AGV intelligent control method and system

Abstract

The invention provides an AGV intelligent control method and system, and belongs to the field of intelligent control, and the method comprises the steps: building an environment grid map based on the pose of an AGV and sensor data, defining the actual cost of the g (s) value of a node s from a starting point to the node s, defining the rhs (s) value of the node s as the advanced estimated cost of the node s to a target point, and judging whether the node s is a non-uniform node or not; according to the relative position of the non-consistent node and the AGV and the environment change type, calculating the updating priority of the non-consistent node; and adding the non-consistent nodes into the priority queue, updating the g (s) value map, and generating and executing a new driving path from the current position of the AGV to the target point. According to the method, the re-planning tasks can be intelligently screened and sorted, computing resources are intensively used for processing the most critical environment change, the decision-making efficiency and response timeliness in a complex environment are improved, and the safety and smoothness of AGV operation and the overall passing efficiency are enhanced.

Description

Technical Field

[0001] The present invention belongs to the field of intelligent control, and in particular relates to an AGV intelligent control method and system. Background Art

[0002] Automated Guided Vehicles (AGVs) are core equipment in modern intelligent logistics and automated production lines. Their path planning and intelligent control technologies are key to their operational efficiency, stability, and safety. In complex and dynamically changing factory or warehouse environments, AGVs need to be able to perceive environmental changes in real time and quickly and accurately plan or re-plan an optimal or suboptimal path from their current location to their destination. Traditional path planning algorithms, such as the A* algorithm, perform well in static environments. However, when dynamic obstacles appear or path information changes, global recalculation is required. This requires a large amount of computation and has poor real-time performance, making it difficult to meet the requirements for efficient AGV operation.

[0003] To address path planning in dynamic environments, the D*Lite algorithm was developed. This incremental heuristic search algorithm maintains path information through a reverse search (from the target point to the starting point). When the environment changes, only the node information in the affected area needs to be updated, without recalculating the entire path. This greatly improves replanning efficiency. However, existing AGV control methods based on D*Lite still have many limitations in practical applications.

[0004] First, in complex factory or warehouse environments, AGVs must be able to flexibly avoid static and dynamic obstacles while maintaining transport efficiency. Classic path planning algorithms such as A* and Dijkstra perform well in static, known environments and can search for optimal paths. In real-world operational scenarios, the environment is dynamic and filled with other AGVs, workers, and temporarily stored goods. This requires path planning algorithms to be able to respond quickly and replan. Compared to global replanning algorithms like A*, when an obstacle in the environment blocks the path, the D*Lite algorithm does not need to recalculate the entire path. Instead, it only updates the cost of the affected area, thereby improving replanning efficiency. However, the replanning trigger mechanism is sensitive and lacks priority distinction. Any minor environmental change can trigger a path update, resulting in an excessive computational burden on the system when environmental disturbances are frequent. Conversely, raising the trigger threshold to reduce computation frequency can result in delayed response to critical obstacles. The calculation model for path cost is overly simplified, typically only considering distance and static obstacles, and failing to fully integrate the AGV's own kinematic characteristics. This can lead to problems such as excessive turns and discontinuous curvature in the planned path, preventing the AGV from executing smoothly and efficiently. It also makes it impossible to quantify and assess some non-explicit traffic risks, such as historically congested areas. This limits the efficiency and safety of AGVs in complex scenarios where humans and machines coexist. Achieving optimal control of AGVs in complex dynamic environments is an urgent issue in this field. Summary of the Invention

[0005] The purpose of the present invention is to provide an AGV intelligent control method and system for solving the problem that existing AGVs are difficult to achieve optimal control in complex dynamic environments, comprising the following steps: An environmental grid map is constructed based on the AGV's position and sensor data. The g(s) value of node s is defined as the actual cost from the starting point to the node s. The rhs(s) value of node s is the estimated cost from node s to the target point. The g(s) and rhs(s) values ​​of each node s are initialized. When the rhs(s) value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted with the local environment complexity and the AGV speed, the node s is determined to be inconsistent. Calculate the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change; add the inconsistent node to the priority queue, and calculate or update its sort key in the queue according to the g(s) value, rhs(s) value, heuristic value and the update priority; Iteratively take out the node u with the smallest keyword from the priority queue for expansion, and update the successor node of the node u with the smallest keyword of The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. ; Based on the updated g(s) value map, generate and execute a new driving path from the AGV's current position to the target point.

