policy network by constructing multi-vehicle transfer records for offline updating of the graph world model and advancing the state in the imagined trajectory through safety projection and actuator response model.

[0015] Secondly, this invention provides a heavy truck platooning control system based on a latent communication graph world model. The system executes the heavy truck platooning control method based on the latent communication graph world model provided by this invention. The system includes a state acquisition and graph construction module, a semantic communication and edge memory module, a graph world model module, a shared cooperative strategy module, a physical safety projection module, a vehicle execution and feedback module, and an offline training module. The platooning control system provided by this invention includes state acquisition, semantic communication, graph world model, shared strategy, safety projection, execution feedback, and offline training modules, with a complete structure and improved overall applicability. Attached Figure Description

[0016] Figure 1 This is a flowchart of a heavy truck platooning control method based on a potential communication graph world model, according to an embodiment of the present invention.

[0017] Figure 2 This is a schematic diagram of a heavy truck platooning application scenario under three-lane mixed traffic conditions, according to an embodiment of the present invention.

[0018] Figure 3 This is a framework diagram of a heavy truck platooning control system based on a potential communication graph world model, according to an embodiment of the present invention. Detailed Implementation

[0019] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.

[0020] Throughout this specification, references to "an embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, the phrases "in an embodiment," "in an embodiment," "an example," or "an example" appearing in various places throughout the specification do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in one or more embodiments or examples in any suitable combination and / or sub-combination. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.

[0021] Please see Figure 1 One embodiment of the present invention provides a heavy truck platooning control method based on a potential communication graph world model, the method comprising the following steps: S1. Acquire the measurement data of the vehicle and the target, V2V communication messages, vehicle capability information and road information, construct mutually independent physical perception relationships and V2V communication relationships, and output the vehicle's observation characteristics, raw physical state quantities and target validity indicators.

[0022] Specifically, such as Figure 2 As shown, taking three heavy trucks on a one-way three-lane mixed traffic road as an example: the lead vehicle 201, the first following vehicle 202, and the second following vehicle 203 form a controlled platoon. A manually driven vehicle 204 and other intelligent connected vehicles 205 constitute the surrounding mixed traffic environment. The intelligent roadside decision-making system 206 provides information such as road geometry, speed limits, and traffic statistics. In the diagram, ICV represents intelligent connected vehicles, and HDV represents manually driven vehicles. The vehicles are traveling in the right direction, and the red rectangle indicates the controlled platoon. Vehicles only change lanes when they comply with traffic rules and meet lane-changing safety conditions. The vehicles, platoon range, and inter-vehicle distances shown are for illustrative purposes only and are not drawn to scale.

[0023] The orange broken line represents the vehicle-to-vehicle (V2V) communication relationship between intelligent connected vehicles. Vehicles in a platoon exchange vehicle status and semantic messages required for collaborative decision-making through this link. The blue solid line represents the direct vehicle-to-infrastructure (V2I) communication link between the intelligent roadside decision-making system 206 and intelligent connected vehicles, used for vehicle information reporting and roadside information dissemination. The blue dashed line represents the indirect information transmission from the intelligent roadside decision-making system 206 to the manually driven vehicle 204, transmitting traffic information through roadside distribution equipment for driver decision-making; this does not represent the establishment of a direct communication link. The yellow curved arrow indicates the trajectory of the manually driven vehicle 204 that may enter the platoon. When 204 enters between the lead vehicle 201 and the first following vehicle 202, the actual preceding vehicle of 202 changes to 204, but its platoon communication lead vehicle remains 201. 202 performs safety calculations based on information such as the relative distance and relative speed of the preceding vehicle obtained from onboard sensors, and maintains communication memory according to the vehicle number of 201, updating the actual preceding vehicle perception relationship and the platoon communication relationship accordingly.

[0024] This invention is applicable to structured roads with known lane topology, obtainable vehicle positioning and forward target measurement, and vehicles equipped with longitudinal drive and braking interfaces. The intelligent roadside decision-making system does not directly control manually driven vehicles; roadside equipment provides optional information and offline training services. Curves, ramp cut-ins, start-stop, and lane congestion are perturbation conditions for the same algorithm. This invention covers car following, start-stop, curve speed limit pre-adjustment, cut-in response, traffic wave absorption, and formation recovery.

[0025] Specifically, each truck is equipped with a positioning device, an inertial measurement unit, a forward-facing radar or equivalent target sensing device, an onboard network interface, V2X communication equipment, a computing unit, and a longitudinal execution controller. Roadside equipment can provide road curvature, speed limits, and traffic conditions; when there is no roadside input, onboard maps and the vehicle's own perception are used, and missing information is expressed through validity indicators, without using zero speed to represent unknown traffic conditions.

[0026] In this embodiment, a centralized training and vehicle-side distributed inference deployment approach is adopted: each vehicle runs the same network parameters and retains its own hidden state and incoming edge memory; during the training phase, the states of the six vehicles can be combined into a batch tensor in parallel, but it is not required that the online vehicle obtains the global true state that has not yet been received. Optionally, in centralized deployment, the formation coordination computing unit outputs instructions to each vehicle, and each vehicle still independently performs the security verification of the physical security projection stage; the communication and computing latency of centralized deployment is included in the response time budget.

[0027] The longitudinal coordinates of the road are used, increasing in the direction along the centerline of the reference road, with the position reference being the projection point of the vehicle's front bumper; the distance on the curve is the arc length along the lane reference line. Targets crossing lanes are first projected onto the common road reference line, while retaining their lateral position and lane markings; forward spacing constraints are only applied when path occupancy conflicts exist. The discrete periodicity is numbered as follows: The period length is The actual time is The convoy of controlled vehicles is assembled as follows: 0 indicates the lead car; To receive or be controlled vehicles, For the vehicle that is the source of the information, The actual preceding vehicle; the actual preceding vehicle can be an externally driven vehicle, and does not necessarily have to be equal to... The prediction step number is... The feature component is numbered as The training iteration number is relation type Take physical sensing or V2V communication.

[0028] To define the longitudinal interaction state between vehicles, the net clearance between the front and rear bumpers and the non-negative approach speed are calculated using the following formula:

[0029] In the formula, The longitudinal position of the front bumper is shown in meters (m). The length of the vehicle is in meters (m). The net distance between the front and rear bumpers, in meters (m). Longitudinal velocity, unit: m / s; The approach speed is non-negative. When there is no preceding vehicle, the target missing flag is recorded and the corresponding losses and constraints are masked; infinity is not input into the network.

[0030] To avoid mixing actions, define The original target acceleration generated by the strategy, Accelerate the target issued by the security layer. The actual average longitudinal acceleration achieved during this period, in m / s². 2 The cap indicates the predicted value, the horizontal line indicates the normalized value, and the asterisk indicates the target value; the superscript is used to distinguish semantics when it is not a power, and all squares are represented by explicit powers of two.

[0031] In this embodiment, multi-source data acquisition and time reference preprocessing are first performed: speed, acceleration, driving and braking status are read from the vehicle CAN or equivalent vehicle network; the positions, relative speeds, lane and target confidence of vehicles ahead and around are obtained from radar and visual target tracking results; and road geometry and traffic information ahead are obtained from maps or roadside messages. The time reference of all messages is unified, and short-time measurements are extrapolated to the control time using kinematics, while an extrapolation error bound is added. Outdated messages are not included in hard safety calculations.

[0032] Secondly, five observation branches are constructed: the movement history of the current vehicle and the actual preceding vehicle, the geometry of the road ahead, the downstream traffic conditions, the forward target in the same lane, and the surrounding targets in adjacent lanes. The current surrounding branch does not read the cut-in probability of the current cycle that has not yet been calculated, to avoid using the predicted output repeatedly as the input for the same step. This embodiment retains 5 historical cycles, 20 road sampling points within 200m, 8 traffic statistics segments, and a maximum of 12 surrounding targets. Surrounding targets are selected in order of predicted conflict time and distance; the sorting is only used for truncation, and the aggregation is not sensitive to the arrangement of the retained targets.

[0033] Table 1 provides a set of fixed input shapes that can be directly implemented. When the target is insufficient for the capacity, zeros are filled in, and the mask is 0; when the target exceeds the capacity, the target is selected based on the conflict priority with the predicted driving path and the distance. All physical quantities are first normalized using the following normalization formula; the mask and discrete indicators are not scaled. Here, MLP represents a multilayer perceptron, the arrows represent the layer widths, the hidden layers use the SiLU activation function, and the final projection layer is a linear layer.

