An unmanned heavy truck team cooperative obstacle avoidance method based on multi-agent cluster

Abstract

The application relates to the technical field of path planning, in particular to a method for cooperative obstacle avoidance of an unmanned heavy truck team based on a multi-agent cluster, which comprises the following steps: acquiring multi-modal sensing data on each unmanned heavy truck, constructing a team communication link according to the multi-modal sensing data, and generating leader vehicle information; acquiring obstacle identification information from each multi-modal sensing data, acquiring passing area information and obstacle information, and generating an obstacle probability distribution graph of the heavy truck team; when an unyielding obstacle is acquired from the obstacle probability distribution graph, generating obstacle avoidance route data of each unmanned heavy truck through the leader vehicle information; acquiring actual obstacle avoidance data corresponding to the obstacle avoidance route data, performing obstacle avoidance scoring according to the actual obstacle avoidance data, and optimizing the strategy for generating the obstacle avoidance route according to the obstacle avoidance scoring. The application has the effect of overcoming the problems that, when the existing unmanned heavy truck team drives in formation, the obstacle avoidance actions are inconsistent, and the formation is prone to confusion and collision.

Description

Technical Field

[0001] This application relates to the technical field of path planning, and in particular to a collaborative obstacle avoidance method for unmanned heavy truck fleets based on multi-agent clusters. Background Technology

[0002] With the rapid deployment of autonomous driving technology in the commercial vehicle sector, unmanned heavy trucks are demonstrating enormous application potential in scenarios such as highway logistics, closed-loop logistics, and port terminal transshipment. In these scenarios, multiple unmanned heavy trucks often travel in platoons to improve transportation efficiency, reduce wind resistance and energy consumption, and alleviate driver shortages. During platooning, vehicles need to navigate various dynamic and static obstacles, including accident vehicles, slow-moving vehicles, road debris, and construction areas. This places extremely high demands on the real-time performance and coordination of perception systems, communication links, and path planning.

[0003] Currently, the industry has accumulated some experience in perception and planning for unmanned heavy truck platooning. At the individual vehicle level, multimodal sensors such as LiDAR, millimeter-wave radar, and high-definition cameras are typically equipped, combined with local path planning algorithms to achieve obstacle avoidance. Vehicles communicate with each other to exchange simple warning information, such as sending a warning to the following vehicle when the preceding vehicle brakes suddenly. For platooning control, a lead vehicle-follower vehicle mode is commonly used. The lead vehicle is responsible for global path planning, while the follower vehicle maintains a preset safe distance through adaptive cruise control, thereby maintaining the basic stability of the platoon.

[0004] However, the above-mentioned scheme still faces a key technical problem in actual operation: because each vehicle relies solely on its own sensors for independent decision-making, and vehicle interaction is limited to one-way warnings or status broadcasts, there is a lack of real-time intent sharing and global optimal negotiation mechanisms among multiple vehicles. When the convoy encounters sudden and complex obstacles, the vehicles are prone to inconsistent obstacle avoidance actions. For example, the leading vehicle may choose to detour to the left, while the following vehicle continues straight due to a blind spot, leading to convoy chaos or even collisions. This fragmented decision-making approach cannot enable the entire convoy to achieve a globally optimal balance between dynamic constraints, convoy formation maintenance, and traffic efficiency. Summary of the Invention

[0005] To overcome the problems of inconsistent obstacle avoidance actions, which can easily lead to formation chaos and collisions when existing unmanned heavy truck fleets are driving in formation, this application provides a cooperative obstacle avoidance method for unmanned heavy truck fleets based on multi-agent clusters.

[0006] The above-mentioned objective of this application is achieved through the following technical solution: A method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters, comprising: Acquire multimodal sensor data pre-installed on each unmanned heavy truck in the heavy truck fleet, construct fleet communication links based on the multimodal sensor data, and generate leader vehicle information in the heavy truck fleet using a distributed election algorithm; Obtain obstacle recognition information from each of the multimodal sensing data, obtain passage area information and obstacle information from the obstacle recognition information, and generate an obstacle probability distribution map of the heavy truck fleet based on the passage area information and the obstacle information; When unavoidable obstacles are obtained from the obstacle probability distribution map, obstacle avoidance route data for each of the unmanned heavy trucks is generated using the information of the lead vehicle. Obtain the actual obstacle avoidance data corresponding to the obstacle avoidance route data, perform obstacle avoidance scoring based on the actual obstacle avoidance data, and optimize the strategy for generating obstacle avoidance routes based on the obstacle avoidance score.

[0007] Preferably, the step of constructing a fleet communication link based on the multimodal sensing data and generating leader vehicle information in the heavy truck fleet using a distributed election algorithm specifically includes: The communication signal strength data and computing resource data of each unmanned heavy truck are obtained from the multimodal sensing data; A comprehensive score is obtained for each unmanned heavy truck based on the communication signal strength data and the computing power resource data. The unmanned heavy truck with the highest navigation compatibility score is identified as the leader vehicle.

[0008] Preferably, the step of comprehensively scoring each unmanned heavy truck based on the communication signal strength data and the computing resource data to obtain a navigation compatibility score specifically includes: The communication delay data and packet loss rate data are obtained from the communication signal strength data, and the central processing unit occupancy rate and remaining memory capacity are obtained from the computing power resource data. A communication quality score is generated based on the communication latency data and the packet loss rate data, and a computing power availability score is generated based on the CPU utilization rate and the remaining memory capacity. The navigation fit score is obtained by weighted summing of the communication quality score and the computing power availability score.

[0009] Preferably, the step of obtaining obstacle recognition information from each of the multimodal sensing data, obtaining passage area information and obstacle information from the obstacle recognition information, and generating an obstacle probability distribution map of the heavy truck fleet based on the passage area information and the obstacle information specifically includes: A local occupancy grid map is obtained from each of the multimodal sensing data, and a grid occupancy probability value is extracted from the local occupancy grid map; The grid occupancy probability value is sent to the lead vehicle corresponding to the leader vehicle information. The lead vehicle then performs a weighted fusion of all the grid occupancy probability values ​​to obtain a global grid occupancy probability value. The obstacle probability distribution map is generated based on the global grid occupancy probability value.

[0010] Preferably, generating the obstacle probability distribution map based on the global grid occupancy probability value specifically includes: Obtain the dynamic grid region corresponding to the dynamic obstacle from the global grid occupancy probability value; Acquire temporal change data of the dynamic grid region, and predict the motion trajectory of the dynamic obstacle based on the temporal change data to obtain predicted trajectory data; Based on the predicted trajectory data, the global grid occupancy probability value is subjected to spatiotemporal filtering to obtain the filtered obstacle probability distribution map.

[0011] Preferably, when unavoidable obstacles are obtained from the obstacle probability distribution map, generating obstacle avoidance route data for each of the unmanned heavy trucks using the leader vehicle information specifically includes: When the unavoidable obstacle is obtained from the obstacle probability distribution map, multiple candidate obstacle avoidance schemes are generated based on the lead vehicle corresponding to the lead vehicle information. The multiple candidate obstacle avoidance schemes are sent to each of the unmanned heavy trucks, and the score data of each unmanned heavy truck for each candidate obstacle avoidance scheme is obtained. Based on the scoring data, a distributed consensus protocol is used for voting. When all the unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the candidate obstacle avoidance scheme that has reached a consensus is determined as the obstacle avoidance route data.

[0012] Preferably, the step of voting based on the scoring data using a distributed consensus protocol, whereby when all the unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the consensus-reaching candidate obstacle avoidance scheme is determined as the obstacle avoidance route data, specifically includes: The maximum steering angle, maximum braking deceleration, and current load data are obtained from the dynamic parameters of each of the unmanned heavy trucks. Based on the maximum steering angle, the maximum braking deceleration, and the current load data, a safety feasibility score is given for each candidate obstacle avoidance scheme to obtain the score data; The scoring data is sent to the lead vehicle, which counts the highest number of scores obtained for each candidate obstacle avoidance scheme and determines the candidate obstacle avoidance scheme with the most highest scores as the obstacle avoidance route data that has reached a consensus.

[0013] Preferably, the step of obtaining the actual obstacle avoidance data corresponding to the obstacle avoidance route data, performing obstacle avoidance scoring based on the actual obstacle avoidance data, and optimizing the obstacle avoidance route generation strategy based on the obstacle avoidance score specifically includes: Obtain formation holding deviation data, braking impact data, and additional energy consumption data from the actual obstacle avoidance data; The comprehensive score for this obstacle avoidance is calculated based on the formation holding deviation data, the braking impact data, and the additional energy consumption data. If the overall score is lower than the preset score threshold, an online fine-tuning instruction is triggered to adjust the strategy parameters for generating the obstacle avoidance route data. If the overall score is higher than or equal to the score threshold, the obstacle avoidance route data corresponding to this obstacle avoidance will be marked as an excellent case and saved.