[0006] Optionally, the threshold adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

[0007] Optionally, the calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; The update priority P is calculated by the formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

[0008] Optionally, the calculating or updating the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ; ; in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

[0009] Optionally, the transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV reaches the node The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

[0010] The present invention also provides an AGV intelligent control system, comprising the following modules: The inconsistent node determination module is used to build an environmental grid map based on the AGV's posture and sensor data, define the g(s) value of node s as the actual cost from the starting point to the node s, and the rhs(s) value of node s as the estimated cost of node s in advance. The g(s) value and rhs(s) value of each node s are initialized. When the rhs value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted according to the local environment complexity and the AGV speed, the node s is determined to be an inconsistent node. An update module is configured to calculate an update priority of the inconsistent node based on the relative position between the inconsistent node and the AGV and the type of environmental change; add the inconsistent node to a priority queue, and calculate or update its sort key in the queue based on the g(s) value, rhs(s) value, heuristic value, and the update priority; The path determination module is used to iteratively extract the node u with the smallest keyword from the priority queue for expansion and update the successor node of the node u with the smallest keyword of Value, where the transfer cost between nodes The calculation is based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risks evaluated based on the machine learning model; based on the updated g(s) value map, a new driving path from the AGV's current position to the target point is generated and executed.

[0011] Optionally, the threshold adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

[0012] Optionally, the calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; The update priority P is calculated by the formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

[0013] Optionally, the calculating or updating the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ; ; in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

[0014] Optionally, the transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV reaches the node The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

[0015] The present invention establishes a re-planning trigger threshold that reflects the complexity of the local environment and the operating speed of the AGV, and determines the priority of updating nodes based on the type of environmental change and relative position. This allows the algorithm to intelligently screen and sort re-planning tasks, concentrate computing resources on processing the most critical environmental changes, reduce unnecessary computing overhead, and improve decision-making efficiency and response timeliness in complex environments. In addition, when calculating the path cost, the present invention combines the kinematic constraints of the AGV, the trajectory prediction results of mobile obstacles in the environment, and the potential traffic risks assessed based on machine learning models. This makes the planned driving path not only easier for the AGV to execute smoothly, but also can avoid possible future collisions and bypass non-explicit risk sources such as historical congestion areas, thereby enhancing the safety, smoothness and overall traffic efficiency of the AGV operation. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the first embodiment; Figure 2 This is a schematic diagram of AGV path planning; Figure 3 Schematic diagram of dynamic threshold adjustment; Figure 4 A schematic diagram of update priority. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the specific embodiments of this application and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application. It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of relevant data need to comply with relevant laws, regulations and standards, and provide corresponding operation entrances for users to choose to authorize or refuse.

[0018] In the first embodiment, the present invention provides an AGV intelligent control method, such as Figure 1 As shown, the following steps are included: S1, based on the AGV's posture and sensor data, builds an environmental grid map, defines the g(s) value of node s as the actual cost from the starting point to the node s, and the rhs(s) value of node s as the estimated cost from node s to the target point, and initializes the g(s) and rhs(s) values ​​of each node s; when the rhs(s) value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted with the local environment complexity and AGV speed, the node s is determined to be an inconsistent node; The environment point cloud data is acquired by LiDAR, and a real-time positioning and map construction SLAM algorithm such as GMapping or Cartographer is used to generate a two-dimensional environment grid map, in which each grid is marked as occupied, idle or unknown. During initialization, the g(s) and rhs(s) values ​​of all nodes s in the map are set to infinity; only the target point rhs( ) value is set to 0 and added to the priority queue.

