Intelligent workshop-oriented agv material handling system, method, device and medium

Abstract

This invention discloses an AGV material handling system, method, equipment, and medium for smart workshops; including: a distributed collaborative sensing network composed of multiple AGVs, where each AGV broadcasts in real time the time window reservation set for the planned path and the uncertainty distribution of travel time, and performs probabilistic conflict prediction based on Gaussian process regression; a digital twin server constructs a three-dimensional virtual workshop; a decision server includes a joint optimization module that incorporates stress safety costs into the task-path joint optimization objective function and uses digital twin forward simulation for weight correction; a differentiable path planning module constructs a differentiable computation graph using the spatiotemporal A* algorithm and integrates it with multi-agent reinforcement learning; and a stress reconstruction and control module reconstructs the internal stress distribution of the goods and outputs the lifting height adjustment amount. This invention solves the problems of multi-AGV conflict, stress damage during handling of weakly rigid materials, and poor adaptability to dynamic environments, improving handling efficiency, safety, and system robustness.

Description

Technical Field

[0001] This invention belongs to the field of automated material handling technology, and specifically discloses AGV material handling systems, methods, equipment and media for smart workshops. Background Technology

[0002] Automated Guided Vehicles (AGVs), as core equipment for material handling in smart manufacturing workshops, have been widely used in automobile manufacturing, 3C electronics, photovoltaic production, precision machining and other fields.

[0003] Currently, typical technical approaches for AGV material handling systems include: adopting a centralized scheduling architecture, with a central server managing task allocation and path planning; avoiding multi-AGV path conflicts through time window models or conflict search algorithms (CBS); introducing A* or Dijkstra algorithms at the path planning level and optimizing turning costs to improve ride smoothness; and utilizing visual sensors and LiDAR to perceive the environment and generating obstacle avoidance strategies using deep learning models for obstacle avoidance control. Furthermore, for handling large or irregularly shaped materials, existing solutions employ dual or multiple AGVs working collaboratively, achieving coordinated operation through preset spacing and synchronous control.

[0004] However, the aforementioned existing technologies still have the following technical shortcomings in practical applications for smart factories:

[0005] First, the collaborative perception capabilities of multiple AGVs are insufficient. The lack of real-time information exchange between AGVs makes it impossible to predict the movement intentions of adjacent AGVs. They often only trigger avoidance after "accidental encounters" at path intersections, leading to frequent stops or even deadlocks.

[0006] Secondly, task scheduling and path planning are disconnected. Traditional solutions allocate tasks first and then plan paths, ignoring the coupling relationship between the two. This can easily lead to excessively long empty runs of AGVs or serious path conflicts, making it difficult to achieve global optimization.

[0007] Third, the ability to avoid obstacles and replan in real time in dynamic environments is poor. When faced with temporary obstacles, existing solutions mostly rely on local deceleration or stopping and waiting, lacking a global path replanning mechanism, making it difficult to efficiently return to the target after deviating from the predetermined route.

[0008] Fourth, the collaborative handling of weakly rigid materials lacks stress control. When handling large or fragile materials, multiple AGVs synchronize at fixed intervals, making it impossible to sense the internal stress of the material, which can easily lead to damage due to excessive local stress.

[0009] Fifth, dynamic changes in the workshop cannot be mapped to the scheduling system in real time. Most existing systems operate in an "open loop" manner and cannot dynamically adjust scheduling strategies based on real-time information such as workstation status and material inventory.

[0010] Therefore, it is necessary to invent AGV material handling systems, methods, equipment, and media for smart workshops to solve the above problems. Summary of the Invention

[0011] To overcome the aforementioned deficiencies in existing technologies, this invention provides an AGV material handling system, method, equipment, and medium for smart workshops. Through a distributed collaborative sensing network composed of multiple AGVs, each AGV broadcasts in real-time the time window reservation set for its planned path and the uncertainty distribution of its travel time, performing probabilistic conflict prediction based on Gaussian process regression. A digital twin server constructs a real-time synchronized 3D virtual workshop, with a built-in offline stress database storing material stress responses under different working conditions. The decision server includes a joint optimization module that incorporates stress safety costs into the task-path joint optimization objective function and uses digital twin forward simulation for closed-loop weight correction. The differentiable path planning module constructs a differentiable computation graph using the spatiotemporal A* algorithm, employs a Leaky ReLU soft cost function, introduces a gradient pruning mechanism, and integrates it end-to-end with multi-agent reinforcement learning. The stress reconstruction and control module reconstructs the internal stress distribution of goods in real-time based on a finite element reduced-order model and Kalman filtering, with a built-in material model library supporting adaptation to multiple material types. Distributed quadratic programming outputs the lifting height adjustment amount, effectively solving the problems mentioned in the background technology.

[0012] To achieve the above objectives, the present invention provides the following technical solution:

[0013] An AGV material handling system for smart workshops specifically includes:

[0014] Multiple automated guided vehicles (AGVs), each equipped with a wireless communication module, an environmental perception sensor, a positioning and navigation module, and a material handling actuator, wherein the wireless communication module supports direct data exchange between the AGVs;

[0015] The distributed collaborative sensing network is dynamically formed by each AGV broadcasting its own status information in real time through the wireless communication module. The status information includes: current pose, velocity vector, path point sequence of the planned path and reserved time window for each path point, and uncertainty distribution parameters of travel time. Each AGV maintains a distributed reservation table locally based on the status information received from neighboring AGVs, recording the occupancy of each grid cell in the workshop by neighboring AGVs at different time periods. The reservation mechanism includes a two-level deadlock resolution strategy: 200ms timeout exit and server arbitration.

