AGV logistics vehicle system based on autonomous algorithm

Abstract

The invention relates to the technical field of industrial automation and intelligent robots, and discloses an AGV (Automatic Guided Vehicle) logistics vehicle system based on an autonomous algorithm, which adopts a cloud side-end collaborative architecture and a cloud execution time window genetic algorithm to perform global scheduling. During genetic evolution, an improved RRT * algorithm is adopted to plan a path for each AGV; according to the improved algorithm, an artificial potential field method is fused to guide an initial path, and a self-adaptive step length strategy is adopted to accelerate convergence. In order to realize no collision, the system establishes a retention time window for a path node, and pre-judges and solves the conflict by detecting the intersection of the time window. The AGV as a terminal realizes autonomous high-precision positioning and path execution by fusing multi-sensor information. Through combination of global optimization and local independence, the problems of path conflict and low planning efficiency are solved, the path planning speed and quality are improved, and efficient and reliable collision-free intelligent logistics scheduling is realized.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and intelligent robot technology, specifically to an AGV logistics vehicle system based on an autonomous algorithm. Background Technology

[0002] Currently, with the deepening development of intelligent manufacturing and industry, Automated Guided Vehicles (AGVs) have become a core component of the internal logistics system of smart factories. AGV systems are responsible for autonomously transporting materials between production lines, warehouses, and workstations, and their operating efficiency and level of intelligence directly determine the production cycle and automation level of the entire factory.

[0003] To address the aforementioned logistics automation needs, existing technologies employ a solution that relies entirely on a central dispatch center. In this solution, the central server receives all tasks and plans a complete travel path for each AGV. The AGV itself merely acts as an execution unit, moving according to the path coordinate sequence issued by the server. All environmental perception and decision-making calculations are centrally handled by the central server.

[0004] However, this centralized scheduling has limitations in scenarios with multiple AGVs operating concurrently. When multiple AGVs run simultaneously, path intersections easily create traffic bottlenecks. The system lacks a predictive conflict resolution mechanism, often only able to passively wait after a conflict occurs, leading to low overall operational efficiency. In task-intensive areas, a stalemate can even occur where multiple AGVs are locked together. The traditional path planning algorithms used also have efficiency bottlenecks. For example, some algorithms take too long to search for paths in complex environments. Other algorithms generate paths containing numerous unnecessary inflections and broken lines, requiring frequent acceleration and deceleration by the AGVs, which not only increases energy consumption but also affects the smoothness of travel. Furthermore, the system's adaptability to dynamic environments is insufficient; temporary obstacles appearing in the factory, such as personnel or temporarily stacked goods, cannot be effectively avoided by this type of system. The AGVs themselves lack autonomous decision-making capabilities; once communication with the central server is interrupted or delayed, they will stop working, paralyzing the entire logistics system and making reliability difficult to guarantee.

[0005] Therefore, the present invention provides an AGV logistics vehicle system based on an autonomous algorithm to address the shortcomings of the prior art. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an AGV logistics vehicle system based on an autonomous algorithm, which solves the problems of frequent path conflicts and low efficiency of global scheduling and planning when multiple AGVs work together in a smart factory environment.

[0007] To achieve the above objectives, the present invention provides an AGV logistics vehicle system based on an autonomous algorithm, comprising: At least one AGV; A cloud-based scheduling platform is used to execute a time-window genetic algorithm for multiple tasks to generate a collision-free path scheme for at least one AGV and other AGVs. The time-window genetic algorithm is characterized by using an improved RRT* algorithm to plan a path for each AGV when planning paths for individuals in the population.

[0008] By adopting the above technical solution, a time-window genetic algorithm is used as the top-level framework for multi-AGV scheduling, while an improved RRT* algorithm is used as the execution unit for bottom-level path generation. This combination enables the system to not only find the optimal task allocation and path combination at the global level through the genetic algorithm, but also to ensure efficiency and quality at the level of individual path planning. This fundamentally solves the path conflict problem when multiple AGVs work together, and improves the operational efficiency and stability of the entire logistics system.

[0009] Preferably, when the cloud scheduling platform performs path and time window planning, it establishes a reserved time window for each key node on the AGV path. The reserved time window defines the start and end times of the AGV occupying the node. The cloud-based scheduling platform determines node conflicts by detecting whether there is a time overlap between the retention time windows of two AGVs at the same node. Furthermore, the cloud-based scheduling platform also determines edge conflicts by detecting whether AGVs are traveling in opposite directions on the same edge and their travel times overlap, and determines following conflicts by detecting whether AGVs traveling in the same direction are less than the safe distance.

[0010] By adopting the above technical solutions, the system expands the AGV's occupation of spatial resources from a purely two-dimensional space to a three-dimensional spatiotemporal space that includes the time dimension, achieving precise quantitative management of node occupation. The conflict detection method based on a time-window reservation provides deterministic data for subsequent path adjudication, ensuring the accuracy of conflict judgment.

[0011] Preferably, when the cloud scheduling platform detects a path conflict, it executes a conflict resolution strategy, which includes: Dynamic solution strategy, the dynamic solution strategy includes: When responding to node or edge conflicts, a dynamic waiting time window is inserted for lower-priority AGVs. When responding to a following conflict, the following vehicle reduces its speed; The replanning strategy calls the improved RRT* algorithm to replan a conflict-free path for one of the AGVs; In addition, a coordinated speed adjustment strategy is provided, which, based on the prediction of the future trajectory of the AGV, fine-tunes the speed of the AGV with potential conflict risk in order to stagger the conflict points in time.

[0012] By adopting the above technical solutions, the system possesses flexible conflict handling capabilities. The waiting strategy handles simple conflicts with minimal computational overhead, while the replanning strategy can fundamentally resolve complex conflicts or those caused by high-priority tasks. The combination of the two enables the system to achieve an effective balance between the real-time performance and optimality of the path.