[0019] By continuously comparing the latest sensor data with the grid map, when it is discovered that the originally idle grid is occupied by an obstacle or the originally occupied grid becomes idle, the transfer cost between the corresponding nodes changes, which in turn causes the rhs(s) value of the adjacent nodes to be updated. A dynamic threshold is calculated. For example, the threshold is equal to a basic constant divided by the weighted sum of the local environment complexity and the AGV speed. Among them, the local environment complexity can be quantified as the density of the obstacle grid within a certain radius around the AGV, and the AGV speed is its current linear speed. When the absolute value of the difference between the g(s) value and the rhs(s) value of node s is greater than this dynamic threshold, the node s is identified as an inconsistent node.

[0020] In an optional embodiment, the threshold value adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

[0021] The dynamic threshold adjustment mechanism enables the AGV to adapt to environmental changes. When the AGV is operating in complex environments or at high speeds, path planning must respond more quickly to even small changes in the costmap to ensure safety. For example, if 10 obstacle grids are detected within a 5×5 neighborhood centered on the node to be determined, the local environment complexity is 0.4. At this point, if the AGV is traveling at 1.5 meters per second, the base threshold is 10, the complexity weighting factor is 2, and the speed weighting factor is 1, the calculated threshold T is approximately 3.03. This low threshold means that even small differences in the g(s) and rhs(s) values ​​will trigger a node update, ensuring high alertness in dangerous or complex situations.

[0022] On the contrary, when the AGV is traveling at a low speed in an open area, the update frequency can be appropriately reduced to save computing resources. For example, if there are only two obstacle grids in the same neighborhood, the complexity is 0.08, and the AGV speed is reduced to 0.5 meters per second, using the same parameters, the calculated threshold T will increase to about 6.02, such as Figure 3 A higher threshold allows a larger tolerance between the g value and the rhs value without triggering an update, thus avoiding unnecessary calculations when the environment is stable and the risk is low, and improving the overall operation efficiency of the algorithm.

[0023] S2, calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change; adding the inconsistent node to the priority queue, and calculating or updating its sorting key in the queue according to the g(s) value, rhs(s) value, heuristic value and the update priority; The update priority P of the inconsistent node s is calculated by a function that combines the distance and the type of change. For example, P is equal to the weight coefficient Divide by the Euclidean distance between node s and the current position of the AGV, plus the weight coefficient Multiply by a change type value. The change type value is preset. For example, when a new obstacle appears, that is, when the grid changes from idle to occupied, the value is a higher value of 1.0, and when a known obstacle disappears, that is, when the grid changes from occupied to idle, the value is a lower value of 0.2. Insert the non-consistent node s into the priority queue U or update its sorting keyword in the queue. The keyword is a two-dimensional or three-dimensional vector used for queue sorting. For example, the first dimension of the keyword is min(g(s), rhs(s)) plus the heuristic cost h from the current AGV position to the node s. The heuristic cost h is preferably calculated using the Manhattan distance; the second dimension of the keyword is the update priority P calculated in the previous step. The smaller the value, the higher the priority during sorting; the third dimension can be set to min(g(s), rhs(s)) as the final tiebreaker criterion.

[0024] In an optional embodiment, the calculating the update priority P of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; The update priority P is calculated by the formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

[0025] The update priority calculation method ensures that the system prioritizes environmental changes that have the greatest impact on AGV operations. Update priority is determined by two key factors: the distance between the change point and the AGV and the nature of the change. The closer the distance, the more direct the potential impact, and therefore, the higher the priority. Furthermore, newly appearing obstacles pose a greater threat to the current path than those that disappear.