[0034] Table 1. Five-branch input and coding network

[0035] The relevant parameters and rules for the encoding process are as follows: torque scale is 2000 Nm, vehicle length scale is 20 m, lateral offset scale is 3.5 m; position, velocity, and acceleration scales can be 200 m, 30 m / s, and 6 m / s², respectively. 2 When turn signals are unavailable, the corresponding mask is set to 0. TTC is calculated by dividing the net distance by the positive approach speed, and is only effective when there is a target, the net distance is non-negative, and the approach speed is greater than 0.1 m / s; otherwise, the mask is 0 to avoid using infinite values. Density and traffic wave velocity are obtained from effective roadside statistics or calibrated vehicle-side statistics; unknown values ​​are not treated as true zero values. Discrete categories use one-hot codes; missing data is set to zero and a mask is added.

[0036] To transform the original physical quantities into normalized features within the network input range, the following normalization formula is used:

[0037] In the formula, For a certain original feature; This is a fixed bias in the same unit; the bias can be 0. Positive scales of the same unit; The dimensionless truncation threshold is... The curvature scale can be taken as 0.01m. -1 The time scale is 1 second, and the traffic density scale is 0.1 veh / m.

[0038] It should be noted that normalization is only used for network input; the security layer always uses uncrunted physical quantities.

[0039] To obtain fixed-dimensional aggregated features independent of the number of targets even when the target capacity is insufficient, the following masked aggregation formula is used:

[0040] In the formula, The normalized observation input vector constructed by the receiving vehicle for the j-th adjacent lane target in the current period includes the 8 target features listed in Table 1 and their 8 component effective masks, totaling 16 dimensions; missing features are set to zero and marked as invalid by the corresponding mask, and the receiving vehicle and period subscripts are omitted in the formula. For the aggregation features of surrounding target branches, Concat indicates concatenation along the feature dimension, and MaskedMax indicates taking the maximum value for each component only for valid targets. For filling capacity, As a valid indicator of the target, To share two layers of a multilayer perceptron. When all targets are invalid, both the mean and the maximum value are defined as zero vectors.

[0041] After completing the encoding and aggregation of each branch, each branch is ultimately projected to 64 dimensions, for a total of 320 dimensions across the five branches; role vectors The system is 16-dimensional, with the first four digits encoding the navigator, first follower, middle follower, and last car, respectively, and the remaining positions padded with zeros; the unique following car in a two-car convoy is coded as the last car. After concatenation, the total dimension is 336. After linear mapping, layer normalization, and SiLU activation, 256-dimensional observation features of the vehicle itself are obtained. .

[0042] Construct physical perception relationships respectively V2V communication relationship Matrix elements The information is indicated by Flow direction The internal tensor size of the formation is External vehicles are recorded as physical targets and included in the observation branch and safety constraints, without fabricating their V2V semantics. If the actual preceding vehicle is an external vehicle, the corresponding internal physical preceding edge is set to zero and the external target number is saved; its motion state is still continuously processed, and the physical perception edge will not be deleted due to V2V packet loss.

[0043] The final output includes normalized embeddings. Original physical state, target valid markers, two types of adjacency relationships, and road preview.

[0044] S2. Determine the information timeliness of each edge in the physical perception relationship and the V2V communication relationship, maintain the edge memory state of each edge by combining the information timeliness and the target validity flag, filter out valid edges, and aggregate the edge memory states of the valid edges to obtain the graph context.

[0045] In this embodiment, to eliminate circular dependencies in same-period communication, during the period... At the end of the cycle, the features filtered in this cycle are encoded into a message for use in subsequent cycles; cycle The model within the system only reads messages that have arrived before the deadline for receiving messages in the current cycle.

[0046] To compress latent features into low-dimensional quantized semantics suitable for V2V link transmission, the following encoding formula is adopted:

[0047] To dequantize the received quantized semantics, the following dequantization formula is used:

[0048] In the formula, The 16-dimensional continuous semantic vector output by the hyperbolic tangent function tanh for vehicle i at period t has not yet been quantized; in equations (4) and (5) This is a shorthand for any component of the vector. This is the reconstructed value of the component after quantization encoding, transmission, and dequantization; round indicates rounding to the nearest integer, and the error bound of equation (5) holds true for each component. The potential feature of this vehicle is defined at the aforementioned potential feature location; For offline learning, the weights and biases of the linear layer; Output unsigned integers from 0 to 255 component by component; To solve the quantization semantics. Both quantization and semantic quantities are dimensionless. Training uses a pass-through gradient approximation, performing true quantization on the forward pass and approximating the gradient to 1 in the unsaturated region on the backward pass; this training approximation does not imply that the floor function is differentiable everywhere.

[0049] The message includes the action intent, which is a normalized encoding of the issued safety command, using a 16-bit unsigned integer. The encoding range is determined by the vehicle's configuration, from the lower limit of emergency braking to the upper limit of positive acceleration, and the decoded unit is m / s². 2 This intention is not considered a guarantee that the other party will necessarily perform. (Encoding lower bound) m / s 2 Upper limit m / s 2 The value sent is The received value is The upper and lower limits are fixed by the compatible model configuration convention.

[0050] The 16-dimensional 8-bit semantics occupies 16 bytes, and the action intent occupies 2 bytes, totaling 18 bytes, which constitutes the semantic and action payload. The message should also include the vehicle number, generation timestamp, sequence number, model version, and verification information. The corresponding payload rate at the 10Hz control frequency is 1440bit / s. Actual link usage should be calculated separately for protocol overhead, retransmissions, and multi-neighbor transmissions.

[0051] To determine the timeliness of information on each side, the true information age is calculated based on the message generation time, and new messages are marked using the following formula:

[0052] In the formula, The indicator function is set to 1 if a new message from vehicle j is received before the reception deadline, and its generation time is later, its sequence number is updated, and it passes the validity check; otherwise, it is set to 0. `max` is the maximum value function used to ensure the information age is non-negative; `t` is the current discrete cycle number. The period length is the product of the two values, which represents the current control time in seconds. The time of generation of the message currently held by the receiving end, in seconds; Age is real information, in seconds; This is a new message flag. When receiving out-of-order messages, only messages generated later and with a valid sequence will be accepted. The upper bound of the time synchronization error should include the security layer response time; message arrival time should not be used to replace generation time.

[0053] To maintain the edge memory state of each edge by combining information timeliness with the validity of the target, a gated cyclic unit is used to update the incoming edge memory according to the following formula:

[0054] In the formula, The 16-dimensional semantic vector is obtained by component-wise dequantization of the vehicle j message currently held by the receiving end using Equation (5). Its time corresponds to the generation time of the message held, rather than the default corresponding to the current period. For relation type indexing, retrieve physically-aware relations or V2V communication relations. This indicates a gated recurrent unit specific to this type, where each type learns and stores network weights separately; For edge memories of vehicles from the same source, receiving vehicles, and relation types in the previous cycle, equation (7) omits the relation type superscript in edge memories, and equation (10) uses... Explicitly indicate the superscript; For the receiving end, the recursive memory stored for this incoming edge. Represents a 128-dimensional real vector space; Divide the decoding intent by 6m / s 2 The dimensionless quantity that follows; s; These are one-hot codes for two types of relationships. The GRU (Gated Recurrent Unit) uses Sigmoid activation for its update and reset gates, and tanh activation for candidate states. New edge memories are initialized to zero; they are continuously retained when the same vehicle number and model version remain unchanged. Physically perceived relationships can project the current target features provided by the aforementioned state acquisition and graph construction stages into 16 dimensions as relationship input, without being disguised as communication messages.

[0055] To make equation (7) reproducible, let the concatenated input be... The previous period's edge memory state was Their lengths are 21 and 128 respectively. The total dimensions are 21: 16-dimensional semantics, 1-dimensional intent, 1-dimensional information age, 1-dimensional update flag, and 2-dimensional relation code. The following gating structure is used, with each relation type holding a set of weights:

[0056]

[0057] In the formula, The dimensions are 128×21. It is 128×128, with a bias of 128 dimensions; For the Sigmoid function, This is element-wise multiplication. All hidden states and gate values ​​are dimensionless. To reset the door, To update the door.

[0058] Accordingly, the semantics of the physical edge is obtained by linearly projecting the 64-dimensional code obtained from the aforementioned observation encoding of the current effective target to 16-dimensional code and then applying tanh. The intended position is filled with the normalized amount of the measured acceleration of the target. The physical edge information age is generated by the sensor and stored separately from the communication edge.

[0059] When no new messages are received, the memory is advanced using the last valid content, the incremented information age, and the zero update flag. This embodiment uses... `s` serves as the semantic edge failure threshold; if the threshold is exceeded or the model version is inconsistent, the communication edge is invalidated. Normalization or truncation of network input is permitted, but the actual information age used for failure determination must never be truncated. Edge memory is reset or re-filtered upon receiving a valid message again.