[0016] The probability conflict prediction module is deployed on each AGV or edge server. It is used to calculate the time distribution overlap area of ​​the vehicle and the neighboring AGV occupying the same grid cell as the conflict probability based on the uncertainty distribution parameters of the travel time using Gaussian process regression. When the conflict probability exceeds a preset threshold, a speed coordination command or a path replanning request is triggered.

[0017] A digital twin server is used to build and update a three-dimensional virtual model that is synchronized with the physical workshop in real time. The virtual model contains the real-time status of each AGV, workstation, material, and passageway. The digital twin server has a built-in offline stress database that stores the pre-calculated stress response results of different materials under different driving speeds, turning angles, and road surface conditions.

[0018] A decision server, communicatively connected to the digital twin server, includes:

[0019] The joint optimization module receives material handling tasks, constructs an objective function that includes task allocation variables and path planning variables, and the objective function includes a stress safety cost term, which is calculated based on the estimated stress safety index of the cargo obtained from the pre-simulation of candidate handling paths by the digital twin server. The joint optimization module sends the generated candidate schemes to the digital twin server for high-fidelity forward simulation verification, and adjusts the weight coefficients of the objective function according to the deviation between the simulation results and the calculated value of the objective function.

[0020] The differentiable path planning module constructs a differentiable computation graph from the spatiotemporal A* path search algorithm. The grid occupancy cost parameter of the differentiable computation graph is used as a learnable parameter and is updated end-to-end through a multi-agent reinforcement learning algorithm. A gradient pruning mechanism is introduced to prevent gradient vanishing. The differentiable path planning module searches for a collision-free spatiotemporal path for each AGV from its current position through the loading point to the target workstation on a three-dimensional spatiotemporal grid map.

[0021] The stress reconstruction and control module utilizes a reduced-order finite element model and Kalman filtering, using data collected by force / torque sensors installed on each AGV handling actuator as boundary conditions to reconstruct the stress distribution inside the transported goods in real time. By solving a distributed quadratic programming problem, it outputs the lifting height adjustment amount for each AGV handling actuator, ensuring that the reaction force at each support point approaches the preset ideal distribution force and constraining the maximum stress of the goods to not exceed a preset allowable stress threshold. The stress reconstruction and control module has a built-in material model library management unit, supporting the import and management of reduced-order finite element models of materials of different sizes, materials, and shapes, automatically matching the corresponding model when a task is issued.

[0022] The vehicle-mounted execution controller, deployed on each AGV, receives the spatiotemporal path output by the differentiable path planning module and the lifting height adjustment output by the stress reconstruction and control module, and performs path tracking, obstacle avoidance, and collaborative transport actions. The vehicle-mounted execution controller adopts a priority scheduling mechanism, executing stress control, path tracking, dynamic obstacle avoidance, and collaborative communication tasks in descending order of real-time performance. Non-real-time reinforcement learning training tasks are uniformly migrated to the edge server for execution.

[0023] An AGV material handling method for smart workshops, applied to the above-mentioned system, includes the following steps:

[0024] S1. The digital twin server initializes the 3D virtual model of the workshop and establishes real-time data synchronization with each AGV and workstation equipment; it preloads the finite element reduced-order models of all commonly used materials into the material model library and generates the corresponding offline stress database.

[0025] S2. The decision server receives material handling tasks, calculates the comprehensive priority for each task, and calls the joint optimization module to generate task allocation schemes and preliminary path schemes. The digital twin server is used to perform high-fidelity forward simulation on candidate schemes and adjust the objective function weights according to the simulation deviation.

[0026] S3, the differentiable path planning module performs differentiable spatiotemporal A* path search for each AGV on a 3D spatiotemporal grid map, and outputs a spatiotemporal path containing a sequence of path points and the estimated arrival time of each path point; during the simulation, the AGV motion simulation and material stress simulation are decoupled, and the stress index is directly obtained from the offline stress database.

[0027] S4. Each AGV broadcasts its own status information in real time through a distributed collaborative sensing network, including the path point sequence of the planned path, the reserved time window, and the uncertainty distribution parameters of the travel time; each AGV updates its local distributed reservation table based on the received information; when the probability conflict prediction module calculates that the probability of conflict between the vehicle and the neighboring AGV exceeds a preset threshold, a speed coordination command is triggered.

[0028] S5. The on-board execution controller of each AGV travels according to the spatiotemporal path generated in step S3, and continuously performs path tracking, dynamic obstacle avoidance and replanning during the travel process; for the task of transporting weak rigid materials, the stress reconstruction and control module updates the reconstruction result of the internal stress distribution of the goods in real time in a closed loop, and outputs the lifting height adjustment amount. The on-board execution controller dynamically adjusts the transport execution mechanism of each AGV according to the lifting height adjustment amount, so that the maximum stress of the goods never exceeds the preset allowable stress threshold.

[0029] S6. After the task is completed, the AGV reports the execution data to the decision server. The digital twin server updates the virtual model synchronously and compares the actual running data with the simulation results to correct the model parameters. At the same time, the state, actions taken and rewards obtained based on this scheduling decision are stored as experience samples in the experience pool for online fine-tuning of the multi-agent reinforcement learning strategy.

[0030] An electronic device includes at least one processor, a memory communicatively connected to the at least one processor, and a computer program stored in the memory and executable by the at least one processor; when the processor executes the computer program, it implements the above-described AGV material handling method for smart workshops.

[0031] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described AGV material handling method for smart workshops.