[0013] Preferably, the time window genetic algorithm uses a fitness function to evaluate the merits of each collision-free path scheme. The value of the fitness function is determined by weighted summation of the total running time required for all AGVs to complete the task, the total path length of all AGVs, and the predicted congestion cost, using set weight coefficients. Its further feature is that the predicted congestion cost is derived from the future traffic flow prediction generated by a spatiotemporal prediction core module.

[0014] By adopting the above technical solution, a clear optimization objective is provided for the evolutionary direction of the genetic algorithm. By adjusting the weight coefficients, the scheduling system can focus on pursuing the shortest overall task completion time or on the shortest overall travel distance to save energy, thereby enabling the optimization results of the algorithm to adapt to different production cycles and management needs.

[0015] Preferably, the improved RRT* algorithm includes an adaptive step size strategy; this strategy dynamically adjusts the expansion step size based on the interior angle of the triangle formed by the random sampling point, the nearest optimal parent node in the random tree, and the final target point, located at the optimal parent node; when the interior angle is greater than 90 degrees, a preset base step size is used; when the interior angle is less than or equal to 90 degrees, a step size that increases according to the cosine value of the interior angle is used.

[0016] By adopting the above technical solution, the stochastic expansion process of the RRT* algorithm is improved. When the expansion direction is towards the target, a larger step size is used to accelerate convergence; when the expansion direction is away from the target, a smaller step size is used to reduce invalid exploration. This dynamic adjustment mechanism improves the convergence speed and computational efficiency of path planning.

[0017] Preferably, the improved RRT* algorithm is an algorithm that incorporates the artificial potential field method: In the node expansion loop of the random tree, the artificial potential field force at the current node is first calculated to define a dominant search direction; Then, a directional sampling region is defined in the dominant search direction; Finally, random sampling is performed within the directional sampling area to generate the next sampling point, and node expansion is performed.

[0018] By adopting the above technical solution, the lack of directionality in the initial random sampling of the RRT* algorithm is overcome. The initial path provided by the artificial potential field method provides a good starting point and directional guidance for subsequent RRT* optimization, making the entire path planning process more efficient and the generated final path of higher quality.

[0019] Preferably, the AGV is used to output a fused positioning result by fusing positioning information provided by the map from laser SLAM, absolute positioning information provided by the visual QR code, and trajectory estimation information provided by the gyroscope.

[0020] By adopting the above technical solutions, the AGV has achieved high-precision autonomous positioning capabilities. The fusion and complementarity of multiple information sources ensure that the AGV can maintain accurate position and attitude recognition even if a sensor temporarily fails or loses information. This is the foundation for its reliable execution of global paths and local adjustments.

[0021] Preferably, the system adopts a cloud-edge-device collaborative architecture, which includes the AGV as the endpoint, the cloud-based scheduling platform as the cloud, and the gateway and network as the edge; the gateway and network are used to realize data transmission between the AGV and the cloud-based scheduling platform.

[0022] By adopting the above technical solutions, a well-defined and highly efficient system architecture was constructed. The cloud is responsible for computationally intensive global optimization, ensuring the global optimality of the scheduling results; the AGVs on the edge side possess a high degree of autonomy, responsible for real-time perception and execution. This architecture balances the intelligence of centralized scheduling with the agility of decentralized control, improving the overall performance and reliability of the system.

[0023] This invention provides an AGV logistics vehicle system based on an autonomous algorithm. It has the following beneficial effects: 1. This invention achieves efficient, collision-free collaborative operation of multiple AGV systems by employing a time-window genetic algorithm for global scheduling and combining it with time-reserved windows for conflict detection and resolution. This scheme extends the management of AGVs' occupation of path nodes from a two-dimensional space to a spatiotemporal dimension. It uses a genetic algorithm to find the globally optimal task allocation and path combination, and proactively eliminates path interference using waiting or replanning strategies. This solves the congestion and conflict problems of multiple vehicles operating simultaneously, improving the overall logistics efficiency of smart factories.

[0024] 2. This invention improves the speed and quality of path planning by employing an improved RRT* algorithm. The algorithm first uses an artificial potential field method to generate an initial path with a clear directional guide, avoiding the blind search in the initial stages of the traditional RRT* algorithm. Then, in the optimization phase, an adaptive step-size strategy is adopted, allowing the algorithm to converge faster with larger step sizes when moving towards the target direction. This combination results in shorter computation time for path planning and a smoother, more optimized path.

[0025] 3. This invention enhances system reliability by constructing a cloud-edge-device collaborative architecture and equipping the AGV with multi-sensor fusion positioning technology. Under this architecture, the cloud handles global optimization, while the AGV, with its autonomous positioning capabilities, performs real-time execution and local adjustments, ensuring the system's adaptability to dynamic obstacles or network fluctuations. Simultaneously, the fusion of laser, vision, and gyroscope positioning ensures the accuracy and continuity of the AGV's position perception in complex environments, providing a solid foundation for the system's stable and reliable operation.

[0026] 4. This invention, by introducing a spatiotemporal prediction core and deeply coupling its prediction results with the original scheduling algorithm, achieves a shift in the system's decision-making mode from reactive to proactive. In the path planning stage, by incorporating the predicted congestion cost into the genetic algorithm, the system can proactively avoid potential future congestion areas. In the conflict resolution stage, the predicted trajectory is used to upgrade from passive waiting to proactive coordinated speed regulation. This not only reduces scheduling conflicts and waiting in dynamic environments at the source but also improves the overall smoothness and resource utilization efficiency of the AGV system. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the artificial potential field method of the present invention; Figure 2 This is a schematic diagram illustrating the calculation angle of the present invention; Figure 3 This is a simulation diagram of the path planning of a small vehicle according to the present invention; Figure 4 This is a flowchart of the time window genetic algorithm of the present invention; Figure 5 This is an environmental model diagram of the present invention; Figure 6 This is a schematic diagram of the time window of the present invention; Figure 7 This is a simulation diagram of the path planning for the two vehicles in this invention; Figure 8 This is a schematic diagram of the spatiotemporal prediction core module architecture and data interaction of the present invention. Detailed Implementation

[0028] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] This invention provides an AGV logistics vehicle system based on an autonomous algorithm, which adopts a cloud-edge-device collaborative architecture.

