A control method and system for a double-sided fork narrow aisle stacker robot

Abstract

The present application relates to a kind of double-sided fork narrow channel stacker robot control method, comprising the following steps: receiving the task of operation containing target position information;Control vehicle body to move to preliminary operation point;After reaching, according to the type of operation, adjust fork to identification height, and obtain three-dimensional point cloud data by laser radar scanning target area;Processing point cloud data, identifying target pose and calculating deviation;According to deviation, adjust vehicle body pose;Control fork transverse movement and execute access operation.The method realizes efficient double-sided operation in narrow channel, high-precision positioning and safe operation.

Description

Technical Field

[0001] This application relates to the field of warehouse logistics automation technology, and more specifically, to a control method and system for a double-sided fork narrow-aisle stacking robot. Background Technology

[0002] As modern warehousing and logistics systems continue to evolve towards high-density storage and full automation, the adoption of narrow aisle layouts and high-bay racking structures to maximize space utilization has become a common industry practice. These operating environments require stacking robots to perform efficient, precise, and safe goods storage and retrieval operations on racks exceeding 8 meters in height within narrow aisles less than 1800 mm wide. This poses a significant challenge to the robot's maneuverability, structural stability, and intelligent control capabilities. Current automated stacking equipment has significant shortcomings in handling these scenarios: while traditional single-sided forklift robots have a simple structure, they need to repeatedly leave the aisle to adjust their direction when handling goods on either side of the aisle, leading to frequent interruptions in the workflow, severely limiting efficiency, and failing to meet the continuous operation requirements of modern logistics.

[0003] To compensate for the shortcomings of single-sided forks, the industry has introduced omnidirectional forklift robots with forks that can move laterally or rotate, such as three-way forklifts. However, this technical solution still faces several obstacles: First, the mast and drive mechanism required to achieve lateral movement or rotation of the forks are complex and bulky. Under high-lift and heavy-load conditions, it is difficult to balance structural strength, operational stability, and positioning accuracy within a limited space, posing potential safety risks. Second, the process of large-scale directional adjustment of the forks in narrow aisles is time-consuming, further compressing the already limited working space and significantly reducing the fault tolerance in extremely narrow environments, making collisions highly likely. Third, existing perception and control systems mainly focus on robot navigation and obstacle avoidance, or only perform rough positioning of fixed-height shelves, generally lacking real-time, high-precision three-dimensional pose recognition capabilities for dynamic vehicles and high-level storage locations, as well as multi-degree-of-freedom collaborative fine-tuning mechanisms for the vehicle body and forks based on the recognition results.

[0004] This technological deficiency forces operators to rely on subjective experience or multiple trial adjustments to achieve alignment in actual high-level stacking operations. This is not only inefficient but also lacks real-time proactive safety verification methods in high-risk scenarios, posing a risk of goods falling or colliding with shelves. Overall, existing technologies cannot achieve efficient, accurate, and continuous storage and stacking of goods on both sides within ultra-narrow aisles while ensuring safe and stable high-level operations. There is an urgent need to overcome the comprehensive technical bottlenecks in addressing narrow aisle space constraints, bilateral storage efficiency, high-level stacking accuracy, and operational safety protection. Summary of the Invention

[0005] The purpose of this application is to provide a control method and system for a double-sided fork narrow-channel stacking robot, which has the advantages of achieving efficient double-sided access, high-precision pose adjustment and enhanced operational safety in narrow channels.

[0006] This application provides a control method for a double-sided fork narrow-channel stacking robot, the technical solution of which is as follows:

[0007] A control method for a double-sided fork narrow-aisle stacking robot includes the following steps:

[0008] S1: Receive job task, which contains information about whether the target position is left or right;

[0009] S2: Control the vehicle to navigate to the preparatory work point, which is either a preparatory pickup point or a preparatory unloading point;

[0010] S3: After arriving at the preparatory work point, the forks are adjusted to the preset recognition height according to the work type, and the target area is scanned by multi-line LiDAR to obtain three-dimensional point cloud data; when picking up goods, the target carrier and forks are scanned, and when unloading goods, the carrier on the forks and the target storage location are scanned.

[0011] S4: Process 3D point cloud data, identify target pose and calculate deviation information for adjustment;

[0012] S5: Based on the calculated deviation information, control the vehicle body to adjust its posture;

[0013] S6: Based on the target position information in the task, control the forks to move laterally to the left or right, and perform storage and retrieval operations including forward alignment, lifting or lowering, and backward retraction.

[0014] Furthermore, this application also proposes that in step S4, when the operation is picking up goods, the three-dimensional point cloud data is processed to identify the end face of the vehicle, specifically including identifying the two support legs of the vehicle and calculating their center pose, thereby obtaining the lateral deviation (y) and heading angle deviation (θ) of the vehicle relative to the forks.

[0015] Furthermore, this application also proposes that in step S4, when the operation is unloading, the three-dimensional point cloud data is processed to identify the target storage location center, specifically including: based on the prior information of the shelf position and width, the point cloud data is detected within the corresponding range and the shelf column features are extracted, and then the storage location center pose is calculated, and finally the relative deviation between the center of the carrier on the fork and the target storage location center is obtained.

[0016] Furthermore, this application also proposes that, in step S3, controlling the forks to adjust to a preset identification height according to the operation type is a closed-loop control based on the fork height information fed back by the pull-rope encoder.

[0017] Furthermore, this application also proposes that, before step S5 is executed, a judgment step is included: judging whether the deviation information calculated in step S4 is within the preset safety adjustment threshold; if yes, then step S5 is executed; if no, then an anomaly is reported and the current operation process is terminated.

[0018] Furthermore, this application also proposes that, before step S2, a safety initialization step is included: determining whether the forks are in a preset safety reset position; if not, then the forks are preferentially moved to the safety reset position.

[0019] Furthermore, this application also proposes that when the operation is unloading and stacking, a safety verification step is included before step S5: based on the point cloud data obtained by scanning the shelf location, the point cloud density in the target stacking area is checked. If there are unknown obstacles, the verification fails to prevent collision.

[0020] Furthermore, this application also proposes a control system for a dual-sided fork narrow-aisle stacking robot, used to execute the above-mentioned control method, including:

[0021] Vehicle body;

[0022] The mast is mounted on the vehicle body;

[0023] The fork mechanism, mounted on the mast, includes a lifting mechanism for driving the forks to rise and fall, and a lateral movement mechanism for driving the forks as a whole to move laterally to the left or right.

[0024] The pose recognition module includes a multi-line LiDAR.

[0025] The navigation and control module includes a main controller and a visual positioning unit for vehicle positioning and path navigation;

[0026] The feedback module includes a drawstring encoder for detecting the absolute position of the fork lifting and lateral movement;

[0027] The main controller is connected to the pose recognition module, navigation and control module, feedback module and forklift mechanism.

[0028] Furthermore, this application proposes that a multi-line lidar is fixedly installed at the bottom of a fixed crossbeam between the two forks on the mast, configured to rise and fall synchronously with the mast, and not move with the lateral movement of the forks; the installation position of the multi-line lidar satisfies the following: when the mast is in a low position, its scanning field of view is not obstructed by the forks, so as to identify the front face of the ground carrier and the support legs; when the mast is in a high position, its scanning field of view can identify the support legs of the carrier on the forks and the uprights and crossbeams of the racking storage location.

[0029] Furthermore, this application also proposes that the visual positioning unit includes a camera positioned facing the ground for identifying navigation code tapes laid on the ground.

[0030] Furthermore, this application also proposes to include a safety detection module, which includes microswitches and photoelectric sensors located at the tips of the forks, as well as anti-collision contact edges and obstacle avoidance lidar located around the vehicle body.

[0031] Furthermore, this application also proposes that the navigation and control module further includes: a navigation lidar installed on the top of the vehicle body, serving as the main sensor for SLAM mapping and global positioning; a blind spot lidar installed at the front of the vehicle body, used to detect low obstacles and obstacles in the blind spot of the navigation lidar; and a high-precision IMU integrated inside the vehicle body, used to measure the vehicle's pitch, roll, and angular velocity.

[0032] As can be seen from the above, the dual-sided fork narrow aisle stacking robot control method and system provided in this application solves the technical problems of low efficiency, poor accuracy and high safety risks of dual-sided storage and retrieval in narrow aisles by receiving work tasks, navigating to the preparation point, adjusting the fork height and scanning the target area, processing point cloud data recognition deviations, adjusting the vehicle body posture, and controlling the lateral movement of the forks to perform storage and retrieval operations. It has the advantages of high efficiency, accuracy and safety. Attached Figure Description

[0033] Figure 1 This application provides a flowchart of the picking task for a dual-sided fork narrow-aisle stacking robot.

[0034] Figure 2 This application provides a flowchart of the unloading task for a dual-sided fork narrow-aisle stacking robot.

[0035] Figure 3 This is a schematic diagram illustrating the identification and detection results of a dual-sided fork narrow-channel stacking robot provided in this application.

[0036] Figure 4 A side view of a dual-sided fork narrow-channel stacking robot provided in this application.