[0032] The technical effects and advantages of this invention are as follows:

[0033] 1. This invention constructs a distributed collaborative perception network that broadcasts real-time status information (including reserved time windows and travel time uncertainties) between AGVs, and combines Gaussian process regression to predict the probability of conflict, which can proactively coordinate speed and path before a conflict occurs;

[0034] 2. This invention abandons the traditional serial mode of "allocation first, planning later" and constructs a unified objective function that includes task waiting time, transportation time, conflict cost, energy consumption and stress safety cost of weakly rigid materials. Closed-loop verification and weight correction are performed through high-fidelity forward simulation of a digital twin server. This scheme effectively avoids long-distance empty runs and path congestion of AGVs.

[0035] 3. This invention utilizes a finite element reduced-order model and Kalman filtering to reconstruct the internal stress distribution of goods in real time using six-dimensional force / torque data at each AGV support point. It also dynamically adjusts the lifting height of each AGV through distributed quadratic programming to achieve a balanced distribution of support reaction forces. When the local stress exceeds the threshold, stress-driven local path replanning is automatically triggered. This scheme controls the maximum stress deviation during the handling of weakly rigid materials, reduces the material damage rate, and broadens the application scope of AGV systems.

[0036] 4. This invention constructs a digital twin system that enables real-time bidirectional synchronization between the physical workshop and the virtual model. The system continuously evaluates the scheduling strategy through a simulation engine and compares the actual operating data with the simulation results to correct the model parameters. At the same time, it uses experience samples to fine-tune the multi-agent reinforcement learning strategy online. This closed-loop mechanism enables the system to have the ability to learn and evolve on its own, and can adaptively respond to dynamic changes in the workshop (such as the insertion of emergency tasks, equipment failures, path congestion, etc.). The scheduling performance is continuously optimized under long-term operation. Attached Figure Description

[0037] The present invention will be further described with reference to the accompanying drawings, but the embodiments in the drawings do not constitute any limitation on the present invention. For those skilled in the art, other drawings can be obtained based on the following drawings without creative effort.

[0038] Figure 1 This is a schematic diagram of the system module composition and signal interaction of the present invention.

[0039] Figure 2 This is the logic diagram of the distributed collaborative perception and probabilistic conflict prediction of the present invention.

[0040] Figure 3 This is a flowchart of the joint optimization and weight correction process of the present invention.

[0041] Figure 4 This is an overall flowchart of the AGV material handling method of the present invention. Detailed Implementation

[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0043] This invention provides, for example Figure 1-3 The AGV material handling system shown for smart workshops specifically includes:

[0044] Multiple automated guided vehicles (AGVs), each equipped with a wireless communication module, an environmental perception sensor, a positioning and navigation module, and a material handling actuator, wherein the wireless communication module supports direct data exchange between the AGVs;

[0045] The distributed collaborative sensing network is dynamically formed by each AGV broadcasting its own status information in real time through the wireless communication module. The status information includes: current pose, velocity vector, path point sequence of the planned path and reserved time window for each path point, and uncertainty distribution parameters of travel time. Each AGV maintains a distributed reservation table locally based on the status information of neighboring AGVs, recording the occupancy of each grid unit in the workshop by neighboring AGVs at different time periods.

[0046] The probability conflict prediction module is deployed on each AGV or edge server. It is used to calculate the time distribution overlap area of ​​the vehicle and the neighboring AGV occupying the same grid cell as the conflict probability based on the uncertainty distribution parameters of the travel time using Gaussian process regression. When the conflict probability exceeds a preset threshold, a speed coordination command or a path replanning request is triggered.

[0047] A digital twin server is used to build and update a three-dimensional virtual model that is synchronized with the physical workshop in real time. The virtual model contains the real-time status of each AGV, workstation, material, and passageway. The digital twin server has a built-in offline stress database that stores the pre-calculated stress response results of different materials under different driving speeds, turning angles, and road surface conditions.

[0048] A decision server, communicatively connected to the digital twin server, includes:

[0049] The joint optimization module receives material handling tasks, constructs an objective function that includes task allocation variables and path planning variables, and the objective function includes a stress safety cost term, which is calculated based on the estimated stress safety index of the cargo obtained from the pre-simulation of candidate handling paths by the digital twin server. The joint optimization module sends the generated candidate schemes to the digital twin server for high-fidelity forward simulation verification, and adjusts the weight coefficients of the objective function according to the deviation between the simulation results and the calculated value of the objective function.

[0050] The differentiable path planning module constructs a differentiable computation graph from the spatiotemporal A* path search algorithm. The grid occupancy cost parameter of the differentiable computation graph is used as a learnable parameter and is updated end-to-end through a multi-agent reinforcement learning algorithm. A gradient pruning mechanism is introduced to prevent gradient vanishing. The differentiable path planning module searches for a collision-free spatiotemporal path for each AGV from its current position through the loading point to the target workstation on a three-dimensional spatiotemporal grid map.

[0051] The stress reconstruction and control module utilizes a reduced-order finite element model and Kalman filtering, using data collected by force / torque sensors installed on each AGV handling actuator as boundary conditions to reconstruct the stress distribution inside the transported goods in real time. By solving a distributed quadratic programming problem, it outputs the lifting height adjustment amount for each AGV handling actuator, ensuring that the reaction force at each support point approaches the preset ideal distribution force and constraining the maximum stress of the goods to not exceed a preset allowable stress threshold. The stress reconstruction and control module has a built-in material model library management unit, supporting the import and management of reduced-order finite element models of materials of different sizes, materials, and shapes, automatically matching the corresponding model when a task is issued.

[0052] The vehicle-mounted execution controller, deployed on each AGV, receives the spatiotemporal path output by the differentiable path planning module and the lifting height adjustment output by the stress reconstruction and control module, and performs path tracking, obstacle avoidance, and collaborative transport actions. The vehicle-mounted execution controller adopts a priority scheduling mechanism, executing stress control, path tracking, dynamic obstacle avoidance, and collaborative communication tasks in descending order of real-time performance. Non-real-time reinforcement learning training tasks are uniformly migrated to the edge server for execution.