[0060] To aggregate the edge memory states of valid edges to obtain the graph context, the following aggregation formula is used:

[0061] In the formula, For vehicle i, define a 128-dimensional graph context in period t; j iterates through the platoon of vehicles that provide information to vehicle i, and r iterates through the physical perception and V2V communication relationships. To complete the adjacency flag after edge validity screening, a valid incoming edge is set to 1, otherwise it is set to 0; The 128-dimensional recursive memory for this type of incoming edge corresponds to Equation (7); the numerator is the sum of the memory vectors of the effective incoming edges, and the denominator is the number of effective incoming edges with a minimum value of 1; For a 128-dimensional aggregated network, the summation only covers valid incoming edges. When there are no valid incoming edges, the input is zero and a communication failure state is attached. The model can still utilize physical observations of the current vehicle and the actual preceding vehicle. Edge memory is used to improve prediction, but it cannot replace the measurements and error bounds required for braking safety boundaries.

[0062] S3. Input the graph context and the actual execution action of the previous cycle into the graph world model, combine the observation features of the vehicle to perform observation fusion, and output the potential features.

[0063] In this embodiment, each vehicle retains a 128-dimensional deterministic state. and a random state consisting of 16 independent 16-category distributions The random state sample is expanded into a 256-dimensional one-hot vector. The latent features are defined as: The latent features, concatenated with the graph context, result in a 512-dimensional array. This includes deterministic state symbols. Safety margin as described later Use different symbols to avoid confusion.

[0064] To advance a deterministic state under the action of an action, the following state transition formula is used:

[0065] In the formula, It is a 128-dimensional GRU state transition network; The world model parameter set; the actual action input is a normalized quantity. ; This is the 128-dimensional deterministic state saved from the previous period. The 256-dimensional random state is saved after filtering the observations from the previous cycle: During online inference, the maximum probability class is selected group by group from the posterior distribution of the previous cycle in equation (12) and encoded as a one-hot vector; during training, it is sampled from the posterior distribution and estimated using the direct gradient method. The starting point of the offline imagined trajectory uses the posterior state of the replay sequence, and subsequent steps without real observations use the prior distribution to generate random states. Initialization is set as follows: The zero vector is generated using the posterior distribution of the initial valid observations through equation (12). ; Starting from the next cycle, recursively follow equation (11) to avoid calling the non-existent random state of the previous cycle.

[0066] To combine the observation characteristics of this vehicle for observation fusion, the posterior and prior distributions are defined as follows:

[0067] In the formula, To integrate the posterior distribution of current observations, For prior distributions without current observation conditions, each classification distribution is treated with Softmax to make the sum of probabilities equal to 1. Online observation filtering also uses the posterior; the prior is only used when there is no imagination to unfold in the absence of future real observations.

[0068] Furthermore, the deterministic GRU input consists of the previous random state (256 dimensions), the actual action (1 dimension), and the graph context (128 dimensions), totaling 385 dimensions, with the hidden state being 128 dimensions; its gating form is the same as equations (8) and (9), only the matrix size is adjusted. The prior MLP structure is 128→256→256, and the posterior MLP structure is 384→256→256, the latter's input consisting of 128-dimensional deterministic states. and 256 visor observation features The outputs are concatenated. The 256 outputs are rearranged into 16×16 groups and Softmax is performed on each group. During the training phase, classification sampling and direct estimation are used. During the online inference phase, the class with the highest probability can be fixed to ensure that the same input can be reproduced.

[0069] make All eight student heads use an MLP structure with a 512→256→ output width and SiLU hidden layers. Each group can have multiple terminal branches; all student heads receive input from the same world model and do not run the complete teacher model separately.

[0070] Table 2 Student Head Output, Supervision Sources, and Output Destinations

[0071] In this context, the displacements of D2 and D5 are relative to the current longitudinal position of the target, and the coordinates of D6 are the longitudinal and lateral displacements relative to the current reference coordinates of the vehicle, all measured in meters (m). When the identity of the preceding vehicle changes, the target prediction cache is reinitialized and cross-identity labels are masked. Three lane intentions are held, left lane change, and right lane change, with labels determined based on the direction of the first crossing of the lane boundary within the next 30 steps. The cut-in label is set to 1 when the target's occupied area enters the vehicle's predicted passage corridor within the time domain, otherwise it is set to 0. Future segments that are lost to tracking are invalidated, and unknown future segments are not marked as not cut-in. Velocity output uses Softplus multiplied by the velocity scale, probability output uses Sigmoid or categorical Softmax activation, and other continuous quantities are output linearly and inversely normalized according to the corresponding scales in Table 2. Predicted acceleration is limited to the calibrated target capability range, and position and velocity predictions are supplemented with kinematic consistency loss.

[0072] To ensure the predicted trajectory satisfies kinematic consistency, the following kinematic consistency loss formula is used:

[0073] In the formula, , and These represent the predicted longitudinal position, predicted longitudinal velocity, and predicted longitudinal acceleration of the same target at the k-th prediction step, respectively, with units of m, m / s, and m / s². 2 The student head output is predicted from the corresponding trajectory in Table 2 and then inversely normalized; for student heads using relative displacement, The displacement relative to the current longitudinal position of the target is taken, and the same origin is used for the entire prediction sequence. , This is the corresponding predicted value for the next prediction step. The time interval between adjacent prediction steps; Mean represents the arithmetic mean of the effective target and the effective prediction steps; and These are the normalized position and velocity consistency residuals, respectively. m、 m / s; the mean is only valid for the target and is not calculated across prediction steps of the parking event. Piecewise kinematics supervision is used near the parking area. All terms are dimensionless; the prediction results are used for performance decision-making, and hard safety constraints still come from the current measurement and conservative error bounds.

[0074] S4. Based on the latent features and the graph context, generate the original action through a shared cooperative strategy network.

[0075] In this embodiment, the Actor network input is shared. , The network consists of a character vector and an 8-dimensional local physical vector, totaling 536 dimensions. The physical vector includes the vehicle's speed, actual acceleration, net distance, relative speed to the preceding vehicle, target speed, target distance, and two-bit flags indicating whether the preceding vehicle is valid and communication is valid. The first six items are normalized according to a predetermined scale. The network structure is 536→256→256→2, and the output mean is... and standard deviation The hidden layer uses SiLU. The policy parameters are defined, and their standard deviations, after being mapped from positive values, are limited to between 0.05 and 1.0; both are dimensionless. The training samples are standard normal noise. Take during deployment .

[0076] To generate the original motion that covers the asymmetric motion range, the following formula is used:

[0077] In the formula, and These represent the mean and standard deviation of the Gaussian action variables output by the shared policy network before the tanh transformation. This is the set of policy network parameters; the standard deviation is limited to between 0.05 and 1.0 after positive value mapping. The noise is a standard normal distribution, sampled during training and set to 0 during deployment; The mean, standard deviation, and noise are dimensionless action variables that are between -1 and 1 after tanh transformation. and These represent the midpoint and half span of the normal acceleration range, respectively, both in m / s². 2 ; and These are the upper and lower limits of normal operating acceleration, in m / s². 2 The midpoint and half-span are calculated using the following formula:

[0078] For example, when taking 1.2 and -3.0, the midpoint is -0.9 and the half span is 2.1. The tanh output cannot be directly considered as having m / s 2 The unit's complete range of motion; the affine mapping described above enables the strategy to cover asymmetric action ranges.

[0079] S5. Combining the physical perception relationship, the original physical state quantity, and the vehicle capability information, perform a safety constraint projection on the original action and output a safety execution command.

[0080] In this embodiment, the achievable capability is obtained based on the current load, gradient, adhesion estimation, braking status, and vehicle calibration table. This embodiment limits the normal acceleration to -3.0 to 1.2 m / s². 2 The maximum emergency braking speed is 6.0 m / s. 2 This value applies only when the vehicle type and road conditions meet the requirements. It guarantees net deceleration. Conservative calibration values ​​should be used, and adhesion and mechanical capability limitations should be met. If braking capability is unknown or fails, close formation should not be maintained.

[0081] In normal small-step prediction, the relationship between the issued command and the periodic average acceleration is calibrated as a first-order discrete model:

[0082] In the formula, The equivalent time constant identified by the periodic average acceleration sequence is in seconds (s), and in this embodiment, it is taken as 0.2 s. Dimensionless. This equivalent discrete model is not directly equivalent to the endpoint acceleration of a continuous actuator; this model is used for prediction in normal operating ranges, while emergency braking uses independently calibrated settling times and guaranteed decelerations, without assuming the braking system can instantaneously reach -6.0 m / s. 2 The model error bound is estimated from the validation data and incorporated into the safety margin; it cannot be claimed that the actual vehicle necessarily meets the constraints based solely on this approximate model.