[0037] Figure 5 A top view of a double-sided fork narrow-channel stacking robot provided in this application. Detailed Implementation

[0038] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments. The components of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0039] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this application, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0040] This embodiment provides a control method for a dual-sided fork narrow-aisle stacking robot. Traditional narrow-aisle stacking robots face technical problems such as low efficiency, insufficient accuracy, and high safety risks when performing dual-sided storage and retrieval operations in high-density warehousing environments. Single-sided fork robots require repeated adjustments to the vehicle's orientation, resulting in frequent operation interruptions; while omnidirectional fork robots can move laterally, their complex structure and poor stability, coupled with a lack of high-precision three-dimensional pose recognition and coordinated fine-tuning capabilities between the vehicle and forks, lead to positioning difficulties, low efficiency, and the risk of collisions or goods falling.

[0041] Example 1:

[0042] like Figure 1 and 2 As shown in Figure 3, this application proposes a control method for a double-sided fork narrow-channel stacking robot, including the following steps:

[0043] S1: Receive job task, which contains information about whether the target position is left or right;

[0044] S2: Control the vehicle to navigate to the preparatory work point, which is either a preparatory pickup point or a preparatory unloading point;

[0045] S3: After arriving at the preparatory work point, the forks are adjusted to the preset recognition height according to the work type, and the target area is scanned by multi-line LiDAR to obtain three-dimensional point cloud data; when picking up goods, the target carrier and forks are scanned, and when unloading goods, the carrier on the forks and the target storage location are scanned.

[0046] S4: Process 3D point cloud data, identify target pose and calculate deviation information for adjustment;

[0047] S5: Based on the calculated deviation information, control the vehicle body to adjust its posture;

[0048] S6: Based on the target position information in the task, control the forks to move laterally to the left or right, and perform storage and retrieval operations including forward alignment, lifting or lowering, and backward retraction.

[0049] For ease of understanding, the following explains some key terms in this embodiment:

[0050] The task refers to the instruction issued by the host system to the stacker robot. It contains the information required to complete a specific logistics operation, such as the identification of the target goods, the picking task, the unloading task, the target storage or picking location, and whether the target location is on the left or right side of the robot aisle.

[0051] A preparatory work point refers to a preset location that a stacker robot needs to navigate to before performing a specific storage or retrieval operation. This location can be a preparatory retrieval point, which is the designated location where the robot is prepared to pick up the carrier from the shelf or the ground; or a preparatory unloading point, which is the designated location where the robot is prepared to place the carrier on the shelf or the ground.

[0052] Multi-line lidar is a sensor that emits multiple laser beams and receives reflected signals, acquiring three-dimensional distance information of the surrounding environment by measuring the round-trip time of the lasers. It can generate high-density three-dimensional point cloud data for environmental perception and target recognition.

[0053] 3D point cloud data refers to a collection of discrete points representing the surface of an object, acquired by multi-line lidar or other 3D sensors. Each point typically contains 3D coordinates (X, Y, Z) and possible intensity information, which can accurately describe the spatial shape and position of the target object.

[0054] Target pose refers to the position and orientation of a target object in three-dimensional space. Position is usually represented by three-dimensional coordinates, while orientation is represented by rotation matrices or Euler angles, used to describe the precise orientation of the target relative to the robot or world coordinate system.

[0055] Deviation information refers to the difference between the target pose and the robot's current pose or fork pose. This information typically includes lateral deviation, longitudinal deviation, height deviation, and heading angle deviation, and is used to guide the robot in making precise adjustments.

[0056] Storage and retrieval operations refer to the entire process by which a stacker robot retrieves or places goods. This process typically involves a series of coordinated actions, such as the forks advancing to align with the target, the forks rising or falling to grab or place the goods, and the forks retracting to safely retrieve the goods.

[0057] The control method in this embodiment specifically includes the following steps:

[0058] First, the robot receives the task. This task, issued by the host system, includes target location information, explicitly indicating whether the target goods are located on the left or right side of the robot's aisle. For example, the task could be a data packet containing instructions such as "pick up, target location A, left side" or "unload, target location B, right side." This information guides the robot's subsequent lateral movement of the forks, avoiding the problems of traditional robots needing to turn around completely or repeatedly drive out of the aisle to adjust their direction.

[0059] Secondly, the robot is guided to a designated work point. This could be a pickup point, where the robot prepares to retrieve the vehicle from a shelf or the ground; or a unloading point, where the robot prepares to place the vehicle onto a shelf or the ground. Vehicle navigation can be achieved in various ways, such as inertial navigation based on pre-set waypoints or positioning navigation by recognizing visual markers on the ground.

[0060] Furthermore, upon reaching the pre-operation point, the forks are adjusted to a preset recognition height according to the operation type, and a multi-line LiDAR scans the target area to obtain 3D point cloud data. Specifically, when the operation type is picking, the forks are adjusted to a height suitable for scanning the target vehicle and the forks themselves (low position), and the multi-line LiDAR scans the target vehicle (e.g., a pallet) and its surrounding environment. When the operation type is unloading, the forks are adjusted to a height suitable for scanning the vehicle on the forks and the target storage location (high position), and the multi-line LiDAR scans the vehicle on the forks and the target storage location. The fork height adjustment can be achieved through a motor-driven lead screw or chain mechanism, with position feedback provided by an encoder.

[0061] Subsequently, the acquired 3D point cloud data is processed to identify the target pose and calculate the deviation information used for adjustment. For example, during a picking operation, the point cloud data is used to identify the overall outline of the target vehicle, and a geometric matching algorithm is used to calculate the relative position and attitude deviation of the vehicle relative to the forks. During an unloading operation, the point cloud data is used to identify the boundary features of the target storage location, and a feature extraction algorithm is used to calculate the relative deviation between the center of the vehicle on the forks and the center of the target storage location. This deviation information can include lateral deviation, longitudinal deviation, height deviation, and heading angle deviation, etc.

[0062] Based on the calculated deviation information, the vehicle body is controlled to adjust its position and orientation. This adjustment can be achieved through differential control of the drive wheels or coordinated control of the omnidirectional wheels. For example, if a lateral deviation is calculated, the vehicle body is controlled to make a slight adjustment to the left or right; if a yaw angle deviation exists, the vehicle body is controlled to rotate at a small angle. This adjustment process aims to ensure that the vehicle body and forks are precisely aligned with the target, providing an accurate initial position for subsequent storage and retrieval operations.

[0063] Finally, based on the target location information in the task, the forks are controlled to move laterally to the left or right, performing storage and retrieval operations including forward alignment, lifting or lowering, and retraction. For example, if the task indicates the target is on the left, the forks are controlled to move laterally to the left to the designated position. Subsequently, the forks perform a forward movement to insert into the carrier or enter the storage location, then lift (retrieve) or lower (unload) depending on the operation type to complete the grabbing or placement of goods, and finally perform a retraction movement to safely retract the forks. Lateral movement of the forks can be achieved through a side-shifting mechanism, such as a gear and rack mechanism driven by a motor or a hydraulic cylinder.

[0064] The control method proposed in this embodiment effectively solves the problems of low efficiency, insufficient alignment accuracy, and potential safety risks in dual-sided storage and retrieval operations on narrow aisle high-level racks by systematically receiving dual-sided operation tasks, providing precise navigation, combining multi-line LiDAR for three-dimensional pose recognition and deviation calculation, and coordinating the control of the vehicle body and forks for multi-degree-of-freedom adjustments. This enables efficient, accurate, continuous, and safe storage, retrieval, and stacking operations of goods on both sides within extremely narrow aisles.

[0065] In some embodiments described above, this application proposes processing 3D point cloud data to identify the target pose and calculate deviation information for precise storage and retrieval operations. However, in picking operations, the lack of a specific identification method for the carrier end face may lead to insufficient identification accuracy, making it impossible to accurately calculate the lateral deviation and heading angle deviation of the carrier relative to the forks, thus affecting the accuracy of subsequent adjustments. Therefore, this application further proposes that in the above method, when the operation is picking, processing the 3D point cloud data to identify the carrier end face specifically includes identifying the two support legs of the carrier and calculating their center pose, thereby obtaining the lateral deviation (y) and heading angle deviation (θ) of the carrier relative to the forks, such as... Figure 3 As shown.

[0066] Specifically, when a dual-fork narrow-aisle stacker robot receives a picking task, its control system determines that a picking operation is required based on the task type. In this scenario, the robot initiates a specific point cloud data processing procedure to identify the end face of the target vehicle. The three-dimensional point cloud data is a dataset containing a large number of discrete three-dimensional coordinate points obtained after multi-line LiDAR scans the target area, reflecting the geometric shape and spatial position of the target object. Identifying the end face of the vehicle refers to extracting the surface features of the side of the vehicle (e.g., pallet, bin, etc.) facing the robot from these point cloud data. This process is the basis for subsequent accurate alignment, aiming to separate the key geometric information of the target vehicle from the complex environmental point cloud. In one implementation, a geometric feature extraction method can be used. For example, a planar or near-planar structure can be identified through point cloud segmentation algorithms (such as RANSAC, region growing, etc.), and then, combined with prior knowledge of the vehicle's known dimensions and shape, the point cloud clusters that best match the end face features of the vehicle can be selected from these planes. In another implementation, a deep learning-based approach can be used to train a 3D point cloud semantic segmentation network (such as PointNet, PointCNN, etc.), which can directly identify and label the points belonging to the "vehicle end face" from the original point cloud data, thereby achieving accurate extraction of the end face.