[0053] Furthermore, in the above technical solution, the status information broadcast by each AGV in the distributed collaborative sensing network also includes reservation requests and reservation confirmations. When any AGV needs to update its planned path, it first broadcasts a reservation request to neighboring AGVs via a wireless communication module. The reservation request includes the identifier of the grid cell to be occupied and a time window. The AGV receiving the request determines whether there is a conflict based on its local distributed reservation table. If there is no conflict, it returns a reservation confirmation. The requesting AGV can only occupy the time window after receiving confirmations from all relevant neighboring AGVs. If the requesting AGV waits for more than 200ms without receiving all confirmations, it automatically abandons the reservation request and replans the path. If the reservation fails three times in a row, it sends an arbitration request to the decision server, which then allocates the time window uniformly.

[0054] Furthermore, in the above technical solution, in the probability conflict prediction module, the uncertainty distribution parameter of the travel time is modeled using Gaussian process regression, the input is the historical operation data of the AGV, and the output is the mean and variance of the actual time to pass each path point; the conflict probability is defined as the overlapping integral of the probability density function of the arrival time distribution of the two AGVs on the time axis.

[0055] Furthermore, in the above technical solution, the objective function of the joint optimization module is expressed as:

[0056] J=min[w1·ΣT wait +w2·ΣT travel +w3·ΣC conflict +w4·ΣEenergy +w5·ΣC stress ]

[0057] Where C stress For the stress safety cost term, C stress =ζ·max(0,S threshold -S safe_pred ), S safe_pred The minimum stress safety index for cargo estimation obtained by simulating candidate paths through a digital twin server is ζ, which is a weighting coefficient. The time speedup ratio of the high-fidelity forward simulation is not less than 10:1. When the deviation between the simulation result and the calculated value of the objective function exceeds 5%, the weighting coefficients w1 to w5 and ζ are automatically adjusted using a Bayesian optimization method. Specifically, with the goal of minimizing the mean square error between the simulation output and the calculated value of the objective function, a finite number of iterations are performed in the weighting coefficient space until the deviation is less than 5% or the maximum number of iterations of 5 is reached. Then, the joint optimization objective function is resolved using the adjusted weighting coefficients.

[0058] Furthermore, in the above technical solution, in the differentiable path planning module, the grid cell size of the three-dimensional spatiotemporal grid map is 0.4 to 0.6 times the width of the AGV body (preferably 0.5, as the grid cell size directly affects the environmental resolution and computational complexity of path planning. Setting the grid size to 1 / 2 of the AGV body width ensures that, under any AGV pose, its body projection covers at least 2×2 grid cells, thus preserving sufficient collision detection redundancy in the discretized map). The time discretization step size is 0.05 to 0.2 seconds (preferably 0.1, as the time discretization step size determines the temporal resolution of AGV motion prediction). The 0.1-second interval matches the typical control cycle (10Hz~20Hz) of the AGV onboard controller and the sampling interval of the environmental perception sensor (LiDAR / IMU is usually 10~100Hz), enabling it to capture the acceleration and deceleration transients of the AGV and the time window overlap with other AGVs without significantly amplifying the prediction error. The evaluation function of the spatiotemporal A* algorithm is f(n,t)=g(n,t)+λ·h(n,t), where g(n,t) includes the movement distance cost, time cost, waiting cost, and turning cost, and λ is a dynamic weighting factor. The grid occupancy cost parameter is used as input to the differentiable computation graph, transformed into a soft occupancy cost through the Leaky ReLU function, and updated through the policy gradient of multi-agent reinforcement learning, enabling the path planning algorithm to adaptively learn avoidance strategies.

[0059] Furthermore, in the above technical solution, in the stress reconstruction and control module, the finite element reduced-order model is constructed offline through principal component analysis, compressing the high-dimensional finite element model of the cargo to no more than 50 modal coordinates (the full-order finite element model (typically with more than 100,000 nodes) cannot be solved in real time on the vehicle-mounted embedded platform. By retaining the first k dominant modes through principal component analysis, a balance can be achieved between accuracy and speed. The basis for selecting k≤50 is as follows: for large-size thin plate-like weakly rigid materials (length-to-width ratio 1:1~3:1), the first 20 modes can describe more than 95% of the deformation energy; the first 50 modes can cover more than 99% of the dynamic stress response, and the stress reconstruction error compared with the full-order model is usually less than 5%). In the online stage, the frequency is not less than 100Hz (the dynamic stress response of weakly rigid materials during the handling process is mainly caused by the unevenness of the AGV driving surface and the acceleration and deceleration impact excitation, and its energy is usually concentrated in the low frequency band below 50Hz. According to the Nyquist sampling theorem, a sampling frequency of 100Hz is sufficient to reconstruct the frequency within 50Hz without aliasing). Stress variation. Meanwhile, the mechanical response bandwidth of the AGV suspension system and lifting actuator generally does not exceed 20Hz, and the 100Hz control frequency is far higher than the effective bandwidth of the actuator, ensuring the stability of the stress closed loop. Six-dimensional force / torque data of each AGV support point are collected, and low-dimensional modal coordinates are estimated using Kalman filtering. Then, the stress values ​​of at least eight key points inside the cargo are reconstructed. The cargo stress safety index Ssafe = 1 - σmax / σallowable is defined, where σmax is the maximum reconstructed stress of the cargo, and σallowable is the allowable stress of the cargo. The system sets three levels of stress control thresholds: scheduling layer penalty threshold Ssafe = 0.7, online early warning replanning threshold Ssafe = 0.3, and execution layer hard constraint threshold Ssafe = 0.2. The distributed quadratic programming specifically uses minimizing the sum of squares of the deviations between the reaction forces of each support point and the ideal distributed force as the objective function, with the constraint that the maximum stress of the cargo does not exceed 0.8 times the allowable stress threshold (corresponding to Ssafe ≥ 0.2).