[0030] The core of this system lies in the fact that highly autonomous AGV units (ends) and cloud-based scheduling platforms (clouds) responsible for global optimization collaborate through a high-speed network to form an organic whole that is both decentralized and centrally optimized.

[0031] This collaborative architecture includes AGVs (endpoints), gateways and networks (edges), and a cloud-based scheduling platform (cloud). The AGVs (endpoints) connect to the cloud platform via the gateways and networks (edges). This architecture enables global scheduling in the cloud, while the AGVs can perform autonomous positioning and real-time obstacle avoidance, working collaboratively.

[0032] Specifically, in one embodiment, the AGV intelligent cloud gateway (edge) is connected to the PLC and the PC with the cloud scheduling platform deployed via a wired connection. The AGV intelligent cloud gateway communicates with the AGV (end) equipped with a WIFI-STA module via wireless WIFI.

[0033] In this architecture, the AGV (the AGV itself) is an independent intelligent unit, equipped with a local computing platform and a complete sensor suite. The AGV (endpoint) achieves core functions such as precise positioning, real-time obstacle avoidance, and local path planning without relying on cloud commands. Its autonomy is the foundation for the system's high reliability and high response speed.

[0034] The cloud (scheduling platform) is responsible for macro-level strategies, including task allocation for multiple AGVs, global path planning, traffic scheduling (based on time windows), and system status monitoring.

[0035] Edge networks and gateways, such as industrial IoT gateways and 5G / WiFi networks, form a high-speed data channel connecting the endpoint and the cloud, ensuring that command and status data can be transmitted with low latency and reliability.

[0036] In a preferred embodiment, to address the scheduling lag and efficiency bottleneck issues arising in the dynamic execution environment due to the original system's lack of predictive ability for future dynamic evolution trends, the cloud scheduling platform was functionally enhanced.

[0037] See attached document Figure 8This enhancement solution adds a spatiotemporal prediction core module between the task scheduler and the path planner within the original cloud scheduling platform.

[0038] The data flow of the core module for spatiotemporal prediction is as follows: Input: Receives status information (position, speed, orientation) and environmental map topology information (i.e., a grid map that defines key nodes) uploaded in real time by the AGV through the edge device, as well as historical task data accumulated by the cloud platform, such as (task start point, end point, path, completion time) records.

[0039] Output: Outputs traffic load predictions for key nodes (e.g., the predicted number of AGVs passing through each key node in the next N minutes) and occupancy probability maps for path segments to the task scheduler and path planner.

[0040] In one embodiment, the spatiotemporal prediction core module consists of a core model based on a spatiotemporal graph neural network (ST-GNN), whose purpose is to predict the future spatiotemporal operating state of the entire AGV system by learning from historical data and real-time status. Its specific construction and operation are as follows: Graph Construction: Each AGV and each critical path point monitored by the system (i.e., the node that retains the time window in the original time window mechanism) is defined as a node in the graph. The feature vector of the node contains information such as the AGV's (position, speed, orientation, and task target point). Simultaneously, three types of edges are constructed to accurately model the spatiotemporal relationships between AGVs: Distance edge: Connects adjacent AGV nodes in space.

[0041] Conflict point edges: At path intersections, establish connections for AGV nodes that may interact with each other, explicitly modeling potential conflict relationships.

[0042] Task-related edges: Connect AGV nodes heading to the same workstation (target point) to model resource competition.

[0043] Core model: Spatiotemporal graph neural network (ST-GNN) is used for prediction.

[0044] Spatial Convolution: Graph Attention Network (GAT) is used to enable the model to dynamically calculate the influence weights between nodes, thereby intelligently judging the degree of mutual influence between nodes. For example, an AGV heading towards an intersection will pay close attention to another AGV that is also heading towards the intersection.

[0045] Temporal convolution: Using gated recurrent units (GRUs), time-series data such as position and speed of each AGV node are modeled to capture its motion trend.

[0046] Key enhancements: During model training and prediction, structured map information (such as passages and restricted areas) can be incorporated as graph attributes; physical constraints (such as maximum turning radius and acceleration) can be added to the loss function; and the task target point can be used as an important feature input to guide the model to make purposeful trajectory predictions.

[0047] Model Training and Output: The ST-GNN model is trained under supervised supervision using historical AGV operation data accumulated in the cloud, enabling it to learn and predict traffic flow patterns in specific factory environments. Finally, the model outputs the predicted trajectory for each AGV and aggregates these data to generate a system-level traffic load prediction and path segment occupancy probability map.

[0048] The AGV robot body in this embodiment of the invention mainly consists of two parts: a motion chassis body and a robotic arm body.

[0049] In one embodiment, the main body of the motion chassis is an industrial heavy-duty dual-wheel differential mobile chassis. Its load-bearing frame is made of seamless rectangular tubing, and the outer casing is made of metal, housing the control cabinet, battery, computing unit, and other components. The chassis dimensions are approximately 500*300*300mm, and an operable display screen is mounted on the front.

[0050] The main drive and power components of this sports chassis utilize two high-power brushless DC motors for dual-wheel differential drive, powered by ternary lithium batteries, supporting manual, automatic, and battery swapping charging modes. The chassis features a multi-wheel adaptive suspension system, with all four omnidirectional wheels employing shock-absorbing technology. The seamless rectangular tube load-bearing frame and adaptive suspension design ensure that all wheels remain in contact with the ground even under heavy loads (e.g., 100kg) through suspension adjustments.