[0083] To predict the position and velocity evolution within the period, the following kinematic recursive formula is used in the interval where the acceleration is approximately constant and the vehicle does not stop:

[0084] If the predicted speed drops to zero, first determine the stopping time, then maintain stillness for the remaining time; it is prohibited to cut off the speed while continuing to integrate the position using negative acceleration over an entire period. Road speed limits are satisfied through advance braking constraints, and physical braking cannot be replaced by numerically cutting off the speed.

[0085] Simultaneously, using the current position of the vehicle in front as a conservative stationary obstacle position, a parking reserve is constructed that does not rely on its future coordination. This is the upper bound of the total response time from the latest valid measurement to ensuring brake establishment, including sensing, calculation, necessary communication, time synchronization error, and brake establishment time. This is the upper bound of the positive acceleration that may continue to occur within this time period; the parking reserve is calculated using the following formula:

[0086] In the formula, This is the net parking space. This represents the combined margin of distance and model uncertainty, both in meters. To ensure net deceleration, the unit is m / s 2 All units are in meters (m). This formula assumes that the obstacle does not move backward longitudinally along the road and that the actual deceleration reaches at least [a certain value] after the response time. And remain stationary after stopping. For targets that can move backward or slide down a ramp, the lower bound of the target motion needs to be reconstructed.

[0087] To determine the target net spacing under the performance objective, the following formula is used:

[0088] In the formula, The nominal formation clearance is 12m in the example. The time interval parameter is set to 1.0s in this embodiment. This formula is used for performance targets, with parking reserve as a separate physical verification condition. A fixed 12m is only applicable when the parking reserve is no greater than 12m at the time, and cannot be uniformly applied across all speeds and communication states. When a cut-in is detected, the actual preceding vehicle and constraint set are updated; a reduced spacing is not used to prevent a cut-in that has already begun.

[0089] Define a safety margin for the time interval during a prediction period that maintains the same actual preceding vehicle identity:

[0090] The margin is required to meet the requirements of the next period. ,in The coefficient of contraction is a dimensionless coefficient, which is taken as 0.2 in this embodiment. To accommodate the distance margin for one-step prediction error, this embodiment uses 0.1m. For targets that will not stop within the cycle, substituting into equation (17) yields:

[0091] In the formula, This is a conservative lower bound for the longitudinal acceleration of the actual preceding vehicle during this period, in m / s². 2 The maximum braking amplitude can be negative; however, the target measurement error must be taken into account when using this lower bound. Numerator unit: m; denominator unit: s. 2 Therefore, the unit of the result is m / s 2 If the preceding vehicle stops within the cycle, the lower bound of its next position is directly calculated using the segmented trajectory of the stopping vehicle, replacing the preceding vehicle displacement term in the numerator, and the negative velocity is no longer used as the true velocity.

[0092] Furthermore, the upper bound of the actual acceleration is transformed into an upper bound of the commands that the actuator can receive, using the following formula:

[0093] In the formula, Division is defined. If the actuator calibration does not support equation (16), the instruction boundary is determined by a lookup table of the verified monotonic response instead of ignoring hysteresis.

[0094] It should be noted that Equation (21) is only a local discrete safety condition. Only when the initial state satisfies the constraints, the disturbance and measurement error are bounded, the solution is continuously feasible and the execution error is bounded, can the preservation of the corresponding safety set be discussed. When the target enters and changes the identity of the preceding vehicle, the constraints need to be re-established, and the conclusion of the invariance of the original preceding vehicle set cannot be applied.

[0095] Based on this, the feasible set is composed of the command amplitude, the speed limit, equation (22), the actual rate of change of acceleration, and the parking reserve; the normal rate of change limit applies to the expected actual acceleration, rather than only to the original strategy command, and adopts the following formula:

[0096] In the command interval, equation (23) corresponds to Two boundaries. Intersecting these boundaries with the normal amplitude range and all safe upper bounds, we obtain... For each target that may intrude into the vehicle's path, a boundary is calculated, and the most stringent one is selected.

[0097] The parking reserve check further verifies that the net spacing between predicted next states is not less than the next state value in equation (18), and checks that the minimum spacing within a cycle is not less than... A segmented constant acceleration trajectory is used, and the minimum distance is calculated for the interval endpoints, the vehicle's stopping time, and the time when the relative velocity is zero. The lower bound of the trajectory after the error limit is verified. If the constraint is monotonic with respect to the command, an additional upper bound can be found using the bisection method until the command interval width is less than 0.01 m / s². 2 This accuracy is the termination condition for the solution; a margin of error must be retained for safety assessment.

[0098] To obtain safely executable instructions within the feasible set, Minimum is the objective, where m / s 2 The following projection formula is obtained:

[0099] In the formula, The feasible command interval is defined as follows: Equation (24) is used when the condition is normal and feasible. The derivative of the interval with respect to the original action is 1, the saturation interval is 0, and the switching point uses the subgradient or stopping gradient. The optional smooth minimum value is only used for training approximation, and the original hard constraints are still verified online in the final step. The braking upper bound cannot be applied first and then the comfort cutoff that would increase acceleration cannot be used to break it.

[0100] When the normal formation is empty, first remove the normal comfort rate of change condition and expand to the calibrated emergency capability range, then re-check safety feasibility; emergency execution still adheres to the mechanical braking build-up rate and adhesion capability. If the emergency formation is also empty, output the maximum available braking request, cancel the continued formation tightening and lane change permission, record the infeasibility reason, and hand it over to the existing minimum risk control function for handling. At this time, the goal is to mitigate risk, without claiming to be able to avoid all physically unavoidable collisions. If the vehicle is already outside the parking reserve, immediately request conservative deceleration; only through subsequent calculations can the next safe step be determined, and it cannot be deduced that the current state is safe. When communication times out but the vehicle's perception is effective, disable the failed semantic edge and increase the response margin; when physical forward perception also fails, do not continue to rely on neural network prediction to maintain tight formation.

[0101] S6. Execute the safety execution instruction and collect the actual execution actions; the actual execution actions are used for the online update of the graph world model, and after the actual execution actions are stored, they are paired with the vehicle observation features collected in the next cycle for the offline update of the graph world model and the shared cooperative strategy network.

[0102] In this embodiment, the safety command carries the vehicle number, control cycle, generation time, validity period, acceleration unit, model version, and safety status code. The validity period in this embodiment is one control cycle; expired, mismatched, duplicate, or version-incompatible commands are rejected and processed by the vehicle's own safety control system. The vehicle-side longitudinal controller converts acceleration requirements into drive torque or braking requests, which are then executed through the calibrated vehicle interface.

[0103] Simultaneously, each cycle records observations, graph relationships, message information age, initial actions, safety instructions, actual acceleration, next observation, reward, termination flag, and safety correction flag. The measured average acceleration is primarily obtained through velocity change and IMU fusion, and the error bound is recorded. Sensor time and execution time are recorded to avoid combining misaligned data into state transitions.

[0104] Multi-vehicle transfer records are constructed using the following formula:

[0105] In the formula, The set of five-branch observation data and their effective masks for each controlled vehicle in period t is used to preserve the original physical quantities, timestamps and target identities in order to reconstruct the observation features of the vehicle according to a fixed coding rule. The next period's set of observations defined for the same field. This is the set of two types of adjacency relationships for this cycle: physical sensing and V2V communication. This is a set of information ages stored according to incoming edges and relation types. The age of communication information is calculated from the message generation time, and the age of physical sensing information is calculated from the measurement sampling time. The unit for both is seconds. Failed edges are shielded by adjacency flags. , and These are vectors composed of the original acceleration, safety command acceleration, and measured period-averaged acceleration of each controlled vehicle, all in m / s². 2 ; The dimensionless full-scale reward corresponding to this transfer; This is a record of multiple vehicle transfers. To continue tagging, State codes are handled for safety. The world model pairs the actual next state with the measured executed actions; actuator identification pairs safety instructions with actual actions; actor imagination training advances the model with actions derived from the original actions through the safety layer and the response model. All three types of actions are retained and do not substitute for each other.