[0014] In summary, this application includes at least one of the following beneficial technical effects: 1. By constructing a multi-vehicle real-time communication link and employing a distributed election algorithm, the lead vehicle is dynamically selected by comprehensively considering the communication signal strength and computing resource data of each vehicle, thus realizing the autonomous selection and optimal establishment of the fleet command node. Compared with fixed assignment or simple rotation, this method ensures that the lead vehicle has optimal data transmission and reception capabilities and real-time processing capabilities, providing a reliable organizational foundation for subsequent collaborative perception and decision-making; 2. By sending the local occupancy grid probability values ​​of each vehicle to the lead vehicle for weighted fusion, and transmitting only the compressed probability data instead of the original point cloud, a global obstacle probability distribution map covering hundreds of meters in front of and behind the entire convoy is generated while significantly reducing communication bandwidth consumption. The lead vehicle dynamically adjusts the fusion weights based on the historical confidence levels of each vehicle's sensors, giving higher confidence levels to obstacles observed simultaneously by multiple vehicles. This overcomes the limitations of single-vehicle perception blind spots and improves the coverage and accuracy of environmental detection. 3. Multiple candidate obstacle avoidance schemes are generated by the lead vehicle. All vehicles score the safety feasibility of each scheme based on their own dynamic parameters such as maximum steering angle, maximum braking deceleration, and current load. A consensus is then reached through multiple rounds of voting using a distributed consensus protocol. The scheme with the highest number of scores is ultimately determined as the obstacle avoidance route. This mechanism allows each vehicle to participate in the decision-making process based on its own physical capabilities, ensuring that the final scheme achieves optimal equilibrium under the safety constraints of all vehicles. This achieves collaborative path planning at the formation level, ensuring coordinated obstacle avoidance actions among all vehicles. 4. By collecting data on formation holding deviation, braking impact, and additional energy consumption during actual obstacle avoidance, a comprehensive score is calculated. This score triggers online fine-tuning or the saving of excellent cases, forming a closed-loop optimization mechanism for the obstacle avoidance strategy. When the score is below a threshold, reinforcement learning algorithms are used to adjust the strategy parameters, enabling the fleet to quickly adapt to specific road sections or operating conditions. When the score is above the threshold, the solution is marked as an excellent case for subsequent reuse. This allows for continuous accumulation of experience and optimization of performance over long-term operation, improving the method's adaptability to diverse scenarios. Attached Figure Description

[0015] Figure 1 This is a flowchart of a collaborative obstacle avoidance method for unmanned heavy truck fleets based on a multi-agent cluster in one embodiment of this application.

[0016] Figure 2 This is a flowchart illustrating the implementation of step S10 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0017] Figure 3 This is a flowchart illustrating the implementation of step S12 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0018] Figure 4 This is a flowchart illustrating the implementation of step S20 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0019] Figure 5 This is a flowchart illustrating the implementation of step S23 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0020] Figure 6 This is a flowchart illustrating the implementation of step S30 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0021] Figure 7 This is a flowchart illustrating the implementation of step S33 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application.

[0022] Figure 8 This is a flowchart illustrating the implementation of step S40 in a multi-agent cluster-based collaborative obstacle avoidance method for unmanned heavy truck fleets in one embodiment of this application. Detailed Implementation

[0023] The present application will be further described in detail below with reference to the accompanying drawings.

[0024] In one embodiment, such as Figure 1 As shown, this application discloses a cooperative obstacle avoidance method for unmanned heavy truck fleets based on multi-agent clusters, which specifically includes the following steps: S10: Acquire multimodal sensor data pre-installed on each unmanned heavy truck in the heavy truck fleet, construct fleet communication links based on the multimodal sensor data, and generate leader vehicle information in the heavy truck fleet using a distributed election algorithm.

[0025] In this embodiment, multimodal sensing data refers to environmental perception data collected by sensors such as LiDAR, millimeter-wave radar, and high-definition cameras installed on each unmanned heavy truck. Fleet communication links refer to real-time data exchange channels between vehicles established through C-V2X or DSRC ad hoc networking technologies. Leader vehicle information refers to the vehicle identification and related parameters selected to assume a temporary command role within the entire heavy truck convoy.

[0026] Specifically, when the unmanned heavy truck fleet prepares to enter highway logistics or closed-loop park pick-up scenarios, each vehicle first activates its onboard multimodal sensors to continuously collect point cloud data, radar echo signals, and visual images of the surrounding environment. Simultaneously, each vehicle broadcasts its communication address and computing resource status to the surrounding area via its C-V2X communication module. After receiving the broadcast information from all vehicles, the system sorts all vehicles according to a preset distributed election algorithm (e.g., a weighted score based on communication signal strength and available computing power), identifying the vehicle with the highest score as the lead vehicle and assigning it a unique leader identifier. This lead vehicle is responsible for subsequent global map fusion and negotiation initiation, while the remaining vehicles act as follower vehicles, maintaining real-time communication with the lead vehicle. In this way, the fleet autonomously forms a communication network with a unified command node without the need for external infrastructure, laying the organizational foundation for subsequent collaborative perception and decision-making.

[0027] S20: Obtain obstacle recognition information from each multimodal sensor data, obtain passage area information and obstacle information from the obstacle recognition information, and generate an obstacle probability distribution map of the heavy truck fleet based on the passage area information and obstacle information.

[0028] In this embodiment, obstacle recognition information refers to the various objects and their attributes in the road ahead identified by each unmanned heavy truck through its own sensors. Passage area information refers to the area of ​​road where vehicles can safely travel, and obstacle information refers to the position, size, and motion state of static or dynamic objects that may obstruct passage. The obstacle probability distribution map is a two-dimensional grid map generated by fusing multi-vehicle perception data, representing the probability of an obstacle existing at each location within the surrounding space of the entire convoy.

[0029] Specifically, the onboard computing unit of each unmanned heavy truck first performs ground segmentation and clustering on the LiDAR point cloud data to identify obstacles such as vehicles, cones, and debris ahead, and marks their outlines and location coordinates. Simultaneously, using semantic segmentation results from visual images, it distinguishes between drivable road surfaces and impassable areas such as shoulders and guardrails. Subsequently, each vehicle converts this information into a local occupancy grid map, where each grid cell stores a probability value between 0 and 1, representing the likelihood that the location is occupied by an obstacle. These local grid maps are uploaded to the lead vehicle via a previously established communication link. After receiving the local grid maps from all following vehicles, the lead vehicle uses a weighted fusion algorithm to superimpose the grid probability values ​​from different vehicles, assigning higher confidence weights to areas observed simultaneously by multiple vehicles, ultimately generating a global obstacle probability distribution map covering hundreds of meters in front of and behind the entire convoy. This map provides a unified environmental awareness foundation for subsequent collaborative path planning, enabling each vehicle to understand the obstacle situation outside its own sensor blind spots.

[0030] S30: When unavoidable obstacles are obtained from the obstacle probability distribution map, obstacle avoidance route data for each unmanned heavy truck is generated using the information of the leader vehicle.

[0031] In this embodiment, unavoidable obstacles refer to obstacles that, based on the current convoy speed and road conditions, cannot be avoided by simple deceleration or slight steering, and require significant detours or convoy reorganization. Obstacle avoidance route data refers to the safe driving trajectory from the current location to the target location planned for each unmanned heavy truck, including a sequence of waypoints, desired speed, and steering angle.

[0032] Specifically, the lead vehicle continuously monitors the global obstacle probability distribution map. When it detects an obstacle ahead with a probability value exceeding a preset threshold (e.g., 0.8) and occupying more than a certain proportion of the lane width, it determines it to be an unavoidable obstacle. At this point, the lead vehicle generates several feasible obstacle avoidance schemes based on the current formation's geometry (vehicle spacing, arrangement), each vehicle's dynamic parameters (maximum steering angle, braking deceleration), and road width. These schemes include shifting the entire formation one lane to the left, having the vehicles in front and behind stagger their positions to bypass the obstacle, or splitting the formation into two groups to pass through separately and then rejoin. Each scheme includes the target position and speed command for each vehicle at every moment. The lead vehicle sends these candidate schemes to all following vehicles via a communication link. Each following vehicle independently evaluates each scheme based on its real-time status (current speed, load, remaining braking force), selects the scheme it considers safest, and provides a score. After collecting all scores, the lead vehicle conducts multiple rounds of voting using a distributed consensus protocol to finally determine a scheme that all vehicles agree on, and then distributes the specific path point sequence corresponding to this scheme to each vehicle. After receiving their respective obstacle avoidance routes, each vehicle begins to execute the corresponding steering, acceleration, or deceleration actions.