[0060] Example 1: Overall System Deployment and Hardware Configuration

[0061] This embodiment uses a smart manufacturing workshop for automotive parts as an example to describe in detail the AGV material handling system for smart workshops provided by the present invention.

[0062] Workshop layout and AGV configuration

[0063] The workshop has a floor plan of 50m (length) × 30m (width), and is arranged with 15 processing stations (including 5 loading stations, 8 processing stations, and 2 unloading stations), a main aisle (2.5m wide), and several branch aisles. The workshop floor is covered with an array of QR code navigation markers, with a spacing of 1m × 1m between the markers. Each QR code stores unique coordinate information (X, Y) and direction angle.

[0064] The workshop is equipped with 8 AGVs. Each AGV uses a differential drive chassis with body dimensions of 1.2m (length) × 0.8m (width) × 0.3m (height). The maximum load is 500kg, the maximum speed is 1.5m / s, the maximum acceleration is 0.5m / s², and the minimum turning radius is 0.6m.

[0065] The hardware configuration of each AGV is as follows:

[0066] Wireless communication module: Utilizing a 5G industrial module, supporting the 5G NR Sub-6GHz band, with an uplink peak rate of 100Mbps, a downlink peak rate of 300Mbps, and an end-to-end communication latency of ≤10ms. This module supports direct data exchange between AGVs via UDP broadcast protocol, eliminating the need for forwarding through a central server.

[0067] Environmental perception sensors include a 16-line lidar (detection range 30m, horizontal angular resolution 0.2°, vertical angular resolution 2.0°), an RGB-D depth camera (resolution 1280×720, frame rate 30fps, horizontal field of view 69°), and four ultrasonic sensors (detection distance 0.1m~3m, used for close-range safety obstacle avoidance).

[0068] Positioning and navigation module: integrates LiDAR, a six-axis IMU (gyroscope zero-bias stability 5° / h, accelerometer range ±8g), and a QR code vision sensor. The IMU sampling frequency is 200Hz, and the QR code vision sensor recognition frequency is 50Hz.

[0069] Material handling actuator: Adopts a lifting structure with a lifting plate size of 0.8m × 0.6m, a maximum lifting height of 50mm, and a lifting speed of 10mm / s. For dual AGV collaborative handling tasks, each AGV's lifting plate is equipped with one laser range sensor (range 0.2m~5m, accuracy ±1mm) and a pair of through-beam photoelectric sensors (detection distance 0.1m~10m, response time ≤1ms) at each of the four corners. For handling tasks involving weakly rigid materials, a six-dimensional force / torque sensor (range Fx / Fy / Fz=0-500N, Mx / My / Mz=0-50Nm, accuracy ±0.5%FS, sampling frequency 100Hz) is integrated at each of the four corner support points of the lifting plate.

[0070] Edge servers and digital twin platforms

[0071] An edge server is deployed in the workshop, configured with: two Intel Xeon Gold 6248R processors (3.0GHz, 40 cores in total), 256GB DDR4 memory, an NVIDIA A10 GPU (24GB VRAM), and a 2TB NVMe SSD. The server's operating system is Ubuntu 20.04, running a digital twin platform and decision server software.

[0072] The digital twin platform is built on the Unity 3D engine. It imports the workshop's CAD layout (DXF format) and equipment 3D models (FBX format) to generate a 1:1 high-fidelity virtual model containing all static elements such as workstations, aisles, columns, and charging piles, with a total facet count of approximately 2 million. It interfaces with the workshop's Manufacturing Execution System (MES) via the OPC UA protocol to obtain real-time processing status (including status codes such as idle, processing, waiting for materials, and fault) for 15 workstations and inventory data for each material buffer zone. It receives real-time pose, speed, battery level, and task status reported by each AGV via the MQTT protocol, synchronously updating the 3D pose display and status labels of the AGVs in the virtual model at 100ms intervals. The digital twin platform has a built-in offline stress database and material model library, supporting batch import of finite element reduced-order models of different materials and automatically generating stress response data for corresponding working conditions.

[0073] Vehicle controller

[0074] Each AGV's onboard controller utilizes the NVIDIA Jetson AGX Orin embedded AI computing platform, boasting a computing power of 275 TOPS and equipped with 32GB of LPDDR5 memory. The controller runs the Ubuntu 20.04 operating system and deploys software modules for the perception and execution layers. The perception layer includes units for multimodal data acquisition, heterogeneous data fusion, dynamic obstacle recognition, and collaborative perception among AGVs; the execution layer includes a path tracking controller, a multi-AGV collaborative handling controller, a stress-adaptive collaborative control module, a safety obstacle avoidance module, and a status reporting module.

[0075] The controller employs a real-time priority scheduling mechanism, with computing power allocated as follows: 30% to the stress reconstruction and control module (100Hz), 20% to the path tracking controller (10Hz), 25% to dynamic obstacle recognition (30fps), 15% to multi-AGV collaborative communication (50Hz), and the remaining 10% as redundancy. All non-real-time tasks, including training of multi-agent reinforcement learning strategies and calibration calculations of digital twin model parameters, are uniformly migrated to the edge server for execution, with the vehicle-mounted terminal only receiving policy parameter updates after training is complete.