[0106] The actual execution actions are used for the online update of the graph world model: During online inference, the graph world model uses the measured actual acceleration from the previous cycle as the state transition input, and only updates the hidden state and edge memory online, without modifying the weights in real time; at the end of the current cycle, the posterior latent state is encoded into semantics, and action intent is generated as a safety instruction, which is broadcast for use in subsequent cycles, thereby eliminating the loop of "the world state of this cycle depends on the semantics not yet generated in this cycle" in terms of timing. Optional tracking error correction is completed by adding first-order model residual prediction, with the residual having m / s 2 The unit's upper bound for error needs to be re-verified; online, the model weights are not directly modified, and offline, after verification by leaving out scenarios, a new version is released. The model versions are synchronized at both ends of the communication to ensure consistent semantic interpretability.

[0107] Based on this, the actual executed actions are stored and paired with the vehicle's observation features collected in the next cycle for offline updates of the graph world model and the shared collaborative policy network. Offline updates include five sub-steps: trajectory cleaning, teacher prediction, supervised label construction, world model optimization, and policy optimization. The inputs from the three teachers all come from trajectories visible up to time t. Training labels can be used for future measurements, but these future labels cannot be read online. In one embodiment, the specific configuration is: a 20-step historical window (2 seconds), where zeros are padded and a mask is attached when there are insufficient targets.

[0108] In this model, the teacher in the leading vehicle uses a directed graph aggregation plus a two-layer LSTM structure: the relative position, speed, and acceleration of the vehicle itself, the actual leading vehicle, and the observable leading vehicle are encoded as 64-dimensional node features, averaged along the direction from the leading vehicle to the following vehicle, and then concatenated with the current node, and projected through 128→64; the 20-step features are fed into a two-layer 128-unit LSTM, and then output the trajectory for the next 30 steps through 128→256→90.

[0109] The surrounding teachers utilize a spatiotemporal graph network: each target's 20-step history is processed using a shared GRU to obtain 64-dimensional features, and two layers of graph attention transfer information between observed targets; each layer has 4 heads, each head has 16 dimensions, and the query-key product is divided by 4 and then processed using effective neighbor Softmax, with the four heads concatenated to obtain 64 dimensions. Each target decodes 3 30-step 2D trajectories, 3 pattern probabilities, and cut-in / lane-change probabilities. During training, the measured future trajectory is assigned to the pattern with the smallest average displacement error, and Smooth-L1 is applied to this pattern, with the pattern probability using the cross-entropy of the selected pattern; no corresponding loss is calculated when no valid target exists.

[0110] The leading teacher uses a similar historical GRU and graph aggregation approach. It takes the main forward target and visible downstream traffic as input and outputs a 50-step prediction as the offline label for the corresponding student head. Three teachers are first trained and frozen using masked real future labels, and then replay teacher labels are generated. Teacher labels are used for student distillation, while the measured futures are still used for teacher verification and student supervision. When no available teacher weights are available, student heads can be trained using the same defined measured future labels.

[0111] To combine the supervision losses of each student group during student distillation, the following distillation loss formula is used:

[0112] In the formula, The dimensionless intra-group average loss between the head output of the m-th student group and the effective supervision labels of the group is given. The supervision labels come from the frozen teacher output or the effective measured labels, specifically corresponding to D1 to D8 in Table 2. When the same group contains multiple output branches, the mean of the effective elements of each branch is calculated first, and then the branches with effective labels are averaged equally. The equally weighted average of the losses for the effective student groups. For student groups, This indicates that the group has valid labels; if there are no valid groups, the entire term is defined as 0. Within a group, the mean is first taken based on the valid entities, prediction steps, and components to avoid the 2160-dimensional D6 dominating the 2-dimensional D1 simply because of its large dimensionality. Continuous quantities use normalized Smooth-L1, binary probabilities use binary cross-entropy, and multi-class or teacher soft distributions use cross-entropy. D6 teacher and student patterns are aligned using a one-to-one matching method with the minimum trajectory cost. A maximum of 6 permutations are enumerated for the three patterns, and then the trajectory and pattern probability loss are calculated.

[0113] To update the world model, the following joint loss is constructed:

[0114]

[0115] In the formula, the expected average is taken over the effective vehicles and time steps, and 0.1 is the dimensionless weight of the kinematic consistency loss; D8 already includes the reward and continuation loss, and will not be calculated again. The reward label is first transformed into... When reasoning, use Restore; Continue Tag The value is 0 for actual terminations such as collisions and successful arrivals, and 1 for normal transitions; when the sampling time is truncated due to the upper limit, the bootstrap function is retained and the truncation flag is saved separately. The sum of the KL divergences of the 16 categorical distributions is calculated using the natural logarithm. To stop the gradient; 0.6 is the free information threshold in NAT terms. , The configurations implemented do not contain any physical dimensions.

[0116] The world model loss simultaneously updates the feature encoder, edge memory, state recursion network, student head, and semantic encoder, so that the "received interaction information" affects the world model through the prediction loss, and then affects the policy through the imagined reward; the offline teacher weights are frozen and cannot change the labels along with the student gradient.

[0117] Based on this, offline policy updates begin from the posterior state and unfold within the shared-world model. The multi-vehicle trajectory is calculated in steps, lasting 3 seconds. Each initial action is processed through a safety projection and actuator response model consistent with the online model to obtain the predicted actual acceleration before being input into the next state transition. The imagined state must not be advanced using uncorrected initial actions, but the strategy must be updated with the benefits of safe actions.

[0118] Each imaginary step first generates a quantized message using the latent state from the previous imaginary step. The message queue and edge memory are advanced according to the sampled communication delay, and then the prior knowledge of each vehicle is advanced by the graph context. The communication delay takes values ​​of 0, 0.1, and 0.2 seconds, with probabilities of 0.85, 0.075, and 0.075, respectively. Extended training may incorporate continuous packet loss, which must be reported separately during validation. External targets are predicted by D2, D5, and D6, and valid masks are retained. When using predicted states for hard constraints, the validated prediction error bound is additionally considered. Future measured values ​​are not input into the imaginary trajectory during training.

[0119] Accordingly, the shared Critic uses the same 536-dimensional input as the Actor, and outputs a scalar value through two layers of 256-unit SiLUs. Each vehicle stores its hidden state and generates its own actions; "shared" means consistent weights, and "cooperation" comes from the incoming edge context, the joint imagined trajectory, and the average reward of the entire formation, but does not mean that all vehicles output the same acceleration.

[0120] To aggregate the rewards from imagined trajectories of multiple vehicles, the following recursive formula for λ rewards is adopted:

[0121] In the formula, , The predicted formation reward is the average of the D8 outputs of each vehicle after inverse transformation. Each vehicle's D8 is supervised by the same measured global reward. The minimum predicted probability of each vehicle continuing is taken as the continuing coefficient for the joint expansion. This represents the average output of the Critic for each vehicle. For the aforementioned 536-dimensional complete input, To share Critic, Its parameters; discount factor Return Mixing Coefficient The continuation label for terminated samples is 0, and time-truncated samples are not mistakenly identified as true terminated samples when bootstrapping is possible.

[0122] To update the policy network, the following policy loss is constructed:

[0123] In the formula, behavior anchoring loss Divide the difference between the current Actor's deterministic original motion and the replayed original motion by 6 m / s 2 The squared mean is calculated only for normal replay samples that have not triggered security corrections; the mean of each role is calculated separately and then averaged equally; roles without corresponding samples are masked. This is the number of strategy updates, starting from 0. Divide the difference between the original action and the safety command by 6 m / s 2 The mean squared value can be taken as... Critic's prediction for each vehicle With the global stopping gradient Calculate the mean square error and then average it over the vehicle and time steps. Safety constraints do not need to rely on increasing the reward / penalty coefficient to take effect.

[0124] Training network parameter updates are separated from online road inference. Only filtering, policy forward propagation, and safety processing are performed per online cycle; the 30-step imagination is primarily used for offline Actor learning. If the deployment platform supports online candidate rolling filtering, this part is counted separately within the cycle time limit; the imagination time domain during training cannot be represented as a computationally unnecessary online search.

[0125] Offline updates use an average reward calculated over the number of vehicles to avoid limitations that prevent expansion once a fixed six-vehicle coefficient is applied. Rewards only assist performance learning; hard safety is handled by projecting the aforementioned safety constraints. The minimum value among the road speed limit, V2I recommended speed, effective downstream average speed, and upper limit of the curve preview is taken; the following vehicle is then compared with the predicted average speed of the vehicle ahead over the next 1 second, and the smaller value is taken. If the prediction of the vehicle ahead is invalid, the current measured speed is used; if there is no vehicle ahead, this item is disabled. The legally mandated road speed limit must be valid; if it is missing, it reverts to the existing minimum risk control. Example: V2I recommended speed 13.89 m / s, road maximum speed limit 33.33 m / s.