[0033] S40: Obtain the actual obstacle avoidance data corresponding to the obstacle avoidance route data, score the obstacle avoidance based on the actual obstacle avoidance data, and optimize the strategy for generating obstacle avoidance routes based on the obstacle avoidance score.

[0034] In this embodiment, actual obstacle avoidance data refers to the measured data recorded by each vehicle during the execution of the obstacle avoidance route, including the deviation between the actual driving trajectory and the planned trajectory, changes in the vehicle's longitudinal and lateral acceleration, and the additional power consumption of the engine or motor. The obstacle avoidance score is a quantitative evaluation of the obstacle avoidance performance based on multiple indicators. The strategy for optimizing the obstacle avoidance route generation refers to adjusting the parameters in the path planning algorithm using machine learning methods to make future obstacle avoidance schemes more reasonable and efficient.

[0035] Specifically, after the entire convoy completes an obstacle avoidance maneuver, each vehicle automatically records key data from the process, including formation keeping deviation (the difference between the actual distance between vehicles and the target distance), braking impact (the rate of change of acceleration during braking), and increased mileage and energy consumption due to detours. This data is aggregated to the lead vehicle, which calculates the overall score for the obstacle avoidance operation using a preset weighted formula. If the score is below a set threshold, it indicates room for improvement in the obstacle avoidance strategy. The system then triggers an online fine-tuning program, using near-end strategy optimization or a soft actor-critic algorithm to slightly update the weight parameters in the path planning strategy, enabling the selection of a better solution for similar scenarios in the future. If the score is above or equal to the threshold, the candidate solution and its environmental characteristics are stored in an excellent case library for direct use in subsequent similar scenarios. Through this closed-loop feedback mechanism, the convoy's obstacle avoidance strategy can continuously evolve with increasing mileage, gradually adapting to the optimal operating mode under different road conditions and weather conditions.

[0036] In this embodiment, by acquiring multimodal sensor data from each unmanned heavy-duty truck and constructing a fleet communication link, a distributed election algorithm is used to generate leader vehicle information, achieving real-time data communication among vehicles within the fleet and establishing a unified command node. Based on this, obstacle recognition information is extracted from the multimodal sensor data to generate passage area and obstacle information, thereby constructing an obstacle probability distribution map of the heavy-duty truck fleet. This allows each vehicle to gain a global environmental awareness, no longer limited to the detection range of its own sensors, but rather by integrating multi-vehicle perception information. When an unavoidable obstacle is detected, obstacle avoidance route data is generated for each vehicle using the leader vehicle information, ensuring that the entire fleet can take coordinated obstacle avoidance actions when facing sudden obstacles, reducing formation chaos and collision risks caused by independent vehicle decisions. Finally, by acquiring actual obstacle avoidance data for scoring and optimizing the obstacle avoidance strategy, the fleet's obstacle avoidance capability can continuously improve with accumulated operational experience, enhancing the method's adaptability to different scenarios. Overall, this method improves the collaborative obstacle avoidance capability of unmanned heavy-duty truck fleets in complex dynamic environments and reduces safety hazards during convoy driving.

[0037] In one embodiment, reference is made to Figure 2 In step S10, namely, constructing the fleet communication link based on multimodal sensor data and generating leader vehicle information in the heavy truck fleet using a distributed election algorithm, the specific steps include: S11: Obtain communication signal strength data and computing resource data for each unmanned heavy truck from multimodal sensing data.

[0038] In this embodiment, communication signal strength data refers to the received signal strength indication value measured by the C-V2X communication module of each unmanned heavy truck when establishing a connection with surrounding vehicles, and the communication link quality and signal-to-noise ratio parameters calculated from it. Computing resource data refers to the current load status of the on-board computing unit of each vehicle, including the real-time utilization rate of the central processing unit, the usage of the graphics processing unit, the remaining memory capacity, and the storage read / write speed.

[0039] Specifically, during the process of establishing communication links within the fleet, each unmanned heavy-duty truck, in addition to broadcasting its own identity, actively measures the communication signal strength between itself and other vehicles, recording the average value and fluctuation range of multiple measurements. Simultaneously, the onboard diagnostic system monitors the load of the computing unit in real time, generating a computing resource report that includes indicators such as CPU utilization percentage, available memory size (in MB), and the number of currently running background tasks. This data is sent via the communication link to the election initiator (usually a temporary coordinating vehicle designated at the initial startup) for subsequent lead vehicle scoring. By acquiring this basic data, the system can comprehensively understand the differences in communication and computing capabilities of each vehicle, providing an objective basis for the scientific selection of the lead vehicle.

[0040] S12: A comprehensive score is obtained for each unmanned heavy truck based on communication signal strength data and computing resource data to determine its navigation compatibility score.

[0041] In this embodiment, the navigation suitability score is a dimensionless numerical value used to quantify the suitability of each unmanned heavy truck to serve as a navigation vehicle. This score comprehensively considers both communication quality and computing resources; a higher score indicates that the vehicle is more suitable for the navigation task.

[0042] Specifically, the system first normalizes the communication signal strength data, mapping it to a score range of 0 to 100, with higher signal strength resulting in a higher score. Similarly, the computing resource data is also normalized; for example, lower CPU utilization and more remaining memory result in a higher score. Then, the two scores are weighted and summed according to preset weight ratios (e.g., communication quality weight 0.6, computing resource weight 0.4) to obtain the navigation suitability score for each vehicle. The weighting can be flexibly adjusted according to the actual application scenario: in highway scenarios with poor communication environments, the weight of communication quality can be increased; while in map fusion scenarios requiring extensive computation, the weight of computing resources can be increased. This comprehensive scoring method avoids the one-sidedness that may result from relying solely on a single indicator, ensuring that the selected navigation vehicle can both guarantee stable communication and handle complex computational tasks.

[0043] S13: The unmanned heavy truck with the highest navigation compatibility score will be identified as the leader vehicle.

[0044] In this embodiment, the leader vehicle information refers to the unique identifier, IP address, communication port, and current position number within the convoy of the officially confirmed leader vehicle. This information will be broadcast to all vehicles in the convoy so that all subsequent communications will revolve around this vehicle.

[0045] Specifically, after calculating the navigation compatibility score for all vehicles, the system sorts the scores in descending order and marks the vehicle with the highest score as the lead vehicle. If multiple vehicles have the same score, the system further compares the standard deviation of their communication signal strength and selects the vehicle with the most stable signal. Once the lead vehicle is determined, it immediately sends an acknowledgment message to the entire convoy, containing its own identity information and timestamp. Upon receiving the acknowledgment message, other vehicles update their communication target addresses to the lead vehicle's address and cease participating in the next round of election, thus forming a star-shaped communication topology centered on the lead vehicle. The lead vehicle then begins to perform its duties, including receiving local perception data from other vehicles, performing map fusion, and initiating obstacle avoidance negotiations. If the lead vehicle loses contact due to malfunction or leaving the convoy, the system will automatically trigger a new round of election to reselect a lead vehicle, ensuring the continuity of convoy control.

[0046] In this embodiment, communication signal strength and computing resource data of each unmanned heavy truck are obtained from multimodal sensor data, and a comprehensive score is calculated to obtain a navigation suitability score. The vehicle with the highest score is identified as the lead vehicle, realizing dynamic selection of the lead vehicle. This method makes the selection of the lead vehicle no longer dependent on fixed assignment or simple rotation, but comprehensively considers the communication quality (signal strength) and computing power (computing resources) of each vehicle, thereby ensuring that the lead vehicle has optimal data transmission and reception capabilities and the ability to process global information in real time. Compared with methods that randomly assign or select lead vehicles according to number order, this scheme improves the stability of the lead vehicle on the communication link, reduces the occurrence of information loss or decision delays caused by excessive communication latency or insufficient computing power of the lead vehicle, and provides a reliable command node foundation for subsequent map fusion and collaborative path planning.

[0047] In one embodiment, reference is made to Figure 3 In step S12, a comprehensive score is calculated based on communication signal strength data and computing resource data to obtain the navigation compatibility score for each unmanned heavy truck. This specifically includes: S121: Obtain communication delay data and packet loss rate data from communication signal strength data, and obtain central processing unit occupancy rate and remaining memory capacity from computing power resource data.