[0076] Example 2: Specific Implementation of the Perception Layer

[0077] Multimodal data acquisition and heterogeneous data fusion

[0078] After the AGV starts, the onboard controller synchronously acquires LiDAR point cloud data, IMU attitude data, QR code visual positioning data, and wheel encoder data at a frequency of 100Hz. The heterogeneous data fusion unit uses the extended Kalman filter algorithm for data fusion, and the algorithm steps are as follows:

[0079] First, the kinematic model (differential drive model) of the AGV is used as the state transition equation, with the state vector being [x,y,θ,v,ω]ᵀ, where (x,y) are two-dimensional coordinates, θ is the heading angle, v is the linear velocity, and ω is the angular velocity. The control input is the speed of the left and right wheels.

[0080] Secondly, IMU data (acceleration and angular velocity) serves as input for the prediction phase, recursively updating the state at a frequency of 200Hz. When the lidar scans the QR code, the QR code vision sensor outputs the absolute position coordinates (xqr, yqr) and heading angle θqr, which are used as observation values ​​for Kalman updates. The fused AGV achieves a real-time pose positioning accuracy of ±3mm and a heading angle accuracy of ±0.5°, meeting the high-precision navigation requirements within the workshop.

[0081] Dynamic obstacle recognition

[0082] The dynamic obstacle recognition unit is equipped with a YOLOv8s object detection model optimized with TensorRT. The model training dataset contains 20,000 labeled images of common obstacles in the workshop, covering eight categories: personnel, other AGVs, forklifts, manual hydraulic trucks, material stacks, racks, open equipment doors, and safety cones. The model input is RGB images captured by an RGB-D depth camera, and the output is the bounding box (coordinates of the top left and bottom right corners) of each detected obstacle, the category confidence score (threshold set to 0.5), and the relative distance (obtained from the median depth value of the corresponding region in the depth image).

[0083] AGV Collaborative Sensing and Distributed Reserved Table

[0084] Each AGV's inter-AGV collaborative sensing unit broadcasts UDP status messages to the multicast address 239.0.0.1:8888 every 50ms. The messages are in Protobuf serialization format, approximately 500 bytes in size, and contain the following fields:

[0085] The unique identifier (ID) of the AGV;

[0086] Current timestamp (in milliseconds);

[0087] Real-time pose (x, y, θ);

[0088] Linear velocity v and angular velocity ω;

[0089] Current task status (0-Idle, 1-Towards loading point, 2-In transit, 3-Unloading, 4-Charging, 5-Fault).

[0090] The planned path is a sequence of path points (each path point contains coordinates x, y and estimated arrival time t, and the sequence length does not exceed 30; long paths with more than 30 path points use a segmented broadcast mechanism, with each segment containing 30 path points, and the order between segments is identified by sequence numbers).

[0091] Uncertainty distribution parameters of travel time: mean and variance of Gaussian distribution for each path point (output in real time through Gaussian process regression).

[0092] Reserve a request or reserve confirmation information (optional).

[0093] Each AGV maintains a distributed reservation table in memory. The data structure is a hash table, with the key being the grid cell ID (encoded as a 32-bit integer using X and Y coordinates), and the values ​​being a list of time windows occupied by that grid cell. Each time window record is [AGV ID, start time tstart, end time tend]. When an AGV receives status messages from neighboring AGVs, it parses the path point sequence and reserved time windows, and updates its local distributed reservation table. The distributed reservation table uses an LRU (Least Recently Used) strategy to clean up time window records that haven't been updated for more than 10 seconds, preventing unlimited memory growth.

[0094] When an AGV needs to update its planned path (e.g., due to a task change or obstacle avoidance), it first generates candidate paths and then extracts all grid cells to be occupied and their corresponding time windows from these candidate paths. Next, the AGV broadcasts a reservation request message to all potentially affected neighboring AGVs (typically other AGVs within a 30m radius) via its wireless communication module. Upon receiving a reservation request, the AGV checks its local distributed reservation table to see if the time window in the request overlaps with the occupied time windows of other AGVs in its local records. If there is no conflict, a reservation confirmation message is returned; if a conflict exists, a rejection message is returned along with information about the conflicting time windows. The requesting AGV can only officially adopt the path after receiving confirmation from all relevant neighboring AGVs; otherwise, the time window needs to be adjusted or the path replanned.

[0095] If the requesting AGV waits for more than 200ms without receiving all confirmations, it will automatically abandon the reservation request and replan the path. If the reservation fails three times in a row, it will send an arbitration request to the decision server, which will then allocate a time window and notify the relevant AGVs to update their local reservation tables.

[0096] Example 3: Specific Implementation of the Decision-Making Layer

[0097] Joint optimization module

[0098] The decision server receives material handling task flows from the MES system. Each task includes the following information: task ID, material type, loading point coordinates, target workstation coordinates, delivery deadline (absolute timestamp), whether it is urgent, material size and weight.

[0099] The task allocation module first calculates the comprehensive priority index P for each task. Taking a task involving the handling of weakly rigid photovoltaic glass as an example: the remaining time before the delivery deadline is 22 minutes, Pdeadline=0.5; the material type is "weakly rigid parts", Pmaterial=0.8; the target workstation is bottleneck workstation #05, which, according to the MES system, is currently idle but has a serious backlog of subsequent tasks, Pproduction=0.9; there is no urgent flag, Pemergency=0. With weighting factors α=0.35, β=0.20, γ=0.25, δ=0.20, the calculated P=0.35×0.5+0.20×0.8+0.25×0.9+0=0.175+0.16+0.225=0.56.