[0051] The control and communication components of the main chassis include a low-level controller and an upper-level computing platform. The low-level controller uses an embedded computing platform with a Cortex-M4 core, responsible for motor control and battery management. The upper-level computing platform uses a 64-bit Intel i5 processor or higher, responsible for running perception and planning algorithms. The communication module supports direct WiFi 6 connection, establishing a connection with the cloud scheduling platform via workshop wireless AP or mobile network (measured average latency of 87ms).

[0052] The safety protection components of the main body of the sports chassis are equipped with dual emergency stop buttons, front and rear safety touch edges, front and rear obstacle avoidance ultrasonic sensors, obstacle avoidance lidar, and front obstacle avoidance binocular structured light (optional).

[0053] In one embodiment, the robotic arm is a seven-axis collaborative robot with a body weight of 4 kg, a rated load of ≤2 kg, and a working radius of 600 mm. This robotic arm, combined with dynamic algorithms, enables single-handed teaching, sensorless collision detection, and drag trajectory reproduction. It supports multiple communication methods including Wi-Fi, Bluetooth, and Ethernet, and can be wirelessly connected and controlled via a dedicated app. Its end effector features a modular gripper interface, allowing for manual replacement of grippers or suction cups.

[0054] In one embodiment, the sensing unit sensors are configured for sensing by the AGV. Each sensor is connected to the upper-level computing platform via a data cable or internal bus for data fusion.

[0055] The sensing unit includes: LiDAR (e.g., RPLIDARM2M1) is responsible for environmental scanning and mapping. Under specific conditions, with a scanning frequency of 15Hz and a detection distance of 0.112m (error ±3cm), it can generate a grid map with a resolution of 5cm.

[0056] Industrial monocular cameras (such as the Hikvision DS2CD3320) are used for cargo identification and positioning. Under standard lighting conditions, they achieve a 90% success rate in recognizing QR codes and support cargo size measurement (error ±2mm). Ultrasonic radar (such as the HCSR04) serves as a near-range blind spot compensation sensor, with a measured effective detection range of 0.32m. It works in conjunction with lidar, using filtering algorithms to optimize near-range obstacle avoidance stability.

[0057] High-precision gyroscopes are used for trajectory calculation and are fused with visual and laser positioning information.

[0058] The AGV body (end) outputs a fused positioning result by fusing the map from laser SLAM, the absolute positioning from visual QR codes, and the trajectory calculation from the gyroscope (the measured repeatability error of the robotic arm is ≤2mm).

[0059] See attached document Figure 1 The single AGV path planning method in this embodiment of the invention first performs step one, which is to search for the initial path using the artificial potential field method.

[0060] like Figure 1 As shown, this step constructs a virtual force field in the AGV's workspace. This force field is composed of a superposition of gravitational and repulsive fields.

[0061] Gravitational field: A gravitational field is generated at the target point. Its function is to produce an attractive force pointing towards the target point, guiding the AGV towards it. In one embodiment, the magnitude of the gravitational force is proportional to the distance from the AGV's current position to the target point. The calculation formula is: ; in: This represents the AGV at the location point. The gravitational force experienced at that time.

[0062] This is the positive gain coefficient of the gravitational field, used to adjust the magnitude of gravity.

[0063] Represents the current position of the AGV With the target point The Euclidean distance between them.

[0064] Repulsive Field: The repulsive field is generated by the obstacle. Its function is to produce a repulsive force away from the obstacle, propelling the AGV to avoid it. This repulsive field has a localized range of effect; that is, the repulsive force only takes effect when the AGV enters the predetermined influence range of the obstacle. In one embodiment, the magnitude of the repulsive force is inversely proportional to the distance between the AGV and the obstacle. When the distance between the AGV and the obstacle is greater than the influence range, the repulsive force is zero. The calculation formula is: ; in: This represents the AGV at the location point. The repulsive force experienced at that time.

[0065] This is the positive gain coefficient of the repulsive field, used to adjust the magnitude of the repulsive force.

[0066] Represents the current position of the AGV and obstacles The shortest distance between them.

[0067] The threshold representing the range of influence of an obstacle, i.e., the maximum distance at which the repulsive field begins to act.

[0068] Resultant force and path generation AGV at any position point Total virtual resultant force It is gravity and repulsive force The vector sum, calculated using the following formula: ; By calculating the resultant force Under the influence of the gradient descent direction, the system can plan an approximate initial path from the starting point to the target point. This process, guided by the force field, effectively avoids the blindness caused by random sampling in the initial search phase of the RRT* algorithm.

[0069] In a preferred embodiment, the combination of step one (artificial potential field method) and step two (improved RRT algorithm) is deeply integrated. This integration method aims to solve the problems of large sampling randomness and slow convergence speed of traditional RRT algorithm, as well as the tendency of traditional artificial potential field method to get trapped in local minima.

[0070] The core of this fusion method lies in integrating the macroscopic guidance of the artificial potential field method (as mentioned above) The defined constraints are embedded into each step of the sampling expansion loop of the RRT* algorithm, specifically implemented through the following ordered steps: Step S110: Real-time direction guidance based on artificial potential field. At a node in the current random tree to be expanded... At this point, calculate the artificial potential field force acting on it. The direction of this resultant force is defined as the dominant search direction vector for this path expansion, providing a clear macroscopic guide for subsequent random sampling.

[0071] Step S120: Perform directional random sampling within the potential field guidance domain. Using the dominant search direction vector obtained in step S110 as the central axis, define a dynamic fan-shaped sampling region (e.g., within 30° on each side of the axis). Randomly generate the next sampling point within this limited fan-shaped region. Instead of sampling completely randomly throughout the free space, this step transforms the macroscopic guidance of the artificial potential field into a real-time, dynamic constraint on the microscopic sampling space of the RRT* algorithm.