[0067] After identifying the vehicle's end face, the method further focuses on identifying the vehicle's two support legs. The support legs (e.g., the fork legs of a pallet or the bottom support structure) are relatively stable, easily identifiable parts of the vehicle's structure with well-defined geometric features. Identifying these two support legs is a crucial step in accurately determining the vehicle's pose, as they provide the vehicle's horizontal positioning reference and attitude information. By identifying the support legs, identification errors caused by irregularities in the vehicle surface or uneven stacking of goods can be effectively avoided. In one implementation, based on the identified point cloud of the vehicle's end face, edge detection or feature point extraction algorithms can be further used to find regions in the point cloud with specific geometric shapes (e.g., elongated or rectangular cross-sections) and relative positional relationships (parallel, fixed-distance intervals), and these regions can be identified as support legs. In another implementation, three-dimensional models or feature templates of different types of vehicle support legs can be pre-established. In the point cloud data, template matching or feature descriptors (e.g., FPFH, SHOT) are used to identify support leg structures that match the pre-defined templates.

[0068] Subsequently, based on the two identified support legs, the system calculates the vehicle's center pose. Calculating the vehicle's center pose means inferring the coordinates and orientation of the entire vehicle's center point in three-dimensional space based on the position and orientation information of the two identified support legs. This center pose is the direct basis for the robot's precise alignment operations, translating the vehicle's physical position and orientation into control commands that the robot can understand and execute. In one implementation, after identifying the geometric center points of the two support legs, the midpoint of these two center points can be taken as the vehicle's lateral center, and the vehicle's heading angle can be determined based on the direction of the line connecting the two support legs or their respective principal axis directions. In another implementation, if the support legs provide richer geometric information (such as cuboid boundaries), the vehicle's center pose can be calculated by fitting a geometric model of the support legs and then using the model's geometric center and principal axis directions, thus obtaining more accurate positioning and orientation information.

[0069] Finally, based on the calculated carrier center pose, the system further obtains the lateral deviation (y) and yaw angle deviation (θ) of the carrier relative to the forks. Lateral deviation (y) refers to the difference in horizontal distance between the centerline of the carrier and the centerline of the forks, reflecting the carrier's offset relative to the forks in the left-right direction. Yaw angle deviation (θ) refers to the angle between the longitudinal axis of the carrier and the longitudinal axis of the forks, reflecting the carrier's rotation angle in the horizontal plane. These two deviation parameters are the core inputs for the robot's fine-tuning operations; they directly quantify the degree of relative mismatch between the forks and the carrier, providing a clear direction and magnitude for subsequent precise alignment. In one implementation, after obtaining the carrier's center pose, the carrier pose is transformed to a coordinate system with the forks as the reference frame through coordinate transformation. Then, the translation in the y-axis direction is directly extracted from the transformed pose data as the lateral deviation, and the rotation angle around the z-axis is extracted as the yaw angle deviation. In another implementation, the precise position and orientation of the forks in the robot coordinate system can be pre-calibrated. After calculating the pose of the vehicle in the robot coordinate system, the lateral deviation and heading angle deviation of the vehicle relative to the forks can be directly obtained through simple vector subtraction and angle calculation.

[0070] Through the aforementioned technical solution, this application, in the picking operation, can extract stable and highly recognizable geometric features from complex 3D point cloud data by specifically identifying the end face of the carrier and further accurately identifying the two supporting legs of the carrier. This identification method, which focuses on key structural features, effectively avoids the problem of insufficient identification accuracy caused by irregularities on the carrier surface or environmental interference. By calculating the center pose of the supporting legs and then quantifying the lateral and yaw angle deviations of the carrier relative to the forks, accurate and reliable input parameters are provided for subsequent vehicle pose adjustments. This enables the dual-side fork narrow-aisle stacking robot to complete the alignment operation with the target carrier more accurately and efficiently, significantly improving the success rate and efficiency of the picking operation, reducing the need for manual intervention or multiple trial adjustments, thereby optimizing the smoothness and stability of the overall operation process.

[0071] The implementation plan for the unloading process lacks a precise identification method for high-level rack storage locations, leading to reliance on experience or multiple trial-and-error adjustments, resulting in low efficiency and potential safety hazards such as goods falling or colliding with the racks in high-risk stacking scenarios. To address this, this application further proposes that in step S4, when the operation is unloading, the 3D point cloud data is processed to identify the target storage location center. Specifically, this includes: based on prior information about the rack position and width, detecting point cloud data within the corresponding range and extracting rack column features, then calculating the storage location center pose, and finally obtaining the relative deviation between the center of the forklift and the target storage location center.

[0072] Specifically, when the operation type is confirmed as unloading, the system will activate the identification process for the target storage location center. Here, "processing the 3D point cloud data to identify the target storage location center when the operation is unloading" refers to the core task of accurately determining the position and orientation of the target storage location in 3D space when the robot performs the operation of placing the carrier from the forks to the rack storage location. This is crucial for ensuring the safe and accurate placement of the carrier, especially in narrow aisle, high-level stacking scenarios where extremely high precision is required to avoid collisions with racks or goods falling. This operation type confirmation can be obtained directly from the received operation task instructions or determined in real time through sensors on the robot itself (e.g., vision sensors identifying specific markings on the carrier or storage location).

[0073] During the recognition process, the system utilizes prior information about the shelf location and width. This prior information consists of known data about the shelf structure and layout pre-stored in the robot control system, such as the approximate location of the shelves in the warehouse coordinate system, the standard width and height of each storage location, and the cross-sectional dimensions of the shelf uprights. This information can be obtained through manual input, importing CAD drawing data from a warehouse management system (WMS), or through a one-time environmental scan modeling during the robot's initial deployment. By utilizing this prior information, the system can effectively narrow down the processing scope of point cloud data, thereby improving recognition efficiency and accuracy.

[0074] Subsequently, the system will "detect point cloud data within the corresponding range." This means that the raw 3D point cloud data acquired by the multi-line LiDAR will not be processed entirely. Instead, based on the aforementioned prior information and the robot's current pose, a local 3D region (such as a bounding box or region of interest) where the target storage location may exist will be defined. Points in the point cloud data located outside this region will be filtered out, thereby reducing the amount of data that needs to be processed, lowering computational complexity, and effectively eliminating irrelevant noise interference from the environment.

[0075] Based on this, the system will "extract shelf upright features." Shelf uprights are relatively stable and easily identifiable geometric elements in the shelf structure, typically appearing as vertical structural components with specific cross-sectional shapes, such as beams and columns. Extracting these features is crucial for accurately calculating warehouse location poses because they exhibit consistency across different warehouse locations and heights, serving as reliable reference points. Feature extraction can employ various methods. For example, the RANSAC (Random Sample Consensus) algorithm can be used to detect linear or planar features in the point cloud, and a set of points matching the upright features can be selected based on their orientation and dimensions. Alternatively, a deep learning model can be used, trained to directly identify and segment the shelf upright points from the point cloud data.

[0076] After successfully extracting the features of the shelf uprights, the system will then calculate the center pose of the storage location. By identifying the centerline or edge of one or more shelf uprights, and combining this with the structural dimensions of the shelf (such as the width of the storage location), the system can accurately calculate the center point of the target storage location. The orientation of the storage location (e.g., its orientation in the horizontal plane) can be determined by the normal vector of the upright or the direction of the line connecting the uprights. This calculation process can employ the least squares method or other optimization algorithms to match the extracted upright feature points with a preset shelf model, thereby solving for the optimal pose parameters of the storage location.

[0077] Ultimately, the system will "obtain the relative deviation between the center of the carrier on the forks and the center of the target storage location." This is the output of the entire recognition process, quantifying the three-dimensional positional deviation (e.g., lateral, longitudinal, and vertical offsets) and attitude deviation (e.g., differences in yaw, pitch, and roll angles) between the current carrier on the forks and the target storage location. This relative deviation information is the direct input for the robot to perform subsequent precise pose adjustments, ensuring that the carrier can be accurately placed in the target storage location. This deviation can be obtained by performing coordinate transformation and comparison between the center pose of the carrier on the forks (which can be obtained from sensors on the forks or by combining preset carrier dimensions with the fork pose) and the calculated center pose of the target storage location.

[0078] Through the above technical solution, this application utilizes prior information about the shelf position and width during unloading operations to limit the processing range of 3D point cloud data to the corresponding area, thereby efficiently detecting and extracting shelf column features. Based on these stable and reliable column features, the center pose of the target storage location can be accurately calculated. Furthermore, the relative deviation between the center of the forklift and the center of the target storage location is calculated, providing a precise and real-time basis for subsequent robot pose adjustments. This technical solution effectively solves the problems of low alignment efficiency and safety hazards caused by the lack of accurate storage location identification in high-level stacking unloading operations. By accurately identifying the storage location, the robot can achieve precise alignment in one go, significantly reducing the number and time of trial adjustments, and improving the efficiency and continuity of unloading operations. Simultaneously, accurate deviation information greatly reduces the risk of goods falling or colliding with the shelf, improving the safety of high-level operations.