[0126] To establish the upper limit of the preview speed for curves, the following formula is used:

[0127] In the formula, curvature Unit m -1 , m -1 Allowing lateral acceleration Unit m / s 2 In this embodiment, 1.0 m / s is used. 2 The actual value is selected based on the vehicle's roll stability and traction; for straight road sections, the smaller value is taken from the road speed limit. Regarding distance... Target speed at the location ,by Construct an upper bound for the currently achievable preview speed, where Normal usable deceleration, unit m / s² 2 In this embodiment, 2.0 m / s is used. 2 And adjust according to the actual vehicle capability; the response distance also needs to be deducted from the preview distance.

[0128] To evaluate the deviations in velocity, spacing, and rate of change of acceleration, the following normalized error formula is used:

[0129] In the formula, m / s is used to prevent division by zero at low speeds; m is the distance scale; m / s 3 This is the scale for the rate of change of acceleration. For vehicles without a preceding vehicle, the distance term is not calculated; this term is set to 0 in the reward formula, and the final calculation still takes the average over all controlled vehicles N.

[0130] To estimate the driving resistance and positive traction power, the following physical model formula is used:

[0131] In the formula, The mass is in kg. m / s 2 , air density kg / m³ 3 , The drag coefficient is a dimensionless coefficient. For the windward area m 2 , The dimensionless rolling resistance coefficient. The slope angle is positive for uphill sections, measured in rad. Transmission efficiency; resistance is measured in N, and power in W. This is a forward traction power proxy, not fuel mass flow rate; fuel quantity needs to be calculated in conjunction with engine speed, torque, and idle fuel consumption calibration charts. Without setting a distance-related drag coefficient, energy savings are not attributed to formation aerodynamic drag reduction not present in the model.

[0132] To construct the cost per vehicle and the global reward, the following formula is used:

[0133] In the formula, The dimensionless cost of a single bicycle; Huber function with threshold 1: takes the value when the absolute value is not greater than 1. Take the rest ; The normalized parking reserve is insufficient; W. Optional weights are 1, 2, 0.2, 0.2, and 5, all dimensionless. Upon reaching the destination and completing the task... A collision results in a value of -100, mission failure and actual termination in a value of -20, and all others in a value of 0. Safety deficiencies due to the absence of a preceding vehicle are also masked; sensor failures are handled separately under safe conditions. If potential function shaping is used, it should be adopted... In accordance with the discount and termination agreement, the potential difference between any two adjacent moments cannot be considered as shaping that necessarily maintains the original optimal strategy; the basic implementation can be carried out without relying on potential function shaping.

[0134] Table 3 shows a set of engineering implementation configurations for this invention. The network dimension corresponds to the input / output dimensions of the aforementioned observation encoder, edge memory network, graph world model, and shared policy network; the control cycle is used for state updates and command issuance; and the nominal vehicle parameters are used for example calculations. Safety reserves, timeout handling, actuator interfaces, and training hyperparameters are set according to the method described in this embodiment. The example values ​​in the table are used to illustrate the implementation method; the vehicle physical parameters should be calibrated and determined based on the specific vehicle model, load, slope, and road surface adhesion conditions.

[0135] Table 3 Project Implementation Configuration

[0136] The network is initialized using a fixed seed of 2026, and the input linear layer is initialized using Xavier with a bias of 0. The GRU uses a standard gating structure. Replay is divided into training, validation, and testing phases based on complete scenes, with ratios of 70%, 15%, and 15%, respectively. Trajectory segments within the same scene and their adjacent segments do not cross sets. A conservative sampling controller takes the velocity error multiplied by 0.5 seconds. -1 Add the effective spacing error multiplied by 0.1s -2 , to obtain m / s 2 The candidate acceleration of the unit is also projected through the aforementioned safety parameters; the time distance without a preceding vehicle is 0. First, use the controller to generate at least 10,000 usable samples for model pre-training, then alternately sample environmental interactions and update the network; freeze the parameters after verifying that the loss is stable and the scenario test passes.

[0137] The offline training steps are as follows: T1: Collect multi-vehicle trajectories with a unified timestamp and construct the playback record of formula (25). The complete road segment scene is divided into training, verification, and testing at 70%, 15%, and 15% respectively; generate future supervision labels separately within each set, and prohibit the same traffic wave trajectory slice from being dispersed into training and testing. Invalid components of sensors, vehicle identities, and model versions are masked.

[0138] T2: Pre-train three teachers using the aforementioned method and freeze them to construct eight sets of student labels. When teachers are unavailable, generate actual test supervision labels directly using the same fields. Offline label generation will not increase online sensor input.

[0139] T3: Pre-train the world model using at least 10,000 conservative control interactions, with a maximum of 50,000 gradient updates in the example; calculate the group-balanced prediction loss on the validation set every 1,000 updates, and stop early if the loss does not decrease by 0.1% after 10 consecutive validations, retaining the best version. This stopping condition is explicitly implemented, not an optimality guarantee.

[0140] T4: Freeze the world model weights to update the Actor and Critic, but allow the Actor's gradient to be backpropagated via the derivative of the world model with respect to the action. Each time, take 16 sequences of 64 steps each; the first 5 steps are only for warm-up, and the posterior starting point is extracted to expand into 30 steps of joint imagination. First calculate the global reward and λ reward, then update the Critic, and finally update the Actor. The model, value, and policy are updated every 16 steps of real-world interaction. The total interaction budget for training can be 300,000 steps, and checkpoints can still be pre-selected for validation. The Critic objective uses the λ reward of the stopping gradient; if a target Critic is used, the target parameters are updated each time by adding 0.01 times the new value to 0.99 times the old value.

[0141] T5: After validating the prediction error, control failure rate, and safety layer infeasibility rate in the reserved scenario, and confirming that the inference time limit meets the 100ms control cycle, freeze the online parameters. The release package includes the input field order, normalization scale, network weights, semantic quantization range, model version, and validation values. All communication vehicles must use the compatible version. Only the hidden state and edge memory are updated online; the weights are not modified in real time.

[0142] Within each control cycle online, the following steps are performed: R1: Lock the sensing and message snapshot at the cycle reception deadline. Complete target association, time alignment, preservation of original physical quantities, and 256-dimensional encoding, and output the physical map and effective communication relationships respectively.

[0143] R2: Read arrived messages by vehicle number, decode and calculate the real information age, update the valid incoming edge GRU, and aggregate into a 128-dimensional graph context; new messages and new physical measurements are marked respectively, and timed-out communication edges are removed.

[0144] R3: The actual acceleration measured in the previous cycle, the previous potential state, and the current graph context advance 128-dimensional deterministic states, which are then fused with 256-dimensional observations to obtain the classification posterior, and then form 384-dimensional potential states and eight sets of predicted outputs.

[0145] R4: Reads the potential state, context, role, and local physics vectors, generates dimensionless actions, and then maps them to -3.0 to 1.2 m / s. 2 The original acceleration. During deployment, only one policy forward pass is performed; there is no retraining or 30-step policy optimization every 100ms cycle.

[0146] R5: Calculate the feasible set using the current measurement without truncation, all path conflict objects, and vehicle capabilities, and project it to obtain a safety instruction; if infeasible, give a normal, emergency feasible, or emergency infeasible status code according to the emergency handling procedure.

[0147] R6: Verify the validity period of the instruction and execute it, recording the actual response; at the end of this cycle, encode the posterior potential state into semantics, generate an action intent with a security instruction, and broadcast it for use in subsequent cycles. After writing and replaying, proceed to the next cycle.

[0148] This sequence ensures that the message read by R2 has arrived before the solution of R3 in this cycle, and the semantics of the current cycle sent are only used in subsequent receiving cycles, thus eliminating the cycle of "the world state of the current cycle depends on the semantics that have not yet been generated in the current cycle" in terms of timing.

[0149] The following are several specific embodiments to illustrate the present invention in detail: In a single-cycle numerical embodiment, the following vehicle speed is taken as 13.89 m / s, the actual preceding vehicle speed is 13.00 m / s, the net distance is 30.00 m, the actual acceleration in the previous cycle is 0, the sampling period is 0.1 s, and the actuator time constant is 0.2 s. A guaranteed net deceleration of 6.0 m / s² is assumed. 2 Total response time: 0.3s; Upper limit of positive acceleration during response period: 1.2m / s² 2 The parking margin is 2m and the error margin is 0.5m.

[0150] According to equation (18), the response distance is 4.221m, the response end speed is 14.25m / s, the subsequent braking distance is 16.921875m, and the total parking reserve is 23.642875m. Therefore, the target spacing is taken as 23.642875m, instead of directly taking the nominal 12m. In this example, the current net spacing of 30m satisfies this conservative reserve.