[0026] For example, assuming the distance weight coefficient is 100, the weight coefficient of the change type is 50. When a new obstacle appears only 2 meters away from the AGV, it is assigned a high change type value, such as 1.0. Its update priority P is calculated to be approximately 99.75. In contrast, if an old obstacle disappears 10 meters away, it is assigned a low change type value, such as 0.1, and its priority P is approximately 14.99. Figure 4 In this way, newly added obstacles in the near future are processed first, as they are directly related to obstacle avoidance safety, while the path optimization opportunities represented by the disappearing obstacles in the far distance are placed in a secondary position.

[0027] In an optional embodiment, the calculating or updating the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value, and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ; ; in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

[0028] The two-dimensional sorting key is used to optimize the processing order of nodes in the priority queue, making the path replanning both efficient and timely. Among them, the function of the main key k1 is similar to The F-value in the algorithm combines the known minimum cost from the starting point to the current node with a heuristically estimated cost from the current node to the destination, aiming to guide the search toward the optimal path. When multiple nodes have the same k1 value, the secondary key k2 becomes decisive. By introducing a negative update priority P, it incorporates the urgency of environmental changes as a secondary sorting criterion. For example, consider two pending nodes, sA and sB, both with a k1 value of 15, resulting in a sorting conflict. If node sA is an inconsistent node caused by the disappearance of a distant obstacle and has a lower update priority P, such as 20, its k2 value is -20. On the other hand, node sB is caused by the appearance of a nearby obstacle and has a very high update priority P, such as 100, its k2 value is -100. During sorting, since -100 is less than -20, node sB is removed from the queue and processed first. This ensures that, given similar path costs, updates are prioritized for areas that have the greatest impact on AGV safety. In an alternative embodiment, in order to make the priority play a greater role, the update priority P is integrated into the calculation of k1. Alternatively, or .

[0029] S3, iteratively take out the node u with the smallest keyword from the priority queue for expansion, and update the successor node of the node u with the smallest keyword of The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. ; Based on the updated g(s) value map, generate and execute a new driving path from the AGV's current position to the target point.

[0030] Take the node u with the smallest keyword from the priority queue U and sort all its successor nodes , according to the formula To update its rhs value, transfer cost It is a composite cost, and its value is obtained by weighted summation of multiple parts. The first part is the kinematic cost calculated based on the Reeds-Shepp curve model, which is used to penalize large-angle turns that do not meet the AGV's minimum turning radius. The second part is the dynamic obstacle collision risk predicted by the long short-term memory network (LSTM) model. If a dynamic obstacle is predicted to occupy a node at some point in the future, , the cost will increase significantly, Figure 2The third part is based on the potential risk cost of the pre-trained gradient boosting decision tree, i.e., GBDT model, which assigns higher risk cost to nodes in congested or accident-prone areas according to historical traffic data. In one embodiment, the updated g-value map is used to generate a new driving path. In another alternative embodiment, when the replanning process, i.e., the expansion iteration of the priority queue is stable, the node at the current location of the AGV is selected. At the beginning, we search for the next path point through the greedy strategy, that is, among all neighbor nodes, we select the node that makes the expression Neighbor node with the smallest value This process is repeated until the target point is reached, forming a path consisting of discrete nodes. This discrete path is smoothed using curve fitting algorithms such as B-spline interpolation to generate a trajectory with continuous curvature, which is then executed by the AGV's underlying motion controller. Optionally, path extraction can be performed after the replanning process, which involves expanding nodes and updating g values, is completed.

[0031] In an optional embodiment, the transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV arrives The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

[0032] The composite cost function aims to generate a path that is not only the shortest but also smoother, safer, and smarter in actual operation; it decomposes the cost of a move into four aspects for comprehensive evaluation, making the AGV's decision more in line with complex environmental requirements.