[0100] The joint optimization module reads the real-time status of the eight AGVs (obtained from the digital twin server), including each AGV's current position, speed, battery level, currently executing task, and remaining path. Then, the module constructs the joint optimization objective function:

[0101] J=min[w1·ΣT wait +w2·ΣT travel +w3·ΣC conflict +w4·ΣE energy +w5·ΣC stress ]

[0102] The definitions of each cost term are as follows:

[0103] T wait The time from when the task is assigned to when the AGV arrives at the loading point includes the AGV's idle waiting time and travel time.

[0104] T travel : The transportation time from when the AGV picks up the material to when it is delivered to the target workstation.

[0105] C conflict Path conflict cost is defined as the sum of the time overlap (in seconds) of the planned paths of two AGVs occupying the same grid cell at the same time.

[0106] E energyThe estimated energy consumption of the AGV to complete the path is calculated based on the travel distance, number of turns, and load.

[0107] C stress : Stress safety cost, Cstress=ζ·max(0,Sthreshold-Ssafe_pred). Where Sthreshold is the preset safety threshold (taken as 0.7), and Ssafe_pred is the estimated minimum stress safety index of the cargo obtained by simulating the candidate path through a digital twin server, ζ is taken as 0.5.

[0108] The initial weighting coefficients are set as follows: w1=0.30, w2=0.25, w3=0.20, w4=0.15, w5=0.10.

[0109] The solver used Gurobi 11.0, with a maximum solution time of 2 seconds. Due to the moderate number of tasks (10) and AGVs (8), the solver returned the optimal solution within 1.8 seconds. The optimal solution assigned the photovoltaic glass handling task to AGV#5 and AGV#6 for collaborative execution (dual AGV mode), while the remaining tasks were assigned to other AGVs. The solution also provided a preliminary path plan for each AGV (containing only key nodes such as the starting point, inflection points, loading points, and target work points).

[0110] The joint optimization module sends candidate solutions to the digital twin server for high-fidelity forward simulation verification. The digital twin server executes the simulation of the entire solution with a 10x time speedup (i.e., 10 seconds of simulation time requires only 1 second of actual time). During the simulation, the AGV kinematics simulation and material stress simulation are decoupled: the kinematics simulation is responsible for calculating the AGV's trajectory, speed changes, and collision detection; the stress simulation no longer solves the finite element model in real time, but directly retrieves the corresponding stress indicators from a pre-generated offline stress database based on parameters such as the AGV's speed, turning angle, and road slope. After the simulation, the actual values ​​of each performance indicator are output. If the deviation between the simulation results and the calculated values ​​of the objective function exceeds 5%, the weight coefficients w1 to w5 and ζ are automatically adjusted. Specifically, a Bayesian optimization method is used: a Gaussian process surrogate model is constructed with the weight vector w=[w1,w2,w3,w4,w5,ζ] as the optimization variable and the mean square error between the simulation output and the calculated values ​​of the objective function as the objective function. The system iterates five times, using the point with the maximum expected improvement as the next set of weights until convergence. The joint optimization objective function is then re-solved using the adjusted weights.

[0111] Differentiable path planning module

[0112] For each AGV, the differentiable path planning module performs path search on a 3D spatiotemporal grid map. The 3D spatiotemporal grid map is constructed as follows:

[0113] The workshop's planar area (X: 0-50m, Y: 0-30m) is divided into grid cells of size 0.4m × 0.4m (half the AGV body width of 0.8m). The time axis is discretized with a step size of 0.1s, and the prediction time domain is set to 120 seconds (i.e., 1200 time steps). The occupancy status of each spatiotemporal grid point (X,Y,T) is determined by the following information:

[0114] Static obstacles (workstation equipment, columns, walls) are marked as occupied on the entire timeline;

[0115] Dynamic obstacles (based on target detection results and Kalman filter predicted trajectories) are marked as occupied within the corresponding time window;

[0116] The planned paths and reserved time windows of other AGVs are recorded in the local distributed reservation table, and the corresponding grids and time windows are marked as occupied.

[0117] The evaluation function of the spatiotemporal A* algorithm is f(n,t) = g(n,t) + λ·h(n,t). Where:

[0118] g(n,t) represents the actual spatiotemporal cost from the starting point to node n at time t, including the cost of moving distance (1.0 per grid cell moved), the cost of time (0.1 per 0.1 seconds), the cost of waiting (0.5 per 0.1 seconds of waiting in place), and the cost of turning (0.5 per 45° change of direction).

[0119] h(n,t) is the heuristically estimated cost from node n to the target point at time t, using the three-dimensional Euclidean distance divided by the maximum speed of the AGV (1.5m / s) as the time estimate.

[0120] λ is a dynamic weighting factor with an initial value of 5.0. It gradually decreases as the search depth increases, and when the search depth exceeds 100 steps, λ decreases to 1.0.

[0121] During the search process, when it is necessary to expand the neighborhood nodes, the algorithm checks whether the spatiotemporal grid (X,Y,T) corresponding to the neighborhood node is occupied. If it is occupied, the node is skipped; if it is free, f(n,t) is calculated and added to the priority queue. The search terminates when the grid containing the target point (loading point or target workstation) is reached.

[0122] The differentiable path planning module constructs the entire search process as a differentiable computational graph. Specifically, the occupancy cost of each grid cell (i.e., the penalty value for whether the grid is marked as occupied) is used as a learnable parameter θgrid. In traditional A*, the occupancy cost is infinite (i.e., impassable) or zero (passable); in this module, the occupancy cost is replaced by a differentiable soft cost function: cost = LeakyReLU(θgrid, α = 0.1), with an output value between 0 and positive infinity. A cost close to 1 indicates a high occupancy probability. Then, the movement distance cost in f(n,t) during the search process is corrected to the original distance cost multiplied by (1 + cost). Thus, the total cost of the entire path is differentiable with respect to θgrid. Simultaneously, a gradient pruning mechanism is introduced to limit the norm of the policy gradient to within 5, preventing gradient explosion or vanishing.