[0048] In this embodiment, communication latency data refers to the time elapsed from the sending to the receiving of a data packet, typically measured in milliseconds. Packet loss rate data refers to the percentage of data packets lost within a certain period relative to the total number of data packets sent. CPU utilization refers to the percentage of CPU currently in use, and remaining memory capacity refers to the unused random access memory capacity.

[0049] Specifically, each unmanned heavy-duty truck sends timestamped probe frames during communication tests with other vehicles. The receivers reply immediately upon receipt, and the sender estimates the one-way latency based on half the round-trip time. One hundred probe frames are sent consecutively, and the number of successfully received frames is counted to calculate the packet loss rate. Simultaneously, the onboard operating system obtains the CPU's time allocation in user mode, kernel mode, and idle mode by reading system files such as ` / proc / stat`, calculating the current CPU utilization. Remaining memory capacity is obtained by reading the `MemFree` field in ` / proc / meminfo`. This data is encapsulated into standard-format data packets and periodically reported to the election initiator. Through these granular metrics, the system can gain a deeper understanding of each vehicle's real-time performance on the communication link and the level of computational resource strain, providing more granular input for subsequent scoring.

[0050] S122: Generate a communication quality score based on communication delay data and packet loss rate data, and generate a computing power availability score based on CPU utilization and remaining memory capacity.

[0051] In this embodiment, the communication quality score is a quantitative evaluation of the reliability of vehicle communication; the lower the communication latency and the lower the packet loss rate, the higher the score. The computing power availability score is a quantitative evaluation of the vehicle's remaining computing power; the lower the CPU utilization and the more memory remaining, the higher the score.

[0052] Specifically, when generating a communication quality score, communication latency data is first mapped to scoring ranges: less than 10 milliseconds earns 100 points, 10 to 20 milliseconds earns 90 points, 20 to 50 milliseconds earns 70 points, and greater than 50 milliseconds earns 50 points. Packet loss rate data is similarly mapped: less than 0.1% earns 100 points, 0.1% to 0.5% earns 85 points, 0.5% to 1% earns 60 points, and greater than 1% earns 40 points. Then, the latency score and packet loss rate score are weighted with weights of 0.4 and 0.6 respectively to obtain the overall communication quality score. When generating a computing power availability score, CPU utilization below 30% earns 100 points, 30% to 50% earns 80 points, 50% to 70% earns 60 points, and above 70% earns 40 points; remaining memory capacity greater than 4GB earns 100 points, 2GB to 4GB earns 80 points, 1GB to 2GB earns 60 points, and less than 1GB earns 40 points. The CPU score and memory score are weighted by 0.5 each to obtain the available computing power score. This tiered scoring method allows vehicles with different performance levels to receive differentiated scores, facilitating subsequent selection of the best option.

[0053] S123: The communication quality score and the computing power availability score are weighted and summed to obtain the navigation adaptability score.

[0054] In this embodiment, the navigation compatibility score is a weighted combination of the communication quality score and the computing power availability score, with the weights dynamically adjusted according to the application scenario. This score comprehensively reflects the degree to which a vehicle simultaneously possesses good communication conditions and sufficient computing resources.

[0055] Specifically, the system defaults to a communication quality score weight of 0.5 and a computing power availability score weight of 0.5, indicating that both are equally important. However, in special scenarios, such as when the convoy is about to enter tunnels or mountainous areas where communication may be obstructed, the system can automatically increase the communication quality score weight to 0.7 to ensure that the selected lead vehicle has the strongest anti-interference capability. Conversely, if large-scale map fusion or complex path planning is about to be performed, the computing power availability score weight can be increased to 0.7. The weighted summation formula is: Lead Vehicle Adaptability Score = Communication Quality Score × Communication Weight + Computing Power Availability Score × Computing Power Weight. The calculated score ranges from 0 to 100. The system ranks all vehicles by score and selects the highest-scoring vehicle as the lead vehicle. This flexible weighting mechanism allows the selection of the lead vehicle to adapt to different operating environments and task requirements, improving the adaptability and robustness of convoy control.

[0056] In this embodiment, by further extracting communication delay and packet loss rate data from communication signal strength data, and extracting CPU utilization and remaining memory capacity from computing resource data, a communication quality score and a computing power availability score are generated respectively. These scores are then weighted and summed to obtain the navigation suitability score, achieving a multi-dimensional and refined evaluation of candidate navigation vehicles. Communication delay and packet loss rate data directly reflect the real-time communication reliability of the vehicle in the current network environment, while CPU utilization and remaining memory capacity reflect the vehicle's remaining computing power available for handling complex computational tasks. Weighting and summing these two types of scores ensures that the final navigation suitability score balances the real-time requirements of the communication link with the sufficiency of computing resources. Compared to relying solely on a single indicator (such as signal strength) for selection, this scheme improves the scientific rigor and accuracy of navigation vehicle selection, reduces the occurrence of navigation vehicles failing to respond or process tasks in a timely manner at critical moments due to poor communication quality or insufficient computing power, and further strengthens the underlying support for fleet collaborative control.

[0057] In one embodiment, reference is made to Figure 4 In step S20, obstacle recognition information is obtained from each multimodal sensor data, and passage area information and obstacle information are obtained from the obstacle recognition information. An obstacle probability distribution map of the heavy truck fleet is generated based on the passage area information and obstacle information. Specifically, this includes: S21: Obtain a local occupancy raster map from each multimodal sensing data, and extract the raster occupancy probability value from the local occupancy raster map.

[0058] In this embodiment, the local occupancy grid map refers to a two-dimensional grid map of a certain range centered on the vehicle, generated by each unmanned heavy truck based on its own sensor data. Each grid cell stores the probability value that the location is occupied by an obstacle. The grid occupancy probability value is a floating-point number between 0 and 1, where 0 represents completely free and 1 represents completely occupied.

[0059] Specifically, the onboard computer of each unmanned heavy-duty truck first projects the point cloud data from the LiDAR onto a two-dimensional plane in the vehicle coordinate system, dividing it into a grid according to a preset grid resolution (e.g., 0.2m × 0.2m). For each grid, the number of point clouds falling within that grid is counted, and combined with radar reflection intensity information, the posterior probability of that grid being occupied by an obstacle is calculated using a Bayesian update formula. Simultaneously, the Doppler effect of millimeter-wave radar is used to detect moving objects, increasing the probability value of dynamic obstacles. Semantic segmentation results from the visual image are also used to correct the probability values; for example, the probability of an area identified as road surface is set to 0.01, and the probability of an area identified as a vehicle is set to 0.95. Finally, a 200m × 200m local occupancy grid map is generated. Each vehicle retains only the core data of this map, namely the coordinates and corresponding probability values ​​of each grid, discarding the original point cloud data to reduce transmission volume. These probability values ​​constitute the smallest data unit for subsequent fusion.

[0060] S22: Send the grid occupancy probability value to the lead vehicle corresponding to the leader vehicle information. The lead vehicle will then perform a weighted fusion of all grid occupancy probability values ​​to obtain the global grid occupancy probability value.

[0061] In this embodiment, weighted fusion refers to the lead vehicle combining grid probability values ​​from different vehicles at the same geographical location according to certain weights to generate a more accurate comprehensive probability value. The global grid occupancy probability value refers to the probability value of each grid in the unified grid map that covers a larger area in all directions of the entire convoy after fusion.

[0062] Specifically, after generating a local occupancy grid map, each following vehicle immediately sends a grid occupancy probability value data packet to the lead vehicle via a C-V2X communication link. The data packet contains the vehicle's own location coordinates (obtained via high-precision GPS), a timestamp, and the latitude / longitude or UTM coordinates and probability value for each grid cell. Upon receiving data from all following vehicles, the lead vehicle first performs time alignment, discarding older data with a delay exceeding 100 milliseconds. Then, for each geographic grid cell, it collects the probability values ​​of all vehicles covering that grid cell and performs a weighted average based on each vehicle's sensor confidence. Sensor confidence is dynamically calculated based on the vehicle's historical detection accuracy, such as the percentage of correctly identified obstacles in the last 100 detections. Vehicles with higher confidence contribute more weight. Furthermore, the lead vehicle itself also has a local grid map, which is also involved in the fusion process. Finally, for each grid cell, a weighted average global grid occupancy probability value is calculated, forming a complete global obstacle probability distribution map. The map typically covers 300 meters in front of the convoy, 100 meters behind it, and 50 meters to the left and right, providing comprehensive environmental information for subsequent route planning.