[0033] For example, suppose the AGV plans to move from node u to an adjacent node, and the geometric distance between the two points is 1 unit. If this movement is slower than the one from the parent node The path to u forms a sharp 90-degree turn with a turning penalty coefficient of If is 5, a steering cost of 5 will be generated to penalize non-smooth motion. Furthermore, if a Kalman filter predicts that a moving pedestrian will appear only 1.1 meters away from the adjacent node when the AGV reaches it, and the maximum penalty value A is 10 and the safety impact range parameter σ is 1, then the dynamic obstacle avoidance cost will be calculated to be approximately 5.46. This high cost will prompt the AGV to choose a path away from the pedestrian.

[0034] In addition, this method also introduces risk assessment based on historical data. If the target node is located near a doorway that people often pass by and is close to a wall, the pre-trained machine learning model outputs a risk score of up to 0.8 based on its historical traversal frequency, distance to obstacles and other characteristics. If the risk weight coefficient is If is 3, the potential risk cost is 2.4. Adding up all the costs, the total transfer cost from u to the target node is 13.86. This comprehensive cost value enables the path planner to generate an optimized path that effectively avoids various known and potential risks.

[0035] In a second embodiment, the present invention provides an AGV intelligent control system, including the following modules: The inconsistent node determination module is used to build an environmental grid map based on the AGV's posture and sensor data, define the g(s) value of node s as the actual cost from the starting point to the node s, and the rhs(s) value of node s as the estimated cost from node s to the target point. The g(s) value and rhs(s) value of each node s are initialized. When the rhs(s) value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted according to the local environment complexity and the AGV speed, the node s is determined to be an inconsistent node. An update module is configured to calculate an update priority of the inconsistent node based on the relative position between the inconsistent node and the AGV and the type of environmental change; add the inconsistent node to a priority queue, and calculate or update its sort key in the queue based on the g(s) value, rhs(s) value, heuristic value, and the update priority; The path determination module is used to iteratively extract the node u with the smallest keyword from the priority queue for expansion and update the successor node of the node u with the smallest keyword of Value, where the transfer cost between nodes The calculation is based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risks evaluated based on the machine learning model; based on the updated g(s) value map, a new driving path from the AGV's current position to the target point is generated and executed.

[0036] In an optional embodiment, the threshold value adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

[0037] In an optional embodiment, the calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; The update priority P is calculated by the formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

[0038] In an optional embodiment, the calculating or updating the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value, and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ; ; in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

[0039] In an optional embodiment, the transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV reaches the node The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

[0040] Through the description of the above embodiments, it can be seen that those skilled in the art can clearly understand that the present application can be implemented by means of software plus a necessary general-purpose hardware platform. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, can be embodied in the form of a software product. The computer software product can be stored in a storage medium, such as ROM / RAM, a magnetic disk, an optical disk, etc., and includes a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in various embodiments or certain parts of the embodiments of the present application.

[0041] Each embodiment in this specification is described in a progressive manner. The same or similar parts between the embodiments can be referred to each other. Each embodiment focuses on the differences from other embodiments. In particular, for system or system embodiments, since they are basically similar to method embodiments, the description is relatively simple. For relevant parts, refer to the partial description of the method embodiment. The system and system embodiments described above are merely schematic, wherein the units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0042] The above describes in detail the method and electronic device for providing commodity object information provided by this application. Specific examples are used herein to illustrate the principles and implementation methods of this application. The description of the above embodiments is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the contents of this specification should not be construed as limiting this application.

Claims

1. An AGV intelligent control method, characterized in that: The following steps are involved: An environmental grid map is constructed based on the AGV's position and sensor data. The g(s) value of node s is defined as the actual cost from the starting point to the node s. The rhs(s) value of node s is the estimated cost from node s to the target point. The g(s) and rhs(s) values ​​of each node s are initialized. When the rhs(s) value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted with the local environment complexity and the AGV speed, the node s is determined to be inconsistent. Calculate the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change; add the inconsistent node to the priority queue, and calculate or update its sort key in the queue according to the g(s) value, rhs(s) value, heuristic value and the update priority; Iteratively take out the node u with the smallest keyword from the priority queue for expansion, and update the successor node of the node u with the smallest keyword of The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. ; Based on the updated g(s) value map, generate and execute a new driving path from the AGV's current position to the target point.