[0072] Step S130: Perform node expansion and optimization of the improved RRT* algorithm. Based on the guiding sampling points obtained in step S120. Continue executing the core process of the improved RRT algorithm. This process includes: finding the distance in the random tree. The most recent node, and the application (described later) of an adaptive step-size strategy to generate new nodes. And optimizations for reselecting parent nodes and rerouting in the RRT algorithm.

[0073] Step S140: Loop Iteration and Termination Judgment. The newly generated nodes... Add it to the random tree and use it as a potential for the next iteration. One, repeat steps S110 to S130 until the random tree expands to the tolerance range of the target point.

[0074] Through the aforementioned cyclical steps of guidance-sampling-expansion, the artificial potential field method provides real-time directional guidance in each iteration of the RRT algorithm, fundamentally solving the problems of blind sampling and slow convergence in the traditional RRT algorithm from a methodological perspective.

[0075] In the step of path optimization using the improved RRT* algorithm, the method includes a target bias strategy.

[0076] The target bias strategy is as follows: during random sampling, a threshold is set. When the generated random number is less than the threshold, the target point is directly used as the sampling point.

[0077] This strategy allows the random tree to grow probabilistically in the target direction, reducing ineffective expansion.

[0078] In a preferred embodiment, the target bias threshold is preferably set to between 0.1 and 0.3. This range is based on the following engineering trade-offs: an excessively high threshold (e.g., greater than 0.3) makes the algorithm overly greedy and prone to early local optima in complex obstacle environments; an excessively low threshold (e.g., less than 0.1) degenerates it into a near-traditional random search, offering limited efficiency gains. Extensive simulations have shown that 0.2 is an optimal value that strikes the best balance between convergence speed and robustness in environment exploration.

[0079] See attached document Figure 2 The single AGV path planning method in this embodiment of the invention further includes an adaptive step size strategy in the step of using the improved RRT* algorithm for path optimization, which is used to dynamically adjust the expansion efficiency of the random tree.

[0080] like Figure 2 As shown, the specific implementation process of this strategy is as follows: First, in one iteration of the algorithm, a random sampling point is selected. Then, in the generated random tree, the distance from the sampling point is searched. Nearest optimal parent node At this time, the sampling point Optimal parent node With the preset final target point Together, they form a virtual triangle. Next, the location of this triangle at its vertices is calculated. interior angle at The interior angle This reflects the degree of deviation of the sampling point from the target direction relative to the current expansion direction. The system uses this interior angle... The size is determined, and the expansion step size is dynamically adjusted. The calculation method for this step size is defined as follows: ; in: This represents the dynamically adjusted step size.

[0081] This represents a preset base step size value.

[0082] Representative from sampling points Optimal parent node and target point In the triangle formed, with Let be the interior angle of the vertex.

[0083] Represents interior angle The cosine value.

[0084] The specific logic of this strategy lies in: when the interior angle When the angle is greater than 90°, it indicates that the angle originates from the optimal parent node. Pointing to sampling point The direction is away from the target point. The direction of the random tree is then considered. At this point, the random tree expands in directions away from the target point, which is usually a fruitless exploration. Therefore, the algorithm uses a small base step size. Expand cautiously to reduce invalid searches.

[0085] Conversely, when the interior angle When the angle is less than or equal to 90°, it indicates that the expansion direction is generally toward the target point. To accelerate convergence, the algorithm uses a relatively large step size. Because within this angle range, ,therefore It is a gain factor ranging from [value] to [value]. When the expansion direction is perfectly aligned with the target direction (i.e., ...), [the gain factor is defined]. When the step size reaches its maximum value. When the expansion direction is perpendicular to the target direction (i.e.) When the step size is restored to the base value, the step size is restored. .

[0086] By implementing this adaptive step size strategy, the random tree can expand rapidly when moving towards the target direction, and expand more cautiously when moving away from the target direction, thereby improving the overall efficiency and convergence speed of path planning.

[0087] The adaptive step size formula The design basis is: The magnitude of the value directly reflects the degree to which the new expansion direction aligns with the target direction.

[0088] when hour, In Interval, step size exist arrive The dynamic scaling between these points demonstrates the algorithm's confidence in expanding toward the target and encourages rapid exploration.

[0089] when hour, If it is negative, then the basic step size is used. (As stated in the original formula,) This reflects the caution when expanding away from the target, avoiding entering the invalid area due to large strides.

[0090] This strategy uses simple geometric relationships to efficiently convert the target point's orientation information into a dynamic control signal for the search step size.

[0091] See attached document Figure 3 The single AGV path planning algorithm in this embodiment of the invention (which combines the artificial potential field search in step one and the improved RRT* algorithm optimization in step two) ultimately generates a smooth, short, and high-quality optimal path.

[0092] The specific reason why this optimization effect can be achieved is as follows: By employing the aforementioned real-time direction guidance based on an artificial potential field (step S110) and performing directional random sampling within the potential field guidance domain (step S120), the search space of the RRT algorithm is limited from the sampling source, enabling it to quickly focus on valuable regions. This fundamentally solves the problems of blind initial search and slow convergence speed in the traditional RRT algorithm.

[0093] Furthermore, by introducing an adaptive step size strategy, intelligent step size adjustment is achieved during the path expansion process, which accelerates when moving toward the target and remains robust when deviating from the target, reducing invalid sampling and redundant nodes.

[0094] Through the above integration and improvement, the path planned by this algorithm has a shorter planning time, fewer turns, better length, and higher smoothness compared with traditional algorithms.

[0095] See attached document Figure 5 The multi-AGV scheduling and collision-free path planning method in this embodiment of the invention first requires environmental modeling.

[0096] In one embodiment, a raster modeling method is used to convert the workshop environment into a virtual electronic map.