[0151] According to equation (20), the time interval margin is 14.11m. The lower limit of the acceleration of the preceding vehicle is taken as -6m / s². 2 With a contraction coefficient of 0.2 and a step distance margin of 0.1m, the numerator of equation (21) is 2.603m and the denominator is 0.105s. 2 The upper limit of the actual acceleration is obtained as approximately 24.79 m / s². 2 This value does not mean that the vehicle can achieve this acceleration; ultimately, it must be compared with 1.2 m / s². 2 The vehicle model instruction limit and other constraints intersect.

[0152] Assume the initial acceleration of the strategy is 0.8 m / s². 2 From equation (16), we can obtain The normal rate of change limits the actual acceleration to increase by no more than 0.15 m / s² per cycle. 2 Therefore, the upper bound of the command is approximately 0.38122 m / s. 2 If the parking reserve and other checks pass, the final speed will be 0.38122 m / s. 2 The actual acceleration for this cycle is expected to be 0.15 m / s². 2 The next velocity is 13.905 m / s, and the displacement is 1.38975 m. Record three independent fields: the original 0.8, the issued 0.38122, and the actual measured acceleration. The model will be updated with the measured values ​​in the next cycle.

[0153] This example also illustrates that: network actions, issued commands, and actual response values ​​may differ; dynamic constraints are more stringent than fixed amplitude truncation; and all distance, velocity, and acceleration calculations are performed in physical units.

[0154] Furthermore, in one embodiment of communication delay and continuous packet loss, the vehicle A semantic message is generated in 1.0 seconds for the vehicle. If the message is held at the 1.2s control time, the actual information age is 0.2s. Edge memory is advanced using dequantized semantics, normalized action intent, the 0.2s information age, and the update flag. If no new message is received by 1.3s, the cached content continues to be used, but the information age is incremented to 0.3s, and the update flag is set to zero. If the control time reaches 1.6s and there is still no update, the information age is 0.6s, exceeding the 0.5s threshold; this communication edge does not participate in aggregation. The actual preceding vehicle detected by the vehicle's radar still participates in physical state coding and safety constraints. It is forbidden to map 0.6s to 0.2s and then continue to claim the applicability of short-latency compensation. After receiving a new version of compatible message, re-filtering is performed to avoid directly restoring tight formation using the stale edge state before the communication interruption.

[0155] In one embodiment of cut-in and formation recovery, when an external vehicle begins to intrude into the driving path of the vehicle in front of it, it is added to the safety constraint object according to the target occupancy range; if necessary, both the original preceding vehicle and the cutting-in vehicle constraints are retained, and the stricter boundary is taken. After the actual preceding vehicle identity is switched, equations (1), (18) to (22) are all recalculated with the cutting-in target, and the leading role identity of the formation communication can remain unchanged temporarily. If the net distance after cutting in is only 10m, which is less than 23.642875m parking reserve under the above speed and braking assumptions, emergency risk handling is immediately initiated, and the 8m target distance is not used to continue accelerating the closed space. After the target leaves, the forward perception, communication and braking capabilities are verified first, and then the formation is gradually restored through the same shared strategy; the restoration process still follows the normal acceleration and rate of change constraints.

[0156] In one embodiment of curve and traffic wave, when the map preview shows an increase in curvature ahead, a speed limit is formed according to equation (31), and then an advance deceleration boundary is determined based on the distance to the curve entrance and the normal deceleration. A radius of 100m and an allowable lateral acceleration of 1.0m / s² are used. 2 For example, the maximum speed limit on a curve is 10 m / s; if the current speed exceeds this value, sufficient preview distance is required for braking, and speed cannot be instantaneously cut off upon entering the curve. Under continuous start-stop disturbances ahead, the graph world model uses historical motion, downstream traffic conditions, and edge memory to predict changes in the speed of the vehicle ahead. The shared strategy reduces speed deviation, spacing deviation, and high-frequency acceleration variations through unified rewards. The string stability performance should be verified using disturbance transfer gain or speed fluctuation metrics; it is not advisable to claim that strict string stability at all frequencies has been proven simply because the error is small.

[0157] In one optional implementation of the lane-changing interface, when a lateral planner has proposed a target lane and predicted trajectory, this invention receives the candidate trajectory, estimated occupancy time, and target states before and after the target lane, updates the physical relationships in the graph, and outputs longitudinal coordination instructions. Lane-changing permission is only granted if the lateral planner's collision check, road marking and lane feasibility check pass, and the longitudinal constraints of this invention are feasible. If any vehicle does not meet the conditions, the entire platoon can delay or cancel the lane change. The interface output includes the target lane number, allow or deny flag, effective time window, and longitudinal safety instructions for each vehicle. Steering wheel angle, yaw stability, and tractor-trailer articulation stability are handled by the existing lateral controller and are not replaced by the aforementioned longitudinal model. The lateral planning interface and teacher distillation are optional implementations and are not prerequisites for the existence of this technology.

[0158] The beneficial effects and verification methods of the above scheme are explained below. It should be noted that low communication load is not equivalent to low end-to-end latency, nor is a limited number of collision-free results equivalent to unlimited operational safety guarantees. For configurations with supplementary safety implementations, control errors and energy consumption should be remeasured.

[0159] Based on this, it is recommended to compare the complete solution, the solution with side memory removed, the original action replay solution, and the independent predictive control baseline under the same road segment, traffic demand, random seed, reward, and input information conditions. Scenarios should be set up for normal communication, 100-200ms latency, continuous packet loss, external intrusion, curve preview, and complex traffic wave, retaining failed and premature termination samples. Simulation should be performed first, followed by hardware-in-the-loop testing, and finally, real-vehicle verification under authorized experimental conditions.

[0160] To evaluate the accuracy of formation spacing control and the rate of change of acceleration, the following index formula is used:

[0161] In the formula, For signs with actual preceding vehicles and valid measurements, The number of sampling times is at least 2. GMAE is in meters (m), and JRMS is in meters per second (m / s). 3 It must be stated whether a dynamic target spacing or a fixed 12m is used; the two values ​​cannot be directly compared. JRMS data for normal and emergency braking phases should be reported separately to avoid using emergency phase data to judge the normal 1.5m / s. 3 Whether the constraint was incorrectly implemented.

[0162] Furthermore, fuel intensity is calculated with the actual driving distance as the denominator; the total fuel consumption of the six vehicles is divided by the distance traveled by the lead vehicle to obtain the platoon-level kg / km. If vehicle kilometers are used, they should be divided by the sum of the driving distances of each vehicle. The difference between the two is approximately the number of vehicles in a platoon, and they should not be used interchangeably. Failure samples use actual driving distances and do not consider the destination mileage as the distance already traveled.

[0163] Simultaneously, the online computation time is recorded for the median, 95th percentile, and maximum values ​​of perception alignment, network inference, security solution, and command issuance. In this example, a 100ms control cycle without automatic operation means the worst-case time is less than 100ms; if the timeout is not completed, expired outputs are rejected, and conservative control of the vehicle is switched to the current vehicle, with the timeout rate calculated.

[0164] For configurations including safety projection and actuator response processing in this embodiment, the acceleration change rate during normal operation and emergency braking should be calculated according to equation (35), and the sample size, maximum absolute change rate, and constraint violation rate for each stage should be reported simultaneously. The normal operation phase is calculated using equation (23) at 1.5 m / s². 3 The rate of change is constrained; if the absolute value of the actual rate of change of acceleration in each effective sampling interval during this stage does not exceed this value, then its root mean square will also not exceed 1.5 m / s². 3 During the emergency braking phase, calibrated mechanical braking is used to establish the rate and adhesion capability boundaries, and the results are statistically analyzed separately. Sampling intervals that cross the transition between normal and emergency states are marked separately and not included in the statistics for the normal phase. Statistical results from different configurations or without distinguishing between operating phases cannot be used as verification that the normal rate of change constraint has been met in this embodiment.

[0165] Finally, the expected beneficial effects and verifiable metrics include: the accuracy of target association after the two types of graph separation; the trajectory error and communication failure recognition rate under delayed conditions corresponding to information age edge memory; the one-step velocity prediction error corresponding to consistent playback of executed actions; and the constraint violation rate, emergency infeasibility rate, and comfort corresponding to physical safety projection. Metrics must be reported separately along with the number of samples, test boundaries, failed samples, and normal / emergency states.