[0063] S23: Generate an obstacle probability distribution map based on the global grid occupancy probability value.

[0064] In this embodiment, the obstacle probability distribution map refers to the final map product that, after fusion and preliminary processing, can be directly used for path planning decisions. The map is presented in grid form, where the value of each grid cell represents the probability of an obstacle existing at that location, and is accompanied by obstacle classification labels (such as static obstacles, dynamic obstacles, and unknown areas).

[0065] Specifically, after obtaining the global grid occupancy probability values, the lead vehicle needs to perform a post-processing step to generate the final obstacle probability distribution map. First, the probability values ​​are thresholded: grids with a probability value greater than 0.6 are marked as obstacle regions, grids with a probability value less than 0.2 are marked as free regions, and grids between 0.2 and 0.6 are marked as unknown regions. Second, connected component analysis is performed on the obstacle regions, aggregating adjacent obstacle grids into independent obstacle objects, assigning each object a unique ID, and recording its center position, bounding box size, and average probability value. Then, Kalman filtering is used to track each dynamic obstacle object, correlating its positional changes over multiple frames to estimate its speed and direction of motion. Finally, this information is integrated into a comprehensive map containing an obstacle list and a grid probability map, i.e., the obstacle probability distribution map. This map is stored in the lead vehicle's memory in a standard format for subsequent obstacle avoidance path planning modules to access. Through this series of processes, the raw multi-vehicle perception data is transformed into a structured, decision-oriented environmental representation.

[0066] In this embodiment, by acquiring local occupancy grid maps and extracting grid occupancy probability values ​​from each multimodal sensing data source, these probability values ​​are sent to the lead vehicle for weighted fusion to obtain a global grid occupancy probability value, thereby generating an obstacle probability distribution map. This achieves lightweight fusion of multi-vehicle perception data. Each vehicle only transmits compressed grid occupancy probability values ​​instead of raw point cloud data, significantly reducing communication bandwidth consumption and enabling high-frequency perception information sharing even under limited V2X channel conditions. The lead vehicle, through weighted fusion, integrates the observation results of multiple vehicles on the same area. Obstacles detected simultaneously by multiple vehicles receive higher confidence, while isolated signals detected by a single vehicle are weakened, thus reducing the impact of sensor false alarms and missed alarms. This scheme improves the detection coverage and confidence of the convoy for environmental obstacles, allowing following vehicles to obtain obstacle information in blind spots in advance with the perception data of the vehicles in front, reducing the occurrence of untimely obstacle avoidance due to limited field of vision of a single vehicle.

[0067] In one embodiment, reference is made to Figure 5 In step S23, which involves generating an obstacle probability distribution map based on the global grid occupancy probability value, the specific steps include: S231: Obtain the dynamic grid region corresponding to the dynamic obstacle from the global grid occupancy probability value.

[0068] In this embodiment, the dynamic grid region refers to the set of grids in the global obstacle probability distribution map that have high probability values ​​and whose positions change over time, typically corresponding to moving vehicles, pedestrians, or other moving objects. These regions are characterized by grid coordinate displacement between adjacent frames.

[0069] Specifically, the lead vehicle maintains a global obstacle probability distribution map while also saving historical map data from the most recent frames (e.g., the last 5 frames). For each obstacle grid in the current frame, the system matches it with nearby grids at the same coordinates in the previous frame, checking for grids with similar probability values. If such grids exist, and their coordinates have shifted, the grid is considered a dynamic obstacle. The system further calculates the offset vector and groups all adjacent grids with the same motion trend into a single dynamic grid region. For example, a grid region corresponding to a vehicle traveling at 80 km / h will move forward approximately 2.2 meters between frames. In this way, the system can separate the static background from the dynamic foreground in the global map, providing a clear target for subsequent trajectory prediction. Information about the dynamic grid region includes the region boundary, current center coordinates, direction of motion, and speed estimate.

[0070] S232: Obtain temporal change data of the dynamic grid area, predict the motion trajectory of the dynamic obstacle based on the temporal change data, and obtain the predicted trajectory data.

[0071] In this embodiment, temporal change data refers to the position sequence of the dynamic grid region over the past few frames, recording the movement history of the obstacle. Predicted trajectory data refers to the position sequence that the obstacle may pass through in the future based on historical movement patterns.

[0072] Specifically, the system extracts the center coordinates of each dynamic grid region over the past 5 frames, forming a time series. Assuming a sampling frequency of 10 Hz, the time span is 0.5 seconds. These historical positions are fitted using either a constant velocity model or a constant acceleration model to calculate the current velocity and acceleration vectors. For example, if the positions in the past five frames are linearly arranged, a uniform velocity model is used; if the positions exhibit a parabolic trend, a uniform acceleration model is used. Based on the fitted motion parameters, the system extrapolates the positions for the next 2 seconds, outputting a predicted point every 0.2 seconds, forming a predicted trajectory consisting of 10 points. For vehicles that may change lanes, the system can also combine road curvature and lane line information to predict their lane-changing intentions, generating multiple possible trajectory branches. This predicted trajectory data is appended to the attributes of dynamic obstacles for subsequent obstacle avoidance scheme evaluation. By predicting the future positions of dynamic obstacles in advance, the system can make more proactive decisions, avoiding last-minute emergency avoidance.

[0073] S233: Based on the predicted trajectory data, the global grid occupancy probability value is spatiotemporally filtered to obtain the filtered obstacle probability distribution map.

[0074] In this embodiment, spatiotemporal filtering refers to using the predicted trajectories of dynamic obstacles to correct the global grid occupancy probability value, making it reflect the environmental state at future moments, while filtering out random errors caused by sensor noise. The filtered obstacle probability distribution map is a time-varying probability field, where the probability value of each grid changes over time.

[0075] Specifically, the system first performs spatial filtering on the global grid occupancy probability values, smoothing the probability values ​​using Gaussian kernel convolution to eliminate isolated noise. Then, it performs temporal filtering: for each grid, it performs an exponential moving average of the probability value of the current frame and the probability value of the previous frame, using the formula: New probability value = 0.7 × Current observation value + 0.3 × Previous frame filtered value. This suppresses flicker noise. The most crucial step is forward filtering using predicted trajectory data: for each dynamic obstacle, based on its predicted trajectory, the probability value of the grid area that the obstacle is expected to occupy in the future is set to a high value (e.g., 0.8), while the probability value of the grid area that the obstacle currently occupies but will leave in the future is gradually reduced. For example, if it is predicted that the vehicle ahead will change lanes to the right in 1 second, the system increases the probability value of the corresponding grid in the adjacent lane on the right in the current map, alerting the path planning module to avoid it in advance. After spatiotemporal filtering, the obstacle probability distribution map becomes smoother and more continuous, and contains predictive information about future situations, providing higher-quality environmental input for subsequent path planning.

[0076] In this embodiment, dynamic grid regions corresponding to dynamic obstacles are obtained from the global grid occupancy probability values. Their temporal change data is extracted for trajectory prediction, and the prediction results are used to perform spatiotemporal filtering on the global grid occupancy probability values ​​to obtain a filtered obstacle probability distribution map. Dynamic obstacles, such as vehicles cutting into adjacent lanes or pedestrians crossing, change positions rapidly over time. Without processing, the original static probability map would lead to incorrect path planning decisions based on outdated information. By analyzing the temporal changes of dynamic grid regions and predicting their future movement trends, potential danger zones can be identified in advance. Spatiotemporal filtering further eliminates sensor noise and false signals from brief flickering, resulting in a smoother and more stable output probability distribution map. This approach improves the adaptability of the obstacle probability distribution map to dynamic environments, enabling subsequent path planning to make decisions based on predictions of future states. This reduces the occurrence of obstacle avoidance scheme failures due to sudden movement of dynamic obstacles, enhancing the smoothness of convoy traffic in complex traffic flows.

[0077] In one embodiment, reference is made to Figure 6 In step S30, when unavoidable obstacles are obtained from the obstacle probability distribution map, obstacle avoidance route data for each unmanned heavy truck is generated using the leader vehicle information. Specifically, this includes: S31: When an unavoidable obstacle is obtained from the obstacle probability distribution map, multiple candidate obstacle avoidance schemes are generated based on the lead vehicle corresponding to the leader vehicle information.

[0078] In this embodiment, candidate obstacle avoidance schemes refer to multiple feasible obstacle avoidance strategies proposed by the lead vehicle for currently detected unavoidable obstacles, taking into account formation configuration, road conditions, and vehicle dynamics constraints. Each scheme includes detailed execution steps, including the target lane, target speed, steering timing, and formation adjustment method for each vehicle.