[0079] In a further implementation, in step S3, the control of adjusting the forks to a preset recognition height based on the job type is a closed-loop control based on fork height information fed back by a drawstring encoder. Specifically, before performing 3D point cloud scanning, the system determines a preset recognition height based on the current job type (e.g., picking or unloading). This preset recognition height is set according to the optimal scanning position of the target object to be scanned in different job scenarios (e.g., the end face of the carrier, the carrier on the forks, or the shelf location). To ensure that the forks can be accurately adjusted to this preset recognition height, this application uses a drawstring encoder to provide real-time feedback on the fork height information. A drawstring encoder is a displacement sensor that accurately obtains the vertical position of the forks by measuring the extension and retraction length of the drawstring.

[0080] Based on real-time fork height information fed back by a drawstring encoder, the system performs closed-loop control of the fork lifting action. Closed-loop control is a feedback control system that continuously compares the actual fork height with a preset recognition height and dynamically adjusts the lifting mechanism motor based on the deviation between the two, so that the actual fork height is as close as possible to the preset value. For example, a proportional-integral-derivative (PID) controller can be used to calculate the control quantity based on the height error, driving the lifting mechanism motor for precise adjustment; alternatively, a fuzzy logic controller can be used to achieve smooth and precise adjustment of the fork height through fuzzy rules and reasoning.

[0081] Through the above technical solution, this application introduces a closed-loop control mechanism based on pull-cord encoder feedback during the process of adjusting the forks to the preset recognition height. Specifically, the system determines the target recognition height according to the current operation type and uses the pull-cord encoder to monitor the actual height of the forks in real time. The controller continuously compares the actual height with the target height and dynamically adjusts the lifting and lowering action of the forks according to the deviation between the two. This closed-loop feedback mechanism can effectively compensate for the influence of mechanical transmission errors, load changes, and external interference on the fork height, ensuring that the forks can accurately and stably stay at the preset recognition height. This significantly improves the accuracy and reliability of 3D point cloud scanning, provides high-quality raw data for subsequent pose recognition, thereby improving the overall operation accuracy and safety, and effectively reducing the risk of collisions or operation failures caused by height deviations.

[0082] In a further embodiment, before step S5 is executed, a judgment step is included: determining whether the deviation information calculated in step S4 is within a preset safety adjustment threshold; if yes, then step S5 is executed; if no, an anomaly is reported and the current work process is terminated. This judgment step aims to verify the validity or security of critical data received by the system to determine the execution path for subsequent operations. Its core is to introduce a decision point to avoid continuing potentially risky operations under unsafe or uncertain conditions. This judgment step can be implemented through a software logic module that receives deviation information as input and outputs a decision result, such as a Boolean value or status code, based on built-in judgment rules. Alternatively, it can be implemented through hardware logic circuits or a programmable logic controller (PLC) to monitor and judge input signals in real time and trigger different output actions based on preset logical conditions.

[0083] The preset safety adjustment thresholds refer to a set of numerical limits pre-set during the system design or debugging phase to measure the acceptable range of deviation information. These thresholds are determined comprehensively based on factors such as the physical characteristics of the robot itself, the limitations of the working environment, the characteristics of the goods, and safety regulations, aiming to ensure the safety and stability of the robot's adjustment actions. These thresholds can be stored in the non-volatile memory of the control system as configuration parameters, allowing modification during system maintenance or upgrades. Simultaneously, the thresholds can also be dynamically configured and loaded through host computer software or a human-machine interface to flexibly adjust according to different operating scenarios or goods types. When the judgment result shows that the deviation information exceeds the safety adjustment threshold, the system will report the anomaly and suspend the current operation process. This is an emergency response mechanism designed to promptly stop potentially dangerous operations, prevent accidents, and notify operators or the monitoring system through the anomaly reporting mechanism for manual intervention or troubleshooting. Anomaly reporting can send anomaly codes or status information to the host computer monitoring system or a remote server via a communication interface, while simultaneously displaying warning information on the robot's local display screen. Suspending the operation process is achieved by stopping all relevant motion control commands and placing the robot in a safe stop state. In addition, anomaly reporting can also be achieved by issuing alarms to on-site personnel through audible and visual alarm devices and sending notifications to mobile terminals via wireless communication modules. The operation suspension procedure can include measures such as immediately cutting off the drive power and activating the emergency braking system to ensure the robot stops quickly and remains stable.

[0084] By introducing a judgment step before execution in step S5, this application can perform real-time verification of the deviation information calculated in step S4. When the deviation information is detected to exceed the preset safety adjustment threshold, the system can immediately report the anomaly and stop the current operation process, thereby effectively avoiding operational risks such as collisions and goods falling that may be caused by blindly performing pose adjustments when the deviation is too large. This mechanism ensures that the robot only makes fine adjustments within a safe and controllable range, significantly improving the safety and robustness of the dual-side fork narrow aisle stacking robot in high-level storage and retrieval operations. For example, when retrieving goods, if the detected deviation of the carrier end face or support leg pose is too large, it may mean that the carrier is placed abnormally or the sensor misidentifies. Stopping the operation at this time can prevent the forks from being forcibly inserted, causing damage to the carrier or goods to scatter. When unloading goods, if the detected relative deviation between the target storage location center and the carrier on the forks exceeds the safe range, it may indicate that the storage location is occupied or the rack structure is abnormal. Timely stoppage can prevent the forks from colliding with the rack. This proactive safety verification mechanism enables robots to achieve more reliable and safer automated operation in complex, narrow-channel, high-position working environments, reducing the need for human intervention and lowering potential economic losses and safety accident risks.

[0085] In a further preferred embodiment, a safety initialization step is included before step S2: determining whether the forks are in a preset safe reset position; if not, prioritizing the movement of the forks to the safe reset position. This safety initialization step aims to ensure that the robot's critical actuators (forks) are in a safe and controllable initial state before performing major work processes (such as navigating to a preparatory work point). Its function is to provide a stable safety baseline for subsequent operations, avoiding potential risks caused by an improper initial state. Specifically, this step can be an independent software module or subroutine, called when the robot starts or receives a new task, responsible for performing a series of pre-check and adjustment operations; or, it can be integrated into the startup sequence of the robot's main controller as part of the system self-check, forcibly executed before the system is ready.

[0086] The purpose of determining whether the forks are in the preset safe reset position is to verify whether the current state of the forks meets the pre-set safe initial position requirements. The preset safe reset position typically refers to the state where the forks are at their lowest position, fully retracted (without lateral movement), and perpendicularly aligned with the mast, to minimize their space occupation and reduce the risk of collision. This determination can be made by detecting the physical position of the forks using limit switches or proximity sensors installed on the fork mechanism. When the forks touch a specific position, the sensor sends a signal indicating whether they have reached or not reached the safe reset position. Alternatively, precision position sensors such as draw-wire encoders or absolute encoders can be used to acquire real-time data on the lifting height and lateral movement of the forks, and compare this data with the numerical range of the preset safe reset position to determine whether the forks are in that position. When the determination result shows that the forks are not in the preset safe reset position, the purpose of prioritizing the movement of the forks to the safe reset position is to forcibly adjust the forks to that safe position. Here, "prioritize" means that this operation has the highest execution priority; other work processes (such as vehicle navigation) will be suspended or blocked before the forks are reset. In practice, the main controller can send control commands to the lifting and lateral movement mechanisms of the fork mechanism, driving the motor or hydraulic system to lower the forks to the lowest point and retract them to the center position until the position sensor reports that the forks have reached the safe reset position. Alternatively, in conjunction with a closed-loop control strategy, the main controller, based on real-time feedback from position sensors (such as a drawstring encoder), precisely controls the speed and direction of the forks to ensure they smoothly and accurately reach the preset safe reset position, continuously monitoring the process until the reset is complete.

[0087] During unloading and stacking operations, the lack of a real-time obstacle detection mechanism in the target stacking area before pose adjustment may lead to unknown objects not being identified, thus causing collision risks. To address this, this application further proposes that when the operation is unloading and stacking, a safety verification step is included before step S5: based on the point cloud data obtained from scanning the shelf location, the point cloud density in the target stacking area is checked; if unknown obstacles exist, the verification fails to prevent collisions.

[0088] Specifically, the limitation "when the operation is unloading and stacking" clarifies the application scenario of the safety verification steps. Unloading and stacking operations typically involve placing goods on high-level racks, which are highly complex operations, and collisions could lead to goods falling or damage to the rack structure, posing a significant risk. Therefore, introducing additional safety verification mechanisms is necessary for such high-risk operations. One implementation is that when receiving operation task S1, the task information explicitly includes the operation type (e.g., picking, unloading, stacking, translation, etc.), and the control system determines whether the current operation is "unloading and stacking" based on this information. Another implementation is that the system infers the operation type based on the height information of the target location and whether the forks are currently carrying goods. For example, if the target location is high and the forks are carrying goods, it is determined to be an unloading and stacking operation.