[0097] like Figure 5 As shown, in this virtual electronic map, white squares are defined as free areas, and green squares are defined as obstacle areas.

[0098] See attached document Figure 4 In the time window genetic algorithm of this invention embodiment, step 1 is first executed: initializing the population.

[0099] In this step, multiple sets of task assignment and path planning schemes are randomly generated; these schemes are referred to as individuals.

[0100] Each individual represents a sequence of AGV task executions.

[0101] See attached document Figure 4 and attached Figure 6 In the time window genetic algorithm of this embodiment, when executing step 2: path and time window planning in its process, the improved RRT* algorithm is first used to plan a path from the starting point to the target point for each AGV in the task allocation scheme.

[0102] Meanwhile, in order to ensure the collision-free characteristics of the path in a multi-AGV environment, this embodiment introduces the concept of a time window to accurately manage and record the time that each AGV occupies key nodes on the path (e.g., grids in a grid map, path intersections in a topology map).

[0103] like Figure 6 As shown, this embodiment defines two types of time windows: 1. A time window is reserved, which records the specific time period during which a node is occupied by a certain AGV. Its set representation is as follows: ; in: It represents the set of all time windows retained in the system.

[0104] This represents a specific reserved event, namely the AGV. For nodes Time occupied.

[0105] It is a closed time interval, with a clearly defined start and end point for the occupation. It is an AGV Arrival or start of occupying a node At that moment, This means that it leaves or ends the node it occupies. The moment. Therefore, this retention time window clearly indicates: AGV In time period Internally occupied node .

[0106] 2. Idle Time Window: This time window records the period during which a node is not occupied by a particular AGV. Its set representation is as follows: ; in: It represents the set of all idle time windows in the system.

[0107] This represents a specific idle event, i.e., a node. Compared to AGV A period of available time.

[0108] It is a time interval, representing the AGV. No nodes were occupied during this period. Therefore, this idle time window clearly indicates that the AGV... In time period Unoccupied nodes .

[0109] By establishing a corresponding retention time window for each AGV's path during the path planning phase, the system can accurately grasp the occupancy status of each node on the future timeline, providing an accurate data foundation for subsequent conflict detection and resolution.

[0110] See attached document Figure 4 and attached Figure 6 In the embodiments of the present invention, the time window genetic algorithm performs step 3 after path and time window planning (step 2): conflict detection and resolution.

[0111] like Figure 4 As shown, this step checks the set of paths generated for each individual in step 2, which includes multiple AGVs, to determine whether there are conflicts between the paths. In one embodiment, a conflict is precisely defined as: in the same task allocation scheme, two or more AGVs attempt to occupy the same node in the same environment model within a certain time period.

[0112] Specifically, the detection process utilizes the retention time window established for each AGV's path node in step 2. To proceed. If the system detects the AGV... At the node Retention window With AGV At the same node Retention window There is a temporal intersection, i.e., an interval. Then determine the AGV and AGV At this node A path conflict occurred.

[0113] When a path conflict is detected, the system will activate the conflict resolution strategy.

[0114] In one embodiment, the strategy can be a waiting strategy. This strategy prioritizes AGVs with lower priority (e.g., based on a preset priority rule). Upon reaching the conflict node The AGV paused and entered a waiting state at a previous node. The waiting period ends when a higher-priority AGV (e.g., ...) is triggered. After completely leaving the conflict node, i.e., when the current system time is greater than the end of its retention time window. Only then is AGVk2 ​​allowed to continue moving forward.

[0115] In another embodiment, the strategy can be a replanning strategy. This strategy is triggered under specific conditions, such as when the expected waiting time caused by the waiting strategy exceeds a preset threshold. Or, when the AGV whose path needs to be adjusted is carrying a high-priority task, the cloud scheduling platform will call the improved RRT* algorithm in this invention to adjust the path for that AGV (e.g., ...). Re-plan a conflict-free path. The newly generated path will be able to achieve one of two effects: completely bypass the conflict node. Alternatively, after querying and confirming the existence of a new idle time window. Then, pass through the node within the new idle time window.

[0116] In a preferred embodiment, the conflict detection and resolution mechanism is further refined, employing a conflict resolution mechanism based on spatiotemporal joint planning. The innovation of this method lies in transforming the conflict from a struggle for spatial resources to a scheduling of temporal resources. By introducing a dynamic waiting time window, the possibility of multi-vehicle collisions is physically isolated from a temporal perspective. The specific steps of this mechanism are as follows: Step S310: Precise definition of conflict type.

[0117] In addition to the aforementioned node-based conflict detection, the system further defines conflicts precisely into multiple types: 1. Node conflict: Detect the retention time window of any two AGVs. Are they on the same node? There is a time overlap.

[0118] 2. Edge Conflict: Detect whether two AGVs are traveling in opposite directions on an edge formed by two adjacent nodes, and whether their time windows occupying the edge overlap.

[0119] 3. Follow-up Conflict: Detect two AGVs traveling on the same path, and determine whether the expected arrival time of the following AGV at a certain node is earlier than the departure time of the preceding AGV, and whether the distance between the two AGVs is less than the safe distance between them.

[0120] Step S320: Priority-based dynamic resolution strategy.

[0121] Once any of the above conflicts is detected, immediately initiate the following resolution strategies to supplement or optimize the original waiting or replanning strategies: If there is a node conflict or an edge conflict, the AGV with the lower task priority will insert a dynamic waiting time window at its current node. The duration of the waiting time window is equal to the overlap duration of the conflict time window plus a safety margin.

[0122] If a collision occurs, the following vehicle will be forced to reduce its speed to maintain a safe distance from the vehicle in front.

[0123] Step S330: Conflict resolution verification and iteration.

[0124] To resolve potential secondary conflicts that the strategy may cause, the system will re-examine the global time window after the strategy is applied, until all conflicts are eliminated.