[0079] Specifically, the lead vehicle continuously monitors the filtered obstacle probability distribution map. When it detects an obstacle with a probability value exceeding 0.8 that occupies more than 70% of the lane width, and the braking distance at the current speed is insufficient to stop safely, it is determined to be an unavoidable obstacle. Subsequently, the lead vehicle activates the path planning module, generating at least three different candidate obstacle avoidance schemes based on the number of vehicles in the current platoon, their spacing, arrangement (straight line or staggered), and the available width and curvature of the road. Scheme 1 is a complete left-side detour: all vehicles simultaneously change lanes to the left, maintaining the platoon formation, and return to their original lanes after bypassing the obstacle. Scheme 2 is a segmented staggered passage: the first half of the vehicles first change lanes to the left, while the second half continues straight and passes the obstacle from the right, then rejoins behind the obstacle. Scheme 3 is a platoon split and regrouping: the lead vehicle leads the first half of the vehicles to change lanes to the left, the second half slows down and increases the distance, and after the first half passes, the second half changes lanes to the left again, finally regrouping. Each proposal comes with safety assessment data, including minimum safe distance, maximum lateral acceleration, estimated travel time, and additional energy consumption. These proposals are packaged in a standardized format and are ready to be sent to all following vehicles for evaluation.

[0080] S32: Send multiple candidate obstacle avoidance solutions to each unmanned heavy truck and obtain the score data of each unmanned heavy truck for each candidate obstacle avoidance solution.

[0081] In this embodiment, the scoring data refers to the numerical score given by each following vehicle after receiving candidate obstacle avoidance solutions, based on its current state and dynamic constraints, after independently evaluating the safety, feasibility, and comfort of each solution. The scoring is usually divided into multiple dimensions and finally summarized into a total score.

[0082] Specifically, the lead vehicle broadcasts candidate obstacle avoidance schemes to all following vehicles via a C-V2X communication link. Upon receiving a scheme, each following vehicle immediately activates its local safety assessment module. This module first reads real-time parameters such as its current speed, load, tire adhesion coefficient, and braking system status. For each candidate scheme, it simulates the steering angle, braking deceleration, and acceleration changes required to execute the scheme and compares them to the vehicle's physical limits. For example, if a scheme requires a vehicle to change lanes with 0.5g lateral acceleration, and the vehicle's current load is too high, resulting in insufficient lateral stability margin, the scheme's safety score will be low. The assessment module also considers the scheme's impact on formation keeping, such as whether it disrupts spacing and increases the risk of rear-end collisions. Finally, each vehicle outputs a comprehensive score from 0 to 100 for each scheme, attaches its vehicle ID and timestamp to the score data, and replies to the lead vehicle. This process typically completes within tens of milliseconds, ensuring real-time negotiation.

[0083] S33: Based on the scoring data, voting is conducted through a distributed consensus protocol. When all unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the consensus-reaching candidate obstacle avoidance scheme is determined as the obstacle avoidance route data.

[0084] In this embodiment, the distributed consensus protocol refers to an algorithmic mechanism that ensures all participating vehicles reach a consensus on the final decision, similar to a consensus algorithm in a distributed system. The obstacle avoidance route data refers to the specific path point sequence and control commands for each vehicle corresponding to the final solution accepted by all vehicles.

[0085] Specifically, after the lead vehicle collects the scoring data from all following vehicles, it initiates a consensus process based on an improved Raft algorithm. First, the lead vehicle counts the highest number of scores received for each candidate solution (i.e., how many vehicles rated the solution as the best). If a solution receives more than half of the highest scores, it is initially designated as the winning solution. However, to prevent the influence of malicious or abnormal vehicles, the protocol requires all vehicles to re-confirm the winning solution: the lead vehicle rebroadcasts the winning solution, and each vehicle checks whether the solution meets its own minimum safety threshold. If it does, it replies "agree"; otherwise, it replies "reject" with a reason. If all vehicles agree, a consensus is reached, and the lead vehicle officially designates the solution as the final obstacle avoidance route data. If any vehicle rejects the solution, the lead vehicle excludes it and selects the second-best solution from the remaining options for re-confirmation until all vehicles agree. After consensus is reached, the lead vehicle decomposes the obstacle avoidance route data into independent trajectory instructions for each vehicle, including the target position, speed, and steering angle at each time point, and sends these instructions to each vehicle via a communication link. Upon receiving the instructions, each vehicle immediately begins execution, thus achieving coordinated obstacle avoidance across the entire convoy.

[0086] In this embodiment, when an unavoidable obstacle is detected, the lead vehicle generates multiple candidate obstacle avoidance schemes and sends them to each unmanned heavy truck. Each vehicle's rating of each scheme is collected, and a distributed consensus protocol is used for voting. Once all vehicles reach a consensus on the same scheme, it is confirmed as the obstacle avoidance route data, achieving collaborative decision-making at the platoon level. This mechanism breaks the traditional model where the lead vehicle unilaterally sets the route and the following vehicles passively execute it. It allows each vehicle to independently evaluate candidate schemes based on its own status (such as current speed, load, and braking performance) and express its opinion through voting. The distributed consensus protocol ensures that all vehicles eventually reach a consensus on the same scheme, thus avoiding platoon fragmentation or collisions caused by different obstacle avoidance paths chosen by the leading and following vehicles. This scheme improves the consistency of platoon obstacle avoidance actions, enabling the entire convoy to move in unison when detouring, slowing down, or changing lanes, reducing the risk of chain accidents caused by decision-making disagreements, and also improving the platoon's passage efficiency in complex scenarios.

[0087] In one embodiment, reference is made to Figure 7 In step S33, voting is conducted based on the scoring data using a distributed consensus protocol. When all unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the consensus-reaching candidate obstacle avoidance scheme is determined as the obstacle avoidance route data. Specifically, this includes: S331: Obtain the maximum steering angle, maximum braking deceleration, and current load data from the dynamic parameters of each unmanned heavy truck.

[0088] In this embodiment, the maximum steering angle refers to the maximum deflection angle that the vehicle's front wheels can achieve, which is usually determined by the mechanical limit of the steering mechanism, and is measured in degrees. The maximum braking deceleration refers to the maximum safe deceleration that the vehicle can produce under full load, and is measured in meters per square second. The current load data refers to the total weight of the cargo currently loaded on the vehicle, measured in tons, and can be obtained through an onboard weighing sensor or a suspension height sensor.

[0089] Specifically, due to the inherent characteristics of unmanned heavy trucks—high mass, large inertia, long braking distance, and high center of gravity—there is a risk of roll during steering. Furthermore, the varying loads of vehicles in a platoon lead to significant differences in maximum steering angle, maximum braking deceleration, and lateral stability margin for each vehicle. Therefore, each unmanned heavy truck's onboard control system performs a self-check upon startup, reading the current maximum mechanical limit via a steering angle sensor and calculating the maximum braking deceleration using brake pressure sensors and wheel speeds. Current load data is measured in real-time by the air suspension's pressure sensors and converted into a weight value. These dynamic parameters are stored in a local vehicle status database and updated periodically at a frequency of once per second. When obstacle avoidance scheme evaluation is required, the vehicle directly reads the latest parameter values ​​from the database, ensuring the assessment is based on the current actual physical capabilities. By acquiring these parameters, the system can accurately determine the safety margin of each vehicle when performing specific obstacle avoidance maneuvers, avoiding requiring vehicles to perform actions beyond their physical limits.

[0090] S332: Based on the maximum steering angle, maximum braking deceleration, and current load data, a safety feasibility score is given for each candidate obstacle avoidance scheme to obtain the score data.

[0091] In this embodiment, the safety feasibility score is a quantitative index calculated for each candidate solution based on the vehicle's dynamic parameters. It is used to measure the degree of safety risk that the solution poses to the vehicle. A higher score indicates that the solution is safer and more feasible.

[0092] Specifically, for each candidate solution, the vehicle first extracts the maximum steering angle and maximum braking deceleration required by the solution. For example, if the solution requires a left lane change, a steering angle of 15 degrees and a deceleration of 0.3g are needed. The vehicle compares the required steering angle with its maximum steering angle. If the required value exceeds 80% of the maximum value, points are deducted; if it exceeds the maximum value, it is directly deemed infeasible, and the score is set to zero. Similarly, the required braking deceleration value is compared with the maximum value; if it exceeds 80%, points are deducted. Furthermore, the impact of the current load on vehicle stability is considered: due to the high center of gravity of heavy trucks, the greater the load, the higher the risk of rollover from the same steering and braking actions. The system obtains a risk coefficient from a table based on the load data and multiplies it by the base score. The final safety feasibility score is calculated as follows: Score = 100 × (1 - Steering Margin Penalty) × (1 - Braking Margin Penalty) × Load Risk Coefficient. The steering margin penalty and braking margin penalty are calculated linearly based on the ratio of the required value to the maximum value. Through this quantitative assessment, each vehicle can objectively determine whether it can safely execute each candidate solution, providing a basis for subsequent voting.