[0089] The safety check performed "before step S5" aims to ensure a thorough safety assessment of the target area before the robot performs any physical movements that could lead to collisions. This pre-check mechanism effectively avoids potential hazards caused by subsequent pose adjustments. One implementation is to treat the safety check step as a precondition judgment module in the control flow; step S5 is only allowed to proceed if the module's check result is satisfactory. Another implementation is to set up a state machine within the control system. When the system is in the "preparing pose adjustment" state, it forcibly jumps to the "safety check" state. After the check is completed, the system decides whether to enter the "performing pose adjustment" state based on the result.

[0090] This "safety verification step" is the core of this solution. Its purpose is to proactively detect potential hazards, such as unknown obstacles, within the work area before the robot performs high-risk tasks. This differs from traditional passive obstacle avoidance; it is a proactive prevention mechanism that significantly improves operational safety. One implementation method is for this step to be coordinated and executed by the central controller in the navigation and control module. This controller acquires point cloud data by calling the pose recognition module and performs data analysis. Another implementation method is to set up an independent safety detection module specifically responsible for executing this verification step and reporting the verification results back to the central controller.

[0091] This verification step is based on point cloud data obtained from scanning the shelf storage locations. Point cloud data is a collection of discrete points in three-dimensional space, which can accurately describe the shape and position of an object. Here, the point cloud data acquired in step S3 by scanning the target area with a multi-line LiDAR is utilized, particularly the data obtained by scanning the forks and target storage locations during unloading. Reusing the acquired data avoids repeated scanning and improves efficiency. One implementation is to store the raw point cloud data acquired by the multi-line LiDAR in memory after step S3, for direct access by the security verification step. Another implementation is to perform preliminary processing and region division on the acquired point cloud data in step S3, extracting the point cloud data related to the target shelf storage location and marking it as "shelf storage location point cloud data" for quick access in subsequent security verification steps.

[0092] By checking the point cloud density within the target stacking area, the presence of unknown obstacles can be determined. Point cloud density refers to the number of points per unit volume or area. Given the known shelving structure and the point cloud distribution characteristics of the expected stacked objects, the point cloud density within the target stacking area should be within an expected range. If an abnormally high point cloud density appears in this area, it may indicate the presence of an unexpected object, i.e., an unknown obstacle. One implementation method is to define a three-dimensional voxel mesh, dividing the target stacking area into several smaller voxels, and then counting the number of points within each voxel. If the number of points in a voxel is significantly higher than a preset threshold, the area is considered abnormal. Another implementation method is to use a neighborhood search-based approach. For each point within the target stacking area, the number of its neighboring points within a certain radius is calculated. If the neighborhood density of a point is abnormally high, it may indicate the presence of an obstacle.

[0093] Ultimately, "if an unknown obstacle exists, the verification fails to prevent a collision." This is the final judgment and response mechanism in the safety verification process. Once an unknown obstacle is detected, the system immediately determines that the verification has failed and takes measures to prevent the robot from performing actions that could lead to a collision. This mechanism ensures that the robot can stop in time in the face of potential danger, thereby avoiding accidents. One implementation is that when the verification fails, the main controller immediately issues a stop command, stops the current work process, and reports the abnormal information. Another implementation is that, in addition to stopping the work process, the system can also control the forklift mechanism or the vehicle body to move to a preset safe position to further mitigate risks.

[0094] Through the above technical solution, this application introduces an active safety prevention mechanism for high-level unloading and stacking operations of a dual-fork narrow-aisle stacking robot. This mechanism fully utilizes the acquired 3D point cloud data, avoiding the need for additional sensor configurations and achieving efficient integration. Through intelligent analysis of the point cloud density of the target stacking area, it can detect unknown obstacles that are difficult to detect using traditional methods in real time and accurately, significantly improving operational safety. When a potential risk is detected, the system can promptly stop the operation, avoiding serious accidents such as cargo falling, shelf damage, or equipment collisions that may be caused by pose adjustments (S5). Especially in high-level operation scenarios, it greatly reduces operational risks and ensures the safety of equipment, goods, and personnel.

[0095] Example 2:

[0096] like Figure 4 As shown in Figure 5, this application provides a double-sided fork narrow aisle stacking robot control system for executing the control method described in Example 1. This system addresses the problem of efficient, precise, and continuous storage, retrieval, and stacking of goods on both sides within an ultra-narrow aisle, while ensuring the safety and stability of high-level operations. This application also discloses a double-sided fork narrow aisle stacking robot control system for executing the aforementioned control method, including:

[0097] Vehicle body 1;

[0098] Mast 2 is mounted on vehicle body 1;

[0099] The fork mechanism 3 is mounted on the mast 2 and includes a lifting mechanism 31 for driving the forks to rise and fall, and a lateral movement mechanism for driving the forks to move laterally to the left or right.

[0100] The pose recognition module includes a multi-line lidar 41;

[0101] The navigation and control module includes a main controller and a visual positioning unit for vehicle positioning and path navigation;

[0102] The feedback module includes a drawstring encoder for detecting the absolute position of the fork lifting and lateral movement;

[0103] The main controller is connected to the pose recognition module, navigation and control module, feedback module and forklift mechanism 3.

[0104] The core innovation of this embodiment lies in combining the multi-line LiDAR 41 with the draw-wire encoder in a closed-loop control manner, and having the main controller coordinate the real-time data interaction between the pose recognition module, navigation and control module, and feedback module, thereby achieving multi-degree-of-freedom coordinated fine-tuning of the vehicle body 1 and the forks within an ultra-narrow channel. Specifically, the pose recognition module uses the multi-line LiDAR 41 to scan the target area, generating high-density three-dimensional point cloud data to identify the precise three-dimensional pose of the dynamic vehicle or high-level storage location in real time; the feedback module accurately detects the absolute position of the fork lifting and lateral movement through the draw-wire encoder, providing closed-loop feedback to the main controller; based on the deviation information output by the pose recognition module and the position data from the feedback module, the main controller coordinates the fine-tuning of the pose of the vehicle body 1 and the lateral movement direction of the forks, ensuring micron-level precision alignment in high-level operations.

[0105] Through the above technical solution, this application effectively solves the problem of efficient dual-sided storage and retrieval under narrow aisle space constraints. The vehicle body 1 does not need to repeatedly drive out of the aisle to adjust its direction; the forklift mechanism 3 can complete dual-sided operations by directly moving laterally left or right through the side-shifting mechanism, significantly improving operational continuity. Simultaneously, the closed-loop control mechanism of the multi-line lidar 41 and the pull-rope encoder, combined with real-time collaborative fine-tuning by the main controller, overcomes the positioning deviation caused by insufficient sensing in traditional systems during high-level stacking, greatly improving storage and retrieval accuracy and safety. Overall, while ensuring structural compactness, this system achieves efficient and accurate storage and retrieval of goods on both sides within ultra-narrow aisles, reducing the risk of goods falling and meeting the continuous operation requirements of modern logistics in high-density warehousing scenarios.

[0106] In some embodiments described above in this application, a control system including a multi-line lidar 41 is proposed for scanning a target area to obtain three-dimensional point cloud data. However, during its implementation, the installation position of the lidar may cause the field of view to be obstructed by the forks or mast 2 structure, making it impossible to effectively identify key target parts when operating at different heights of the mast 2. For example, at a low position, the front face of the ground carrier and the support legs cannot be clearly scanned; at a high position, the support legs of the carrier on the forks and the uprights and beams of the rack storage location cannot be accurately identified, thus affecting the position and pose recognition accuracy and increasing adjustment errors and collision risks. In response, this application proposes a scheme to optimize the installation and motion configuration of the multi-line lidar 41. Specifically, the multi-line lidar 41 is fixedly installed at the bottom of the fixed crossbeam 21 between the two forks on the mast 2. It is configured to rise and fall synchronously with the mast 2 and not move with the left and right lateral movement of the forks. The installation position of the multi-line lidar 41 satisfies the following conditions: when the mast 2 is in a low position, its scanning field of view is not obstructed by the forks, so as to identify the front face of the ground carrier and the support legs; when the mast 2 is in a high position, its scanning field of view can identify the support legs of the carrier on the forks and the uprights and crossbeams of the racking position.

[0107] In detail, the multi-line lidar 41 is a sensor capable of simultaneously emitting and receiving multiple laser beams, acquiring three-dimensional spatial information of the target object by measuring the flight time of the laser beams. It is fixedly mounted on the bottom of the fixed crossbeam 21 between the two forks on the gantry 2, aiming to provide a relatively centered and stable observation angle, effectively avoiding direct obstruction of the lidar's scanning field of view by the forks during storage and retrieval operations. This fixed installation can be achieved by welding or bolting, firmly fixing the base of the multi-line lidar 41 to the bottom of the fixed crossbeam 21 of the gantry 2, ensuring its structural stability during robot operation; alternatively, a dedicated mounting bracket can be designed, which is connected to the bottom of the fixed crossbeam 21 of the gantry 2 by bolts or clips, and the multi-line lidar 41 is then mounted on this bracket to achieve precise positioning and convenient maintenance.