2. The method according to claim 1, characterized in that The threshold adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

3. The method according to claim 1, characterized in that The calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; The update priority P is calculated by the formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

4. The method according to claim 1, wherein The calculating or updating of the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ； ； in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

5. The method according to claim 1, wherein The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated by the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV reaches the node The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

6. An AGV intelligent control system, characterized in that: Includes the following modules: The inconsistent node determination module is used to build an environmental grid map based on the AGV's posture and sensor data, define the g(s) value of node s as the actual cost from the starting point to the node s, and the rhs(s) value of node s as the estimated cost from node s to the target point. The g(s) value and rhs(s) value of each node s are initialized. When the rhs(s) value of any node s is updated due to environmental changes, if the absolute value of the difference between the updated rhs(s) value and the g(s) value exceeds a threshold adjusted according to the local environment complexity and the AGV speed, the node s is determined to be an inconsistent node. An update module is configured to calculate an update priority of the inconsistent node based on the relative position between the inconsistent node and the AGV and the type of environmental change; add the inconsistent node to a priority queue, and calculate or update its sort key in the queue based on the g(s) value, rhs(s) value, heuristic value, and the update priority; The path determination module is used to iteratively extract the node u with the smallest keyword from the priority queue for expansion and update the successor node of the node u with the smallest keyword of The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. ; Based on the updated g(s) value map, generate and execute a new driving path from the AGV's current position to the target point.

7. The system according to claim 6, characterized in that The threshold adjusted according to the complexity of the local environment and the AGV speed includes: With the node to be determined s as the center, select an N×N grid neighborhood and count the number of obstacle grids in the neighborhood , calculate the local environment complexity ; Get the current linear speed of the AGV ; The threshold T is given by the formula Calculate, where is the basic threshold, is the complexity weighting coefficient, is the speed weighting coefficient, and N is 5.

8. The system according to claim 6, wherein: The calculating the update priority of the inconsistent node according to the relative position between the inconsistent node and the AGV and the type of environmental change includes: Calculate the Euclidean distance between the inconsistent node s and the current position of the AGV ; If node s changes from a non-obstacle to an obstacle, then is a high priority value; if node s changes from an obstacle to a non-obstacle, then is a low priority value; Update priority P by formula Calculate, where is the distance weight coefficient, is the change type weight coefficient, and ε is a small positive constant to avoid the denominator being zero when the distance is zero.

9. The system according to claim 6, wherein: The calculating or updating of the sorting key in the queue according to the g(s) value, the rhs(s) value, the heuristic value and the update priority includes: The sorting key is a two-dimensional vector K=[k1,k2], and the calculation method is: ； ； in, From the current position of AGV The heuristic value to node s, P(s) is the calculated update priority.

10. The system according to claim 6, wherein: The transfer cost between nodes is calculated based on the AGV kinematic constraints, the spatiotemporal trajectory prediction results of dynamic obstacles, and the potential passage risk evaluated based on the machine learning model. , specifically: The transfer cost between the nodes It is calculated by adding the base geometry cost, steering cost, dynamic obstacle avoidance cost, and potential risk cost: ; Among them, the basic geometric cost From node u to node The geometric distance; The steering cost Calculated by: Get the parent node of node u , calculate the vector With vector The angle θ between them, ,in is the steering penalty coefficient; The dynamic obstacle avoidance cost Calculated by: Using Kalman filter to predict dynamic obstacles when AGV reaches the node The estimated position at the time of calculation node The distance from the location at the estimated time , , where A is the maximum penalty value, is the safety impact range parameter; The potential risk cost Calculated by: The feature vector is input into the pre-trained machine learning model to obtain the traffic risk score , and then through Calculated; wherein, the feature vector includes nodes The historical traversal frequency, the distance to the nearest static obstacle, and the historical pedestrian flow data in the area, is the risk weight coefficient.

Images