[0125] By performing conflict detection and resolution steps, it can be ensured that each evaluated individual (path scheme) in the genetic algorithm is collision-free, thus guaranteeing that the final output optimal solution is an executable and safe set of multi-AGV paths.

[0126] In a further embodiment, the conflict resolution strategy was enhanced to upgrade from post-event remediation to pre-event prevention.

[0127] This enhancement strategy introduces a dynamic conflict risk assessment based on predicted trajectories. Specifically, the system utilizes the future predicted trajectories output by the aforementioned spatiotemporal prediction core module (ST-GNN) to identify potential conflict pairs in advance that have non-overlapping static time windows but whose predicted trajectories are very close in time and space, posing a risk. For these potential conflict pairs, the system no longer employs a passive waiting strategy but instead initiates proactive collaborative speed adjustment. Through cloud-based commands, the speeds of the relevant AGVs are slightly adjusted, allowing them to smoothly stagger conflict points in time, achieving uninterrupted passage. This reduces the number of AGV waiting times and the total waiting time, improving the system's smoothness in dynamic environments.

[0128] See attached document Figure 4 In the time window genetic algorithm of this embodiment, after executing step 3 (conflict detection and resolution), step 4 is then executed: calculating fitness.

[0129] In this step, the system calculates the fitness value for each individual (i.e., a complete set of multi-AGV task allocation and conflict-free path schemes). In one embodiment, this fitness It is defined as an indicator for evaluating the overall performance of the scheme, used to measure the merits of the scheme.

[0130] For example, this fitness It can be defined as the total running time and total path length The weighted sum is calculated using the following formula: ; in: This represents the fitness value of the scheme.

[0131] This represents the total running time required for all AGVs in the scheme to complete their respective tasks.

[0132] This represents the total path length of all AGVs running in this scheme.

[0133] and For example, the preset weighting coefficients can be set to satisfy... The optimization goal of this step is to find a way to improve fitness. The smallest individual.

[0134] In a preferred embodiment, the fitness function was improved to enable the evolution from offline planning based on static environments to forward-looking planning based on dynamic predictions.

[0135] This improved scheme adds a congestion prediction cost term to the original fitness function. The new fitness function (i.e., the total cost) can be defined as: ; in: This represents the path length.

[0136] For time cost.

[0137] This represents the improved fitness value (total cost) of the scheme.

[0138] This represents the predicted congestion cost, calculated from the path segment occupancy probability map output by the aforementioned spatiotemporal prediction core module. When a planned path needs to pass through a critical node with high predicted traffic load, its corresponding congestion cost will increase accordingly.

[0139] These are the preset weighting coefficients.

[0140] This improvement allows the genetic algorithm to preferentially eliminate schemes that are likely to enter future congested areas from the early stages of evolution, guiding the algorithm's evolution from the source. Next, as Figure 4 As shown, proceed to step 5: genetic evolution.

[0141] In this step, the system calculates the fitness based on the fitness calculated in step 4. The value is used to perform genetic operations on the current population to generate a new generation of population. In one embodiment, genetic operations include selection, crossover, and mutation.

[0142] Selection operation: based on individual fitness Individuals are selected based on their value. For example, an elite retention strategy can be used to select individuals from the current population. One or more individuals with the best (i.e., the smallest) fitness value are directly copied into the next generation population; alternatively, a roulette wheel selection method can be used, based on the reciprocal of the fitness value. The proportion of individuals determines the probability of being selected, in order to select other individuals.

[0143] Crossover operation: with a preset crossover probability The selected pairs of individuals are then cross-crossed. For example, this can be done by swapping parts of the task assignment sequences represented by the two individuals.

[0144] Mutation operation: with a preset mutation probability Mutation can be performed on individuals within a population. For example, mutation can be achieved by randomly changing the execution order of a task in the task sequence represented by that individual.

[0145] Next, proceed to step 6: terminate the judgment.

[0146] The system determines whether a preset termination condition is met. In one embodiment, the termination condition is: the number of iterations of the algorithm reaches a preset maximum number of iterations. .

[0147] If the judgment result is that the termination condition is met (e.g., the number of iterations has reached the specified value), then... If the system outputs the optimal solution, then proceed to step 7: Output the optimal solution. The system will select the population with the best fitness. (Right now The individual with the smallest value is output as the final multi-AGV task allocation and collision-free path planning scheme.

[0148] If the result indicates that the termination condition is not met, the algorithm returns to step 2: path and time window planning, and replaces the current population with the new generation population generated in step 5, and continues to the next iteration.

[0149] See attached document Figure 7This figure illustrates a specific example of the final collision-free path scheme generated for a scheduling task involving multiple AGVs after executing a time-window genetic algorithm and outputting the optimal solution. In this simulation environment, it is assumed that two vehicles, AGV1 and AGV2, are performing the task simultaneously.

[0150] For example, AGV1's task is to move from the starting grid (0,0) to the target grid (9,0). Through the global scheduling optimization of the algorithm in this invention, the system plans an executable path for AGV1, namely, along the ground... Figure 1 The path from (0,0)→(1,0)→…→(9,0).

[0151] Meanwhile, another AGV2 may be performing other tasks (e.g., from (0,5) to (9,5)). The scheduling algorithm of this invention has detected and resolved all potential path intersections or resource occupation conflicts in advance by establishing a reservation time window for the path nodes of the two AGVs. Figure 7 The path of AGV1 shown is ultimately determined as part of the optimal solution, under the premise of ensuring that there is no collision with AGV2 and all other vehicles.

[0152] therefore, Figure 7 This intuitively demonstrates that the multi-AGV scheduling method of the present invention can successfully plan safe, collision-free and efficient executable paths for multiple AGVs.

[0153] An example of the system operation process in this embodiment of the invention is as follows: Task Received: The cloud-based scheduling platform loads the environment model and receives multiple tasks. Several AGVs within the workshop are in standby mode.