[0108] Furthermore, the multi-line lidar 41 is configured to rise and fall synchronously with the gantry 2, meaning its height position remains consistent with the vertical movement of the gantry 2. This configuration ensures that the lidar's scanning plane is always matched to the current operating height. Whether identifying ground vehicles at a low position or identifying shelving locations at a high position, the lidar can be at a suitable observation height, thereby acquiring effective three-dimensional point cloud data of the target area. In terms of implementation, the multi-line lidar 41 can be directly mounted on the lifting part of the gantry 2, making it move as part of the gantry 2; alternatively, a mechanical linkage mechanism, such as a slide rail, chain, or gear transmission system, can be used to connect the multi-line lidar 41 to the lifting mechanism of the gantry 2, ensuring synchronous vertical movement of both.

[0109] Meanwhile, the multi-line lidar 41 does not move with the lateral movement of the forks, meaning its horizontal position remains unchanged during the lateral movement of the forks. This characteristic avoids dynamic obstruction of the lidar's field of view or displacement of the scanning reference point that may occur during the lateral movement of the forks, ensuring the stability and consistency of the scanning data. This can be achieved by mounting the multi-line lidar 41 on a fixed part of the mast 2, such as the outer mast of the mast 2 or a crossbeam that does not move with the forks, ensuring its decoupling from the lateral movement mechanism; or by designing an independent mounting structure that is fixed to the main body of the mast 2 and independent of the lateral movement mechanism of the forks, so that the horizontal position of the multi-line lidar 41 is not affected by the lateral movement of the forks.

[0110] Furthermore, the installation position of the multi-line lidar 41 is optimized to meet specific field-of-view requirements. When the mast 2 is in a low position, its scanning field of view is not obstructed by the forks, and it can clearly identify the front face and support legs of the ground carrier. This is achieved by performing optical simulation and mechanical structure optimization during the design phase to determine the precise installation height and front-to-back position of the multi-line lidar 41 relative to the forks and mast 2, so that when the mast 2 is in a low position, the scanning sector can avoid physical obstruction by the forks; or, after actual installation, by adjusting the pitch angle of the lidar or the fine-tuning mechanism of the mounting bracket, and combining with actual scanning tests, its field of view range in the low position is verified and optimized. When the mast 2 is in a high position, its scanning field of view can identify the support legs of the carrier on the forks and the uprights and beams of the racking position. This can be achieved by selecting a model with a sufficient vertical field of view when choosing the multi-line lidar 41, and by combining its installation position to ensure that when the mast 2 is in a high position, it can scan downwards to the carrier support legs on the forks and upwards to the uprights and beams of the racking position; or by adjusting the installation angle of the multi-line lidar 41 so that when the mast 2 is raised to a high position, its scanning range can simultaneously cover the bottom features of the carrier on the forks and the top structure of the racking position.

[0111] Through the above technical solution, this application effectively solves the problem of decreased pose recognition accuracy and increased operational risks caused by the obstruction of the LiDAR field of view by the forks or mast 2 structure during high-level stacking operations in narrow aisles. The multi-line LiDAR 41 is fixedly installed at the bottom of the fixed crossbeam 21 between the two forks on the mast 2, and its movement is synchronized with the lifting and lowering of the mast 2, but does not move with the lateral movement of the forks. This configuration ensures that the LiDAR is always in a centrally symmetrical and stable observation position, avoiding interference with the field of view when the forks move laterally, and guaranteeing the continuity and accuracy of the scanning data. More importantly, by precisely designing its installation position, it ensures that when the mast 2 is in a low position, the LiDAR's scanning field of view can completely avoid the obstruction of the forks, clearly identifying the front face and support legs of the ground carrier, thus providing accurate pose information for low-level picking. When the mast 2 is in a high position, its scanning field of view can effectively cover the support legs of the carrier on the forks and the columns and beams of the target racking location, providing key structural feature data for high-level unloading or stacking operations. This unobstructed, high-precision scanning capability across the entire height range significantly improves the quality and reliability of the 3D point cloud data acquired by the pose recognition module. This, in turn, enhances the accuracy of multi-degree-of-freedom collaborative fine-tuning of the vehicle body and forks, effectively reducing adjustment errors and collision risks in high-level operations. It ensures the operational efficiency and safety of the dual-side fork narrow-channel stacking robot in complex high-level scenarios.

[0112] In a further embodiment, this application proposes a visual positioning unit including a camera positioned facing the ground for recognizing navigation code strips laid on the ground. Specifically, the visual positioning unit is a key component in the control system of the dual-sided fork narrow-aisle stacking robot. Its main function is to determine the precise position and posture of the robot body through visual perception of the environment, thereby providing basic data for the navigation and path planning of the vehicle body 1. Its implementation methods may include, but are not limited to: matching and positioning images acquired by visual sensors with a pre-built environmental map; or employing Simultaneous Localization and Mapping (SLAM) technology to simultaneously perform self-localization and environmental map construction in an unknown environment.

[0113] A ground-facing camera is a specially designed image acquisition device whose lens optical axis is typically perpendicular or nearly perpendicular to the ground, aiming to focus on acquiring visual information about the ground area. This setup effectively reduces interference from complex environmental factors such as high-level shelving and aisle walls on the positioning images, improving the stability and clarity of image acquisition. The camera can be a high-resolution industrial-grade CMOS camera to ensure the clarity of image details; or it can be a camera with a global shutter to avoid image blurring or distortion during high-speed robot movement, thus ensuring the accuracy of the image data.

[0114] Navigation strips are visual markers with specific codes or patterns pre-laid on the ground in the work area. These strips serve as fixed reference points for robot navigation, carrying precise position information or path guidance. The process of identifying navigation strips typically involves image processing techniques, such as image segmentation and feature extraction algorithms (e.g., SIFT, SURF) to identify the geometry and internal patterns of the strips; or using deep learning models to analyze camera-captured images in real time, directly resolving the absolute coordinate information represented by the strips. By identifying these standardized, high-contrast strips, the robot can obtain high-precision absolute position information, effectively correct accumulated errors, and ensure accurate movement along a pre-defined path.

[0115] Through the above technical solution, the visual positioning unit, using a camera positioned facing the ground, focuses on recognizing navigation code strips laid on the ground, effectively solving the problem of insufficient positioning accuracy when navigating in ultra-narrow aisles. Specifically, the ground-facing camera orientation minimizes interference from complex environmental factors such as high-level shelves and aisle walls, ensuring the stability and effectiveness of the acquired images. Simultaneously, the navigation code strips serve as pre-set visual reference points with high-precision position information, enabling the robot to obtain accurate absolute position information by recognizing these standardized markers, effectively correcting cumulative errors that may arise from inertial navigation or odometer readings. This high-precision positioning capability ensures that vehicle 1 can navigate to the preparatory work point with extreme accuracy in step S2, providing a solid foundation for subsequent steps such as fork height adjustment in step S3, 3D point cloud data processing in step S4, and vehicle 1 pose adjustment in step S5. This significantly improves the overall operational accuracy and efficiency of the dual-fork narrow-aisle stacking robot in high-level, narrow-aisle working environments, reducing the risk of collisions and operational interruptions caused by positioning deviations.

[0116] In addition, the aforementioned control system also includes a safety detection module. This module comprises microswitches and photoelectric sensors located at the fork tips, as well as anti-collision contact edges and obstacle avoidance lidar located around the vehicle body 1. Specifically, the safety detection module is an integrated hardware and software unit designed to monitor the robot and its operating environment in real time, identify potential hazards, and trigger corresponding safety responses. This module can be composed of a standalone microcontroller or programmable logic controller (PLC), responsible for receiving signals from various sensors and making judgments and outputting control commands according to preset safety logic; alternatively, it can also be part of the main controller, implementing its functions through specific software threads or hardware interfaces, operating in parallel with other control logic to ensure the priority of safety functions. Its core function is to provide comprehensive safety assurance, preventing equipment damage, cargo loss, and personnel injury.

[0117] The microswitch located at the fork tip detects whether the fork tip is in contact with an object. When the fork tip touches an obstacle, the microswitch is immediately triggered, sending a signal to the control system to prevent damage caused by physical collision. This microswitch can be a mechanical microswitch, changing the circuit state through physical contact, and features rapid response and simple structure; alternatively, it can be a piezoresistive or thin-film pressure sensor, whose resistance changes when the fork tip is subjected to pressure, thus being detected. The photoelectric sensor detects the presence or distance of an object by emitting and receiving light beams. Located near the fork tip, it is used for non-contact detection of obstacles ahead, providing early warning and preventing the forks from accidentally contacting goods or shelves during movement. This photoelectric sensor can be a through-beam photoelectric sensor, with the transmitter and receiver mounted on opposite sides of the fork tip, determining the presence of an obstacle when the light path is blocked; it can also be a diffuse reflection photoelectric sensor, with the transmitter and receiver integrated, determining the presence of an object by detecting light reflected from its surface; or it can be a laser rangefinder sensor, accurately obtaining the obstacle distance by measuring the laser's round-trip time.

[0118] The anti-collision edges installed around the vehicle body 1 provide physical cushioning and contact detection. When the vehicle body 1 collides slightly with an external object, it can absorb part of the impact force and immediately send a collision signal to the control system to trigger an alarm or emergency braking. These anti-collision edges can be flexible edges made of rubber or foam material with integrated microswitches or pressure sensors, triggering a signal when the edge is deformed under pressure; alternatively, they can be elastic materials with a conductive coating, forming a circuit and sending a signal when the conductive layer contacts under pressure.