[0093] S333: The scoring data is sent to the lead vehicle, which counts the highest number of scores obtained by each candidate obstacle avoidance scheme and determines the candidate obstacle avoidance scheme with the most highest scores as the obstacle avoidance route data that has reached a consensus.

[0094] In this embodiment, the highest number of scores refers to the number of unmanned heavy trucks that rated a particular candidate solution as the best among all solutions. The lead vehicle selects the solution with the most first-place rankings as the final consensus solution by counting the number of times each solution achieved first place.

[0095] Specifically, after each following vehicle completes its safety feasibility assessment of each candidate solution, it identifies the solution with the highest score and sends its solution number along with its score to the lead vehicle. The lead vehicle collects the first-place votes from all vehicles and tallies the votes for each solution. For example, if solution A receives 5 votes, solution B receives 3 votes, and solution C receives 2 votes, then solution A is the preliminary winner. To prevent ties, the lead vehicle also collects the total score and average score for each solution as backup indicators. If two solutions have the same number of votes, their average scores are compared, and the solution with the higher average score is selected. After determining the winning solution, the lead vehicle uses it as the final obstacle avoidance route data and performs a second confirmation according to the consensus confirmation process in claim 6. If all vehicles agree, it officially takes effect; if any vehicle objects, the solution is excluded, and the first-place votes are recounted from the remaining solutions until a consensus is reached. Through this voting-based mechanism, the final obstacle avoidance solution represents the safety preferences of the majority of vehicles while also considering the special circumstances of a minority of vehicles, achieving fair and efficient collaborative decision-making.

[0096] In this embodiment, the maximum steering angle, maximum braking deceleration, and current load data are obtained from the dynamic parameters of each unmanned heavy truck. Based on this, the safety feasibility of each candidate obstacle avoidance scheme is scored. Then, the lead vehicle counts the highest number of scores obtained for each scheme, and the scheme with the most highest scores is determined as the consensus obstacle avoidance route data. This achieves differentiated safety assessment and consensus decision-making based on the physical constraints of heavy trucks. Due to the inherent characteristics of unmanned heavy trucks, such as large mass, high inertia, long braking distance, and high center of gravity, there is a risk of rollover during steering. At the same time, the different loads of each vehicle in the formation lead to significant differences in the maximum steering angle, maximum braking deceleration, and lateral stability margin of each vehicle. If the above dynamic constraints are ignored and a conventional uniform voting mechanism is directly adopted, the obstacle avoidance route that is finally reached may exceed the safe execution capability of some vehicles, causing loss of control or rollover. This scheme incorporates data on maximum steering angle, maximum braking deceleration, and current load, enabling each vehicle to score the safety feasibility of candidate solutions based on its own actual physical limits. The closer the steering angle or braking deceleration is to the vehicle's limit, and the higher the current load leading to a greater risk of rollover, the lower the score. The scheme with the most highest scores, as determined by the lead vehicle, essentially selects the most balanced solution under the overall vehicle safety constraints. This scheme improves the adaptability of obstacle avoidance routes to the actual physical capabilities of each vehicle, reduces the occurrence of loss of control or rollover due to a scheme exceeding the dynamic limits of a particular vehicle, and enhances the safety of platoon obstacle avoidance.

[0097] In one embodiment, reference is made to Figure 8In step S40, the actual obstacle avoidance data corresponding to the obstacle avoidance route data is obtained, an obstacle avoidance score is calculated based on the actual obstacle avoidance data, and the strategy for generating the obstacle avoidance route is optimized based on the obstacle avoidance score. Specifically, this includes: S41: Obtain formation holding deviation data, braking impact data, and additional energy consumption data from actual obstacle avoidance data.

[0098] In this embodiment, formation keeping deviation data refers to the absolute value of the difference between the actual distance between each vehicle and the target distance during the entire obstacle avoidance process, as well as the offset between the actual lateral position and the planned lateral position of the vehicles. Braking impact data refers to the absolute value of the rate of change of acceleration during braking, reflecting the smoothness of braking. Additional energy consumption data refers to the additional energy consumption caused by obstacle avoidance operations, relative to the energy consumption increment of normal straight-line driving.

[0099] Specifically, during obstacle avoidance, each unmanned heavy-duty truck's onboard recorder continuously recorded the following data at a frequency of 10 Hz: the actual distance to the vehicle in front measured by millimeter-wave radar was compared with the target distance given by the planning module to calculate the deviation value; longitudinal and lateral accelerations were collected by the inertial measurement unit, and the rate of change of acceleration was calculated as the impact intensity; instantaneous power was read by the motor controller or engine ECU, subtracted from the baseline power for normal driving at the same speed, and integrated to obtain the additional energy consumption. After obstacle avoidance, all recorded data were aggregated to the lead vehicle. The lead vehicle statistically analyzed the formation maintenance deviation data, calculating the maximum deviation, average deviation, and root mean square deviation. The braking impact intensity data was taken as its maximum and average values. The additional energy consumption data were summed to obtain the total additional energy consumption value. These data constitute the basic material for evaluating the obstacle avoidance effect, providing a quantitative basis for subsequent scoring and optimization.

[0100] S42: Calculate the overall score for this obstacle avoidance based on formation maintenance deviation data, braking impact data, and additional energy consumption data.

[0101] In this embodiment, the comprehensive score is a dimensionless numerical value used to comprehensively evaluate the obstacle avoidance operation in terms of safety, comfort, and economy. A higher score indicates a better obstacle avoidance effect.

[0102] Specifically, the system first maps the three indicators to sub-scores ranging from 0 to 100. Formation maintenance deviation sub-score: 100 points for a maximum deviation less than 0.5 meters, 80 points for 0.5 to 1 meter, 60 points for 1 to 2 meters, and 40 points for greater than 2 meters. Braking impact sub-score: 100 points for a maximum impact less than 2 meters per cubic second, 80 points for 2 to 4 meters, 60 points for 4 to 6 meters, and 40 points for greater than 6 meters. Extra energy consumption sub-score: 100 points for extra energy consumption less than 0.5 kilowatt-hours, 80 points for 0.5 to 1 kilowatt-hour, 60 points for 1 to 2 meters, and 40 points for greater than 2 meters. Then, the scores are weighted and summed according to preset weights: safety 0.5, comfort 0.3, and economy 0.2, emphasizing the importance of safety. The overall score = Formation maintenance deviation sub-score × 0.5 + Braking impact sub-score × 0.3 + Extra energy consumption sub-score × 0.2. The calculated score ranges from 40 to 100 points. This comprehensive score directly reflects the overall quality of the obstacle avoidance attempt, providing clear reward and penalty signals for subsequent strategy optimization.

[0103] S43: If the overall score is lower than the preset score threshold, an online fine-tuning command is triggered to adjust the strategy parameters for generating obstacle avoidance route data.

[0104] In this embodiment, the scoring threshold is a pre-set limit value, such as 75 points. A score below this value indicates that the obstacle avoidance performance is unsatisfactory and needs improvement. The online fine-tuning command refers to the control signal that triggers the reinforcement learning algorithm to make small updates to the parameters in the path planning strategy.

[0105] Specifically, the lead vehicle compares its calculated comprehensive score with a preset score threshold. If the score is below the threshold, the system considers the obstacle avoidance strategy to have significant flaws and requires optimization. This triggers an online fine-tuning command, initiating the local reinforcement learning fine-tuning module. This module employs a near-end policy optimization algorithm, using the complete obstacle avoidance data (including environmental state, candidate solutions, execution results, and scores) as a training sample to perform a gradient update on the policy network. The update magnitude is constrained by the trust domain to ensure that each update does not lead to drastic policy changes, thus guaranteeing safety. The goal of fine-tuning is to adjust the weight parameters in path planning, such as increasing the penalty coefficient for formation keeping deviations or reducing the focus on additional energy consumption when generating candidate solutions. After fine-tuning, the updated policy parameters take effect immediately for the generation of the next obstacle avoidance strategy. Through this online fine-tuning mechanism, the fleet can quickly adjust its behavior for recently encountered specific scenarios (such as frequently occurring construction areas), achieving a learning-while-running effect.