[0154] Multi-AGV scheduling and global path planning: The cloud-based scheduling platform executes a time-window genetic algorithm to schedule multiple AGVs and plan collision-free paths for the received tasks.

[0155] Step 1: This process first initializes the population and generates multiple initial schemes for task allocation and path planning.

[0156] Step 2: Next, for each scheme, the improved RRT* algorithm is used to plan the path and time window for each AGV, and a retention time window for the path node is established.

[0157] Step 3: Then, the system performs conflict detection and resolution. It determines path conflicts by checking for overlapping time windows and uses waiting or replanning strategies to eliminate the conflicts.

[0158] Step 4: After ensuring that there are no path conflicts, calculate the fitness of each solution. This fitness is used to evaluate the overall performance of the solutions.

[0159] Step 5: Subsequently, genetic evolution is performed based on the fitness value, and a new generation of the scheme population is generated through operations such as selection, crossover, and mutation.

[0160] Step 6: The above process (steps 2 to 5) is iterated continuously until the termination condition is met, such as reaching the maximum number of iterations.

[0161] Step 7: Finally, the algorithm outputs the individual with the best fitness, which is the final multi-AGV task allocation and collision-free global path scheme.

[0162] Path execution and local path adjustment: The cloud-based scheduling platform distributes the optimal global path generated in step 2 to the designated AGVs.

[0163] The AGV begins to move according to the received path instructions.

[0164] During operation, if the AGV encounters an obstacle not marked in the environment model, it will perform real-time local path adjustment using an initial path search method based on an artificial potential field. The obstacle will generate a repulsive force field, guiding the AGV to detour around it.

[0165] After successfully bypassing the obstacle, the AGV continued to execute the global path issued by the cloud scheduling platform.

[0166] Task completed: After each AGV reaches its target point, it reports that the task is completed, and the system then prepares to execute a new task cycle.

Claims

1. An AGV logistics vehicle system based on an autonomous algorithm, characterized in that, include: At least one AGV; A cloud-based scheduling platform is used to execute a time-window genetic algorithm for multiple tasks to generate a collision-free path scheme for at least one AGV and other AGVs. The time-window genetic algorithm is characterized by using an improved RRT* algorithm to plan a path for each AGV when planning paths for individuals in the population.

2. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, When the cloud scheduling platform plans paths and time windows, it establishes a reserved time window for each key node on the AGV path. The reserved time window defines the start and end times when the AGV occupies the node. The cloud-based scheduling platform determines node conflicts by detecting whether there is a time overlap between the retention time windows of two AGVs at the same node. Furthermore, the cloud-based scheduling platform also determines edge conflicts by detecting whether AGVs are traveling in opposite directions on the same edge and their travel times overlap, and determines following conflicts by detecting whether AGVs traveling in the same direction are less than the safe distance.

3. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, When the cloud-based scheduling platform detects a path conflict, it executes a conflict resolution strategy, which includes: Dynamic solution strategy, the dynamic solution strategy includes: When responding to node or edge conflicts, a dynamic waiting time window is inserted for lower-priority AGVs. When responding to a following conflict, the following vehicle reduces its speed; The replanning strategy calls the improved RRT* algorithm to replan a conflict-free path for one of the AGVs; In addition, a coordinated speed adjustment strategy is provided, which, based on the prediction of the future trajectory of the AGV, fine-tunes the speed of the AGV with potential conflict risk in order to stagger the conflict points in time.

4. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The time window genetic algorithm uses a fitness function to evaluate the merits of each collision-free path scheme. The value of the fitness function is determined by weighted summation of the total running time required for all AGVs to complete the task, the total path length of all AGVs, and the predicted congestion cost in the scheme, using set weight coefficients. Its further feature is that the cloud-based scheduling platform also includes a spatiotemporal prediction core module, and the predicted congestion cost is derived based on the future traffic flow prediction generated by the spatiotemporal prediction core module.

5. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The improved RRT* algorithm includes an adaptive step-size strategy, characterized by: The expansion step size is dynamically adjusted based on the interior angle of the triangle formed by the random sampling point, the nearest optimal parent node in the random tree, and the final target point, located at the optimal parent node. When the interior angle is greater than 90 degrees, a preset base step size is used; When the interior angle is less than or equal to 90 degrees, a step size that increases based on the cosine value of the interior angle is used.

6. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The improved RRT* algorithm is an algorithm that incorporates the artificial potential field method: In the node expansion loop of the random tree, the artificial potential field force at the current node is first calculated to define a dominant search direction; Then, a directional sampling region is defined in the dominant search direction; Finally, random sampling is performed within the directional sampling area to generate the next sampling point, and node expansion is performed.

7. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The AGV is used to output a fused positioning result by fusing the following three types of positioning information: Location information provided by the map from laser SLAM; Absolute positioning information provided by visual QR codes; And trajectory calculation information provided by the gyroscope.

8. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The sensing unit carried by the AGV includes: LiDAR is used for environmental scanning and mapping. Industrial monocular cameras are used for cargo identification and positioning. At least one ultrasonic radar is used as a blind spot compensation sensor for near-range blind spots.

9. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The main body of the AGV's motion chassis includes: A load-bearing frame using seamless rectangular tubing; A multi-wheel adaptive suspension system, wherein the adaptive suspension system is used to ensure that all wheels are always in contact with the ground through suspension adjustment under heavy loads.

10. The AGV logistics vehicle system based on an autonomous algorithm according to claim 1, characterized in that, The intelligent factory AGV logistics vehicle system based on autonomous algorithms adopts a cloud-edge-device collaborative architecture, which includes: The AGV serves as the end point; As a cloud-based scheduling platform; As a gateway and network, the gateway and network are used to enable data transmission between the AGV and the cloud scheduling platform.

Images