[0119] Obstacle avoidance lidar constructs a 3D or 2D point cloud map of the surrounding environment by emitting laser beams and measuring reflection time. This allows for real-time identification of the location, size, and distance of obstacles, ensuring that the vehicle 1 does not collide with surrounding objects when moving through narrow passages. This obstacle avoidance lidar can be a 2D lidar, typically mounted around the vehicle 1, scanning the horizontal plane to detect obstacles around the vehicle 1; it can also be a 3D lidar, capable of acquiring richer environmental information for more complex obstacle avoidance and environmental modeling; or it can be a solid-state lidar, achieving electronic scanning rather than mechanical rotation, offering higher reliability and a smaller size.

[0120] Through the above technical solution, this application introduces a multi-layered, multi-dimensional safety detection mechanism into the control system of a dual-fork narrow-aisle stacking robot. Microswitches and photoelectric sensors installed at the fork tips can perform real-time, proactive detection of close-range contact and potential obstacles during storage and retrieval operations, effectively preventing collisions between the forks and goods or shelves, and avoiding goods falling or equipment damage. Simultaneously, anti-collision edges and obstacle avoidance lidar installed around the vehicle body 1 provide comprehensive safety assurance for the movement of the vehicle body 1 within narrow aisles. The anti-collision edges provide buffering and trigger alarms in the event of slight contact, while the obstacle avoidance lidar can scan the environment in real time, identify obstacles, and enable the robot to plan its path in advance or brake suddenly, thereby avoiding collisions between the vehicle body 1 and aisle walls or other equipment. These safety features work together to cover the key risk points of fork operation and vehicle body 1 movement, significantly improving the safety and reliability of the robot in high-position, narrow-aisle operating environments, and effectively reducing operational risks and potential losses.

[0121] Based on the aforementioned solution, this application further optimizes the perception architecture of the navigation and control module to address the challenges of high-precision positioning, real-time obstacle avoidance, and vehicle stability control in complex environments within ultra-narrow channels. Specifically, the navigation and control module also includes: a navigation LiDAR 42 mounted on the top of the vehicle body 1, serving as the main sensor for SLAM mapping and global positioning; a blind-spot LiDAR 43 mounted on the front of the vehicle body 1, used to detect low-lying obstacles and obstacles in the blind spots of the navigation LiDAR 42; and a high-precision IMU 44 integrated inside the vehicle body 1, used to measure vehicle pitch, roll, and angular velocity.

[0122] In detail, the navigation LiDAR 42, mounted on the top of the vehicle body 1, is the core sensor for the robot to achieve autonomous localization and mapping. It typically employs a mechanical rotating or solid-state LiDAR with 16 or more lines, featuring a 360° horizontal field of view and a large vertical field of view. As the primary sensor for SLAM (Simultaneous Localization and Mapping), the navigation LiDAR 42 scans the surrounding environment at high speed, acquiring dense 3D point cloud data. The system utilizes this point cloud data to match it with a pre-built high-precision point cloud map, achieving centimeter-level localization of the robot in the global coordinate system. Furthermore, it can build or optimize the environmental map in real time in unknown areas or when the map is updated. Mounted at the highest point of the vehicle body 1, it aims to maximize the scanning field of view and reduce occlusion from the vehicle's own structure (such as the mast 2 and the high-lift forks), ensuring stable capture of features from shelves on both sides of the aisle, the ceiling, and distant environmental characteristics, providing a reliable environmental model for path planning and global navigation.

[0123] The blind spot filler lidar 43, installed at the front of vehicle body 1, primarily addresses the blind spot problem of the navigation lidar 42. Due to its high mounting position, the navigation lidar 42's scanning beam may create a blind spot near the ground at the front of vehicle body 1, failing to effectively detect low-lying, suddenly appearing obstacles (such as scattered pallets, small cargo boxes, ground protrusions, etc.). The blind spot filler lidar 43 typically employs a 2D or 3D lidar with a narrow vertical field of view but a high scanning frequency, and is mounted low and facing downwards. It is specifically designed to scan the near-field, low-height area in front of vehicle body 1, enabling timely detection of obstacles entering this area. Simultaneously, it also covers other static blind spots caused by the navigation lidar 42's mechanical structure or mounting position (such as areas very close to the sides of the vehicle body). The data from the blind spot filler lidar 43 is fused with the data from the navigation lidar 42, providing the obstacle avoidance module with a blind-spot-free, full-coverage near-field environmental perception capability, greatly enhancing the robot's safety when navigating congested, dynamic, narrow passages.

[0124] The high-precision IMU44 (Inertial Measurement Unit) integrated inside the vehicle body 1 is a key component for improving navigation accuracy and vehicle motion control stability. It typically includes a three-axis gyroscope and a three-axis accelerometer, enabling continuous measurement of the vehicle body 1's angular velocity and linear acceleration in three axes (pitch, roll, and yaw) without relying on external signals. In navigation and positioning, the data from the high-precision IMU44 is tightly or loosely coupled and fused with data from the navigation lidar 42 and visual positioning units (such as cameras). The high-frequency (typically several hundred hertz) motion increment information provided by the IMU44 effectively smooths and predicts changes in vehicle pose within the scanning interval by the lidar or visual positioning system, compensating for positioning jumps or losses caused by vehicle bumps, slippage, or brief perception failures (such as loss of lidar features due to traversing reflective surfaces), thus providing a more continuous, stable, and smooth pose estimation. In terms of vehicle motion control, the pitch and roll angles measured in real time are used to evaluate the vehicle's attitude stability when lifting heavy objects, turning at high speed, or on uneven ground, providing feedback to the control algorithm to prevent overturning; the measured angular velocity is used to achieve more precise heading control.

[0125] Through the above technical solutions, this application constructs a multi-layered, multi-sensor fusion advanced perception and navigation system. The navigation lidar 42 acts as the "eagle eye," providing global, macroscopic environmental perception and high-precision positioning; the blind spot lidar 43 acts as the "tentacles," focusing on detecting nearby, low-lying obstacles in blind spots to ensure driving safety; and the high-precision IMU 44 acts as the "inner ear vestibule," providing high-frequency, stable proprioceptive motion perception to enhance the system's robustness and control quality. These three components work in conjunction with the aforementioned visual positioning unit (which identifies ground code strips) to form a complete technical closed loop of "global absolute positioning (visual code strip + lidar SLAM) + local relative positioning and obstacle avoidance (multi-lidar fusion) + high-frequency motion prediction and compensation (IMU)." This solution effectively addresses the issues of insufficient reliability of single-sensor positioning, blind spots in perception, and inaccurate motion state estimation in warehouse environments characterized by ultra-narrow, high-level, complex lighting, and potential dynamic obstacles. This ensures that the dual-fork narrow-channel stacking robot can achieve safe, accurate, and efficient autonomous navigation and operation in all weather conditions and scenarios, serving as a crucial foundation for achieving its core intelligent and unmanned high-level storage and retrieval functions.

[0126] The following example will provide a more detailed explanation of the above technical solution:

[0127] In a high-density automated warehouse, there are extremely narrow aisles with a width of 1.6 meters and racks with a height of up to 10 meters. A double-sided forklift narrow aisle stacker robot receives a task that requires the robot to first retrieve a carrier from a high-bay rack location (e.g., 5 meters high) on the left side of the aisle and then place it in another high-bay rack location (e.g., 6 meters high) on the right side of the aisle.

[0128] Before the robot begins its task, the system performs a safety initialization. First, it checks if the forks are in the preset safety reset position. If the forks are not in this position, for example, if they remained at a higher position after the last operation, the system will prioritize controlling the lifting mechanism 31 to lower the forks and the lateral movement mechanism to move the forks laterally to the safety reset position. This ensures the safety of subsequent navigation and operations, preventing accidental collisions during movement within the aisle.

[0129] Subsequently, based on the task information, the robot identifies the navigation code tape laid on the ground through the visual positioning unit in the navigation and control module, and, in conjunction with the main controller, performs path planning and vehicle navigation to accurately drive towards the designated pickup point. This designated pickup point is a specific location in the aisle in front of the target pickup location, ensuring that the robot can perform subsequent precise alignment operations at this location.

[0130] Once the vehicle body 1 navigates to the designated pickup point and comes to a stable stop, the robot enters the precise recognition and alignment phase. The main controller adjusts the fork mechanism 3 to the preset recognition height based on the current task type (pickup). This height adjustment is a closed-loop control based on fork height information fed back from the pull-cord encoder in the feedback module, ensuring the forks precisely stop at a height sufficient to effectively scan the target area. For example, for a 5-meter-high pickup bay, the forks will be precisely raised to a height slightly below the bottom of the vehicle.

[0131] After the forks are adjusted into position, the multi-line LiDAR 41 in the pose recognition module begins scanning the target area. Since it's a picking operation, the LiDAR scans the target vehicle and the forks below it. The multi-line LiDAR 41 is fixedly installed on the mast 2 at the bottom of the fixed crossbeam 21 between the two forks. Its installation position has been optimized to ensure that when the mast 2 is in a high position, its scanning field of view can clearly identify the support legs of the vehicle on the forks and the uprights and crossbeams of the racking position. Furthermore, it does not move with the lateral movement of the forks, ensuring the stability and accuracy of the scan.