[0106] S44: If the overall score is higher than or equal to the score threshold, the obstacle avoidance route data corresponding to this obstacle avoidance will be marked as an excellent case and saved.

[0107] In this embodiment, an excellent case refers to a successful obstacle avoidance operation record, including the global environment state at the time, the obstacle avoidance scheme adopted, key data during the execution process, and the final comprehensive score. These cases are saved in a local case library for direct use in subsequent similar scenarios.

[0108] Specifically, if the overall score is higher than or equal to the threshold, the system considers the obstacle avoidance operation to have reached a satisfactory level. The lead vehicle packages the entire obstacle avoidance process data into a case file, including: a snapshot of the global obstacle probability distribution map before obstacle avoidance begins, the final obstacle avoidance route data, the actual execution trajectory of each vehicle, the formation keeping deviation curve, the braking impact curve, the cumulative value of additional energy consumption, and the overall score. The case file is named with a timestamp and geographical location and stored in the lead vehicle's solid-state drive. At the same time, the feature summary of the case (such as obstacle type, road curvature, fleet size, and weather conditions) is indexed into the case library's search table. When encountering similar scenarios in the future, the path planning module can first query the case library. If a high-matching, excellent case is found, its obstacle avoidance solution can be directly reused, or it can be slightly adjusted based on it, thereby saving time on regeneration and negotiation. By accumulating excellent cases, the fleet's obstacle avoidance experience is consolidated and passed on, making the handling of common scenarios increasingly efficient and stable.

[0109] In this embodiment, by extracting formation holding deviation data, braking impact data, and additional energy consumption data from actual obstacle avoidance data, a comprehensive score for the obstacle avoidance operation is calculated. Based on the comparison between the score and a preset threshold, online fine-tuning or saving as an excellent case is triggered, achieving closed-loop optimization of the obstacle avoidance strategy. Formation holding deviation reflects the degree of disorder in the convoy formation during obstacle avoidance, braking impact reflects the loss of comfort for passengers or cargo, and additional energy consumption measures the impact of obstacle avoidance operations on economics. The comprehensive score normalizes these multi-dimensional indicators into a single value, facilitating quantitative evaluation of each obstacle avoidance performance. When the score is below the threshold, an online fine-tuning command is triggered to adjust the strategy parameters, enabling the convoy to gradually optimize decision preferences for specific road sections or operating conditions. When the score is above the threshold, the solution is marked as an excellent case and saved for reference and reuse in subsequent similar scenarios. This solution improves the adaptive evolution capability of the convoy obstacle avoidance strategy, allowing the method to continuously accumulate experience and optimize performance over long-term operation, reducing the occurrence of poor performance in new scenarios due to strategy rigidity.

[0110] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0111] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A cooperative obstacle avoidance method for unmanned heavy truck fleets based on multi-agent clusters, characterized in that, The cooperative obstacle avoidance method for unmanned heavy truck fleets based on multi-agent clusters includes: Acquire multimodal sensor data pre-installed on each unmanned heavy truck in the heavy truck fleet, construct fleet communication links based on the multimodal sensor data, and generate leader vehicle information in the heavy truck fleet using a distributed election algorithm; Obtain obstacle recognition information from each of the multimodal sensing data, obtain passage area information and obstacle information from the obstacle recognition information, and generate an obstacle probability distribution map of the heavy truck fleet based on the passage area information and the obstacle information; When unavoidable obstacles are obtained from the obstacle probability distribution map, obstacle avoidance route data for each of the unmanned heavy trucks is generated using the information of the lead vehicle. Obtain the actual obstacle avoidance data corresponding to the obstacle avoidance route data, perform obstacle avoidance scoring based on the actual obstacle avoidance data, and optimize the strategy for generating obstacle avoidance routes based on the obstacle avoidance score.

2. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters as described in claim 1, characterized in that, The step of constructing a fleet communication link based on the multimodal sensing data and generating leader vehicle information in the heavy truck fleet using a distributed election algorithm specifically includes: The communication signal strength data and computing resource data of each unmanned heavy truck are obtained from the multimodal sensing data; A comprehensive score is obtained for each unmanned heavy truck based on the communication signal strength data and the computing power resource data. The unmanned heavy truck with the highest navigation compatibility score is identified as the leader vehicle.

3. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 2, characterized in that, The process of comprehensively evaluating the communication signal strength data and computing resource data to obtain a navigation compatibility score for each unmanned heavy truck specifically includes: The communication delay data and packet loss rate data are obtained from the communication signal strength data, and the central processing unit occupancy rate and remaining memory capacity are obtained from the computing power resource data. A communication quality score is generated based on the communication latency data and the packet loss rate data, and a computing power availability score is generated based on the CPU utilization rate and the remaining memory capacity. The navigation fit score is obtained by weighted summing of the communication quality score and the computing power availability score.

4. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 1, characterized in that, The step of obtaining obstacle recognition information from each of the multimodal sensing data, obtaining passage area information and obstacle information from the obstacle recognition information, and generating an obstacle probability distribution map of the heavy truck fleet based on the passage area information and the obstacle information specifically includes: A local occupancy grid map is obtained from each of the multimodal sensing data, and a grid occupancy probability value is extracted from the local occupancy grid map; The grid occupancy probability value is sent to the lead vehicle corresponding to the leader vehicle information. The lead vehicle then performs a weighted fusion of all the grid occupancy probability values ​​to obtain a global grid occupancy probability value. The obstacle probability distribution map is generated based on the global grid occupancy probability value.

5. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 4, characterized in that, The step of generating the obstacle probability distribution map based on the global grid occupancy probability value specifically includes: Obtain the dynamic grid region corresponding to the dynamic obstacle from the global grid occupancy probability value; Acquire temporal change data of the dynamic grid region, and predict the motion trajectory of the dynamic obstacle based on the temporal change data to obtain predicted trajectory data; Based on the predicted trajectory data, the global grid occupancy probability value is subjected to spatiotemporal filtering to obtain the filtered obstacle probability distribution map.

6. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 1, characterized in that, When unavoidable obstacles are obtained from the obstacle probability distribution map, obstacle avoidance route data for each of the unmanned heavy trucks is generated using the information of the lead vehicle, specifically including: When the unavoidable obstacle is obtained from the obstacle probability distribution map, multiple candidate obstacle avoidance schemes are generated based on the lead vehicle corresponding to the lead vehicle information. The multiple candidate obstacle avoidance schemes are sent to each of the unmanned heavy trucks, and the score data of each unmanned heavy truck for each candidate obstacle avoidance scheme is obtained. Based on the scoring data, a distributed consensus protocol is used for voting. When all the unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the candidate obstacle avoidance scheme that has reached a consensus is determined as the obstacle avoidance route data.

7. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 6, characterized in that, The process of voting based on the scoring data using a distributed consensus protocol, whereby when all the unmanned heavy trucks reach a consensus on the same candidate obstacle avoidance scheme, the consensus-reaching candidate obstacle avoidance scheme is determined as the obstacle avoidance route data, specifically includes: The maximum steering angle, maximum braking deceleration, and current load data are obtained from the dynamic parameters of each of the unmanned heavy trucks. Based on the maximum steering angle, the maximum braking deceleration, and the current load data, a safety feasibility score is given for each candidate obstacle avoidance scheme to obtain the score data; The scoring data is sent to the lead vehicle, which counts the highest number of scores obtained for each candidate obstacle avoidance scheme and determines the candidate obstacle avoidance scheme with the most highest scores as the obstacle avoidance route data that has reached a consensus.

8. The method for cooperative obstacle avoidance of unmanned heavy truck fleets based on multi-agent clusters according to claim 1, characterized in that, The steps of obtaining the actual obstacle avoidance data corresponding to the obstacle avoidance route data, performing obstacle avoidance scoring based on the actual obstacle avoidance data, and optimizing the obstacle avoidance route generation strategy based on the obstacle avoidance score specifically include: Obtain formation holding deviation data, braking impact data, and additional energy consumption data from the actual obstacle avoidance data; The comprehensive score for this obstacle avoidance is calculated based on the formation holding deviation data, the braking impact data, and the additional energy consumption data. If the overall score is lower than the preset score threshold, an online fine-tuning instruction is triggered to adjust the strategy parameters for generating the obstacle avoidance route data. If the overall score is higher than or equal to the score threshold, the obstacle avoidance route data corresponding to this obstacle avoidance will be marked as an excellent case and saved.

Images