[0166] Please see Figure 3In an optional embodiment, the present invention provides a heavy truck platooning control system based on a latent communication graph world model. The system includes a state acquisition and graph construction module 101, a semantic communication and edge memory module 102, a graph world model module 103, a shared cooperative strategy module 104, a physical safety projection module 105, a vehicle execution and feedback module 106, and an offline training module 107, used to execute specific steps of the relevant embodiments of the heavy truck platooning control method based on a latent communication graph world model provided by the present invention. Dashed lines represent offline supervision or parameter publishing, solid lines represent online data and control signals. Modules 101 to 106 constitute an online control closed loop. Module 107 uses playback sequences to train and publish verified model parameters. Table 4 shows the verifiable interfaces between modules. Within the same control cycle, processing is performed in the order of 101, 102, 103, 104, 105, and 106. The results acquired by 106 are used by 101 and offline 107 in the next cycle. Among the six online modules of the present invention, 103 has a built-in predictive student decoding head; offline teacher knowledge injection is incorporated into 107. The heavy truck platooning control system based on the potential communication graph world model provided by this invention has a complete structure, is objective and stable, and improves the overall applicability and practical application capability of this invention.

[0167] Table 4 Module Composition and Input / Output Relationships

[0168] In summary, this invention reduces the interference of communication loss on the perception of the actual vehicle ahead by separating two types of relationships; it maintains continuous updates of interactive predictions under short latency and packet loss by using information age-driven edge memory; it reduces the semantic mismatch between model learning and actual vehicle response by sharing a graph world model and action feedback; and it ensures that the policy output falls within the constraints achievable by the current vehicle through explicit physical projection and infeasibility handling. The specific modules and processing steps corresponding to these effects can be verified through ablation and scenario comparison.

[0169] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications 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 the present invention, and they should all be covered within the scope of the claims and specification of the present invention.

Claims

1. A method for heavy truck platooning control based on a latent communication graph world model, characterized in that, Includes the following steps: Acquire vehicle and target measurement data, V2V communication messages, vehicle capability information and road information, construct mutually independent physical perception relationships and V2V communication relationships, and output vehicle observation characteristics, raw physical state quantities and target valid indicators; Determine the information timeliness of each edge in the physical perception relationship and the V2V communication relationship, maintain the edge memory state of each edge by combining the information timeliness with the target validity flag, filter out valid edges, and aggregate the edge memory states of the valid edges to obtain the graph context. The graph context and the actual execution action of the previous cycle are input into the graph world model, and observation fusion is performed by combining the observation features of the vehicle to output the potential features. Based on the latent features and the graph context, the original action is generated through a shared cooperative policy network; By combining the physical perception relationship, the original physical state quantity, and the vehicle capability information, a safety constraint projection is performed on the original action, and a safety execution command is output. The safe execution instruction is executed to collect actual execution actions; the actual execution actions are used for online updates of the graph world model, and after the actual execution actions are stored, they are paired with the vehicle observation features collected in the next cycle for offline updates of the graph world model and the shared cooperative strategy network.

2. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The process involves acquiring measurement data of the vehicle and the target, V2V communication messages, vehicle capability information, and road information; constructing independent physical perception relationships and V2V communication relationships; and outputting the vehicle's observation characteristics, raw physical state quantities, and valid target indicators, including: Based on the vehicle and target measurement data and the road information, five-branch coding and normalization aggregation are performed to output the vehicle observation features; Based on the measurement data of the vehicle and the target, the original physical state quantity is output, and the target validity flag is generated simultaneously; The physical perception relationship is constructed based on the measurement data of the vehicle and the target. The physical perception relationship includes physical perception edges, which correspond to the motion state of the perceived target. The validity of the edge is determined by the target validity flag. The V2V communication relationship is constructed based on the V2V communication messages. The V2V communication relationship includes V2V communication edges, and the V2V communication edges correspond to the semantic interaction states of vehicles within the platoon.

3. The heavy truck platooning control method based on a latent communication graph world model according to claim 2, characterized in that, Determining the timeliness of information on each side of the physical sensing relationship and the V2V communication relationship includes: The information timeliness of the V2V communication edge is determined based on the difference between the generation time of the V2V communication message and the current time, and the communication information age is generated; the information timeliness of the physical sensing edge is determined based on the difference between the sampling time of the vehicle and target measurement data and the current time, and the sensing information age is generated; the communication information age and the sensing information age are calculated and used independently.

4. The heavy truck platooning control method based on a latent communication graph world model according to claim 3, characterized in that, The process of maintaining the edge memory state of each edge by combining the timeliness of the information with the target validity flag, filtering out valid edges, and aggregating the edge memory states of the valid edges to obtain the graph context includes: The recursive edge memory state of the V2V communication edge is maintained by combining the communication information age, and the recursive edge memory state of the physical sensing edge is maintained by combining the perception information age. The validity of the V2V communication edge is determined based on the age of the communication information. When the age of the communication information exceeds a preset failure threshold or the model version is incompatible, the V2V communication edge is determined to be an invalid edge and removed from the aggregation range. The validity of the physical sensing edge is determined based on the target validity flag. When the target validity flag indicates that the corresponding sensing target is invalid, the physical sensing edge is determined to be an invalid edge. After removing the invalid V2V communication edges and physical sensing edges, the valid edges are obtained. The edge memory states of all valid edges are summed and normalized, and the graph context is obtained through aggregation transformation. When the V2V communication edges are invalid, the physical sensing edges still participate in the aggregation calculation independently.

5. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The process of inputting the graph context and the actual actions executed in the previous cycle into the graph world model, combining the observation features of the vehicle itself for observation fusion, and outputting latent features includes: The graph context, the actual actions performed in the previous cycle, and the deterministic and random states of the previous cycle are input into the graph world model to recursively obtain the deterministic state of the current cycle. The prior distribution of the random state is obtained based on the deterministic state of the current period; the posterior distribution of the random state is obtained by fusing the observation features of the current vehicle with the deterministic state of the current period, and the prior distribution is only used when there is no imagination to unfold in the future; The latent feature is constructed by combining the current periodic deterministic state and the random state corresponding to the posterior distribution using the random state corresponding to the posterior distribution.

6. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The generation of original actions based on the latent features and the graph context through a shared cooperative policy network includes: The latent features, the graph context, the role vector, and the local physics vector are input into the shared cooperative policy network, and the shared cooperative policy network outputs the mean and standard deviation of the original actions; Based on the mean and the standard deviation, an action variable is generated, and the action variable is mapped to an asymmetric action interval between the upper and lower limits of normal operating acceleration to obtain the original action.

7. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The method of combining the physical perception relationship, the original physical state quantity, and the vehicle capability information to project a safety constraint onto the original action and output a safety execution command includes: Based on the physical perception relationship, the original physical state quantity, and the vehicle capability information, calculate the parking reserve and time interval safety margin; based on the time interval safety margin and the kinematic recursive relationship, calculate the upper limit of the actual acceleration; Based on the first-order discrete response relationship, the upper bound of the actual acceleration is transformed into the upper bound of the command, and the first-order discrete response relationship is characterized by the equivalent time constant to represent the dynamic response of the actuator. By combining the upper bound of the instruction and the parking reserve to determine the feasible range, the original action is projected onto the feasible range, and the safe execution instruction is output.

8. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The execution of the security execution command is carried out, and the actual execution actions are collected. The actual actions performed are used for the online updating of the graph world model, including: Verify the validity of the security execution command and execute it; collect and store the actual execution actions for this cycle. The actual actions performed in the current cycle are used as state transition inputs for the graph world model in the next cycle for online updates. Online inference only updates the hidden states and the edge memory states, without modifying the weights in real time.

9. The heavy truck platooning control method based on a latent communication graph world model according to claim 1, characterized in that, The actual executed actions are stored and paired with the observed features of the vehicle collected in the next cycle for offline updates of the graph world model and the shared cooperative policy network, including: The actual actions performed are paired with the vehicle's observation features collected in the next cycle to construct a multi-vehicle transfer record; Using the multi-vehicle transfer records, the actual executed actions and the vehicle's observation features collected in the next cycle are used as real state transfer samples to update the graph world model offline; Using the multi-vehicle transfer records, starting from the random state corresponding to the posterior distribution, an imagined trajectory is unfolded in the graph world model. The original action at each step is processed by the safety constraint projection and actuator response model to obtain the predicted actual action. The predicted actual action is used to advance the next state transition. The shared cooperative strategy network is updated offline based on the reward of the imagined trajectory.

10. A heavy truck platooning control system based on a latent communication graph world model, characterized in that, The system includes a state acquisition and graph construction module, a semantic communication and edge memory module, a graph world model module, a shared cooperation strategy module, a physical security projection module, a vehicle execution and feedback module, and an offline training module, and executes the heavy truck platooning control method based on the latent communication graph world model as described in any one of claims 1-9.

Images