[0132] After the lidar acquires 3D point cloud data of the target area, the pose recognition module processes this data. When the operation is picking up goods, the system processes the point cloud data to identify the end face of the vehicle, specifically identifying the two support legs of the vehicle and calculating their center pose. By analyzing the geometric features of the support legs, the system can accurately calculate the lateral deviation (y) and yaw angle deviation (θ) of the vehicle relative to the forks. This process avoids the coarse estimation of the vehicle position or reliance on manual experience adjustments in traditional methods, significantly improving the accuracy of the initial alignment.

[0133] After calculating the deviation information, the system first determines whether this deviation information is within the preset safety adjustment threshold. For example, if the calculated lateral deviation or heading angle deviation is too large, exceeding the range that the robot can safely and effectively adjust, the system will judge it as abnormal, immediately report it, and terminate the current operation process to prevent potential collisions or operation failures. If the deviation is within the safety threshold, the main controller will control vehicle 1 to adjust its posture based on the calculated deviation information. This includes fine-tuning the lateral position and heading angle of vehicle 1 to achieve preliminary and accurate alignment between vehicle 1 and the target vehicle.

[0134] After the robot body 1 completes its pose adjustment, based on the target position indicated in the task as "left side," the main controller controls the lateral movement mechanism of the fork mechanism 3 to drive the forks to move laterally to the left, aligning them with the target carrier. Subsequently, the robot performs a series of storage and retrieval operations: first, forward alignment, where the robot body 1 slowly moves forward until the forks are fully inserted under the carrier; next, a lifting action, where the lifting mechanism 31 drives the forks to slightly lift, lifting the carrier from the shelf; finally, a backward retraction action, where the robot body 1 slowly reverses, safely removing the carrier from the shelf location and retrieving it back onto the robot body 1. This series of actions is completed efficiently within a narrow aisle, avoiding the inefficiency of traditional single-sided forklift robots that need to repeatedly drive out of the aisle to adjust their direction.

[0135] After picking up the goods, the robot carries the carrier and navigates to the designated unloading point. Upon arrival at the unloading point, the forks are readjusted to the preset recognition height, and the LiDAR scans the carrier on the forks and the target storage location to obtain 3D point cloud data.

[0136] The pose recognition module processes this point cloud data to identify the center of the target storage location. Specifically, based on prior information about the shelf position and width, the system detects point cloud data within the corresponding range and extracts shelf upright features. By accurately identifying the shelf uprights, the system can calculate the pose of the storage location center, and thus obtain the relative deviation between the center of the forklift and the center of the target storage location.

[0137] After calculating the deviation information, the system will also determine whether the deviation is within the safety adjustment threshold. Furthermore, since this is an unloading and stacking operation, the system will also perform a safety verification step: based on the point cloud data obtained from scanning the shelf location, the system checks the point cloud density within the target stacking area. If an unknown obstacle is detected within the target stacking area (e.g., a previous carrier being improperly placed causing a protrusion), the verification fails, the system reports the anomaly, and stops the operation to prevent collisions between the forks or carrier and the obstacle, ensuring the safety of high-level stacking operations. If the verification passes and the deviation is within the safety threshold, the main controller, based on the calculated deviation information, controls vehicle 1 to adjust its position, ensuring precise alignment between the carrier and the target location.

[0138] After the robot body 1 has been positioned, and the target location indicated in the task is "right side", the main controller controls the lateral movement mechanism to drive the forks to move laterally to the right, aligning them with the target storage location. Then, the robot performs storage and retrieval operations: forward alignment, the robot body 1 moves slowly forward, delivering the container into the storage location; next is the descent action, the lifting mechanism 31 drives the forks to descend, smoothly placing the container on the shelf; finally, the backward retraction action, the robot body 1 slowly reverses, safely retrieving the forks from the storage location.

[0139] Through the above process, the robot can accurately, efficiently, and safely access goods at high positions on both sides within extremely narrow aisles using high-precision 3D perception and multi-degree-of-freedom collaborative control. Compared with existing technologies, this solution significantly improves alignment accuracy and operational efficiency through real-time, high-precision 3D pose recognition and collaborative fine-tuning of the vehicle body 1 and forks based on deviation information, avoiding the reliance on human experience and repeated trial-and-error adjustments in traditional solutions. Simultaneously, the introduction of multiple safety verification mechanisms (such as safety reset, deviation threshold judgment, and stacked obstacle detection) effectively reduces the risk of collisions and goods falling during high-position operations, solving the comprehensive technical challenge of balancing efficiency, accuracy, and safety in narrow-aisle high-position operations.

[0140] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A control method for a double-sided fork narrow-channel stacking robot, characterized in that, Includes the following steps: S1: Receive a job task, the task containing information that the target position is left or right; S2: Control the vehicle to navigate to the preparatory work point, which is either a preparatory pickup point or a preparatory unloading point; S3: After arriving at the preparatory work point, the forks are adjusted to a preset recognition height according to the work type, and the target area is scanned by a multi-line lidar to obtain three-dimensional point cloud data; wherein, when picking up goods, the target vehicle and forks are scanned, and when unloading goods, the vehicle on the forks and the target storage location are scanned. S4: Process the three-dimensional point cloud data, identify the target pose, and calculate the deviation information for adjustment; S5: Based on the calculated deviation information, control the vehicle body to adjust its posture; S6: Based on the target position information in the task, control the forks to move laterally to the left or right, and perform storage and retrieval operations including forward alignment, lifting or lowering, and backward retraction.

2. The method according to claim 1, characterized in that, In step S4, when the operation is picking up goods, the three-dimensional point cloud data is processed to identify the end face of the vehicle. Specifically, this includes identifying the two support legs of the vehicle and calculating their center pose, thereby obtaining the lateral deviation (y) and heading angle deviation (θ) of the vehicle relative to the forks.

3. The method according to claim 1, characterized in that, In step S4, when the operation is unloading, the three-dimensional point cloud data is processed to identify the center of the target storage location. Specifically, this includes: based on prior information about the shelf position and width, detecting point cloud data within the corresponding range and extracting shelf column features, then calculating the pose of the storage location center, and finally obtaining the relative deviation between the center of the carrier on the forks and the center of the target storage location.

4. The method according to claim 1, characterized in that, In step S3, the control of adjusting the forks to a preset identification height according to the operation type is a closed-loop control based on the fork height information fed back by the pull-cord encoder.

5. The method according to claim 1, characterized in that, Before step S5 is executed, a judgment step is also included: judging whether the deviation information calculated in step S4 is within the preset safety adjustment threshold; if so, step S5 is executed. If not, report the error and suspend the current workflow.

6. The method according to claim 1, characterized in that, Before step S2, a safety initialization step is also included: determining whether the forks are in a preset safety reset position; if not, the forks are preferentially moved to the safety reset position.

7. The method according to claim 1, characterized in that, When the operation is unloading and stacking, a safety verification step is also included before step S5: based on the point cloud data obtained by scanning the shelf location, the point cloud density in the target stacking area is checked. If there are unknown obstacles, the verification fails to prevent collision.

8. A control system for a double-sided fork narrow-channel stacking robot, used to execute the control method as described in any one of claims 1-7, characterized in that, include: Vehicle body (1); The mast (2) is mounted on the vehicle body (1); The fork mechanism (3) is mounted on the mast (2) and includes a lifting mechanism (31) for driving the forks to rise and fall, and a lateral movement mechanism for driving the forks to move laterally to the left or right. The pose recognition module includes a multi-line lidar (41); The navigation and control module includes a main controller and a visual positioning unit for vehicle positioning and path navigation; The feedback module includes a drawstring encoder for detecting the absolute position of the fork lifting and lateral movement; The main controller is communicatively connected to the pose recognition module, navigation and control module, feedback module and forklift mechanism (3).

9. The control system according to claim 8, characterized in that, The multi-line lidar (41) is fixedly installed at the bottom of the fixed crossbeam (21) between the two forks on the mast (2). It is configured to rise and fall synchronously with the mast (2) and not move with the left and right lateral movement of the forks. The installation position of the multi-line lidar (41) satisfies the following: when the mast (2) is in a low position, its scanning field of view is not blocked by the forks, so as to identify the front face of the ground carrier and the support leg; when the mast (2) is in a high position, its scanning field of view can identify the support leg of the carrier on the forks and the uprights and crossbeams of the rack storage position.

10. The control system according to claim 8, characterized in that, The visual positioning unit includes a camera facing the ground for identifying navigation code tapes laid on the ground.

11. The control system according to claim 8, characterized in that, It also includes a safety detection module, which includes microswitches and photoelectric sensors set at the tips of the forks, as well as anti-collision contact edges and obstacle avoidance lidar set around the vehicle body (1).

12. The control system according to claim 8, characterized in that, The navigation and control module also includes: a navigation lidar (42) installed on the top of the vehicle body (1) as the main sensor for SLAM mapping and global positioning; a blind spot lidar (43) installed at the front of the vehicle body (1) for detecting low obstacles and obstacles in the blind spot of the navigation lidar (42); and a high-precision IMU (44) integrated inside the vehicle body (1) for measuring the vehicle body pitch, roll and angular velocity.

Images