[0123] This differentiable computation graph is connected to a multi-agent reinforcement learning policy network. The policy network outputs an adjustment Δθ to the θgrid, then recalculates the path cost, and feeds back the negative value of the cost as a reward signal to update the policy network through a policy gradient algorithm. This enables the path planning algorithm to adaptively learn avoidance strategies, such as automatically increasing the occupancy cost in areas where conflicts frequently occur, guiding the AGV to detour.

[0124] Stress Reconstruction and Control Module

[0125] For collaborative tasks involving two AGVs transporting weakly rigid materials (such as photovoltaic glass), the stress reconstruction and control module operates according to the following steps:

[0126] First, an offline finite element model (FEM) is established for the goods. All material FEM models are stored in a unified material model library, which supports adding, deleting, and modifying models and interfaces with the MES system. When a task is issued, the corresponding model is automatically loaded based on the material ID. Taking photovoltaic glass as an example, its dimensions are 2.0m × 1.0m × 0.0032m, the material is soda-lime glass, the elastic modulus is 70 GPa, Poisson's ratio is 0.22, the density is 2500 kg / m³, and the allowable stress is 35 MPa. A tetrahedral mesh is used, with approximately 100,000 nodes. The first 50 modes are extracted through modal analysis, and principal component analysis (PCA) is used to compress the nodal displacement field to 50 modal coordinates. The input to the reduced-order model is the force load at the support points (4 support points, each with 3 force components and 3 moment components, totaling 24 dimensions), and the output is the stress values ​​at 8 key locations inside the goods. The calculation speed of the reduced-order model is approximately 2000 times faster than that of the full-order finite element model. At the same time, an offline stress database is generated based on this reduced-order model, covering all common working conditions of AGV driving speed 0~1.5m / s, turning angle 0~90°, and road slope 0~5°.

[0127] During the online phase, the AGV's six-dimensional force / torque sensor acquires data at a frequency of 100Hz. After low-pass filtering (cutoff frequency 20Hz), the data is input into a Kalman filter. The Kalman filter's state vector consists of 50-dimensional modal coordinates, and the observation vector consists of 24-dimensional force / torque data. The state transition matrix is ​​determined by the modal frequencies, and the observation matrix is ​​determined by the modal force-stress mapping relationship of the reduced-order model. The filtered modal coordinates are used to reconstruct the stress values ​​at eight key points inside the cargo.

[0128] The problem of distributed quadratic programming is described as follows: The objective function is to minimize the actual reaction force Fi at each support point and the ideal distributed force Fi. target The sum of squared deviations between them, i.e., minΣ||Fi-Fi target The constraints are: ① The sum of the reactions at each support point equals the total weight of the cargo; ② The maximum stress of the reconstructed cargo, σmax, ≤ 0.8 × σallowable (σallowable is 35 MPa, meaning the maximum constraint stress does not exceed 28 MPa); ③ The adjustment amount Δhi of the lifting height of each AGV is within the range of [-5mm, +5mm]. This quadratic programming problem is solved using the OSQP solver on the onboard controller, with a solution time of less than 1ms. The solution results output the lifting height adjustment commands for each AGV lifting mechanism.

[0129] When the reconstructed stress shows that the maximum stress of the cargo on a certain road segment is consistently higher than 0.7 × σallowable (i.e., 24.5 MPa), stress-driven local path replanning is triggered. The AGV reports a stress warning to the decision server, which then calls the differentiable path planning module to search for a gentler path near the current location (reducing turning angles and gradients) or instructs the AGV to reduce its travel speed.

[0130] Example 4: Specific Implementation of the Execution Layer

[0131] Path tracking controller

[0132] The path tracking controller in the onboard execution controller employs a model predictive control (MPC) algorithm. The AGV's kinematic model uses a bicycle model, with state variables [x, y, θ, v]ᵀ, and control inputs of acceleration a and steering angle δ. The MPC prediction time domain N... p =20 (corresponding to 2 seconds), controlling the time domain N c =5 (corresponding to 0.5 seconds). In each control cycle (100ms), MPC solves the following optimization problem:

[0133]

[0134] Constraints: vmin≤vk≤vmax, amin≤ak≤amax, δmin≤δk≤δmax, and kinematic recursive equations.

[0135] The weight matrix is ​​set as follows: Q pos =10, Q θ =5, Q v =1, R u =0.1. This optimization problem is transformed into a quadratic programming problem, which is solved within 10ms using the OSQP solver. The acceleration and steering angle commands output by the controller are sent to the drive motor and steering motor of the AGV via the CAN bus.

[0136] Multi-AGV Collaborative Handling Controller

[0137] Taking the collaborative handling of photovoltaic glass by AGV#5 (master AGV) and AGV#6 (slave AGV) as an example. The master AGV receives the overall handling path (path point sequence length 78 points, time span 78 seconds) issued by the decision-making level. The master AGV decomposes the path into local paths for the slave AGVs: the path points of the slave AGV maintain a fixed relative offset from the path points of the master AGV, with an offset of (1.6m, 0), meaning the slave AGV is located 1.6m behind the master AGV. The slave AGV measures the actual distance (dactual) between itself and the master AGV in real time using a laser ranging sensor, with a target distance (dtarget) = 1.6m. The parameters of the PID controller are: Kp = 0.8, Ki = 0.05, Kd = 0.1. The controller outputs a speed correction. Where e = dtarget - actual. The corrected speed command vcmd = vmaster + Δv, where vmaster is the current speed of the master AGV. The measured accuracy of the dual-vehicle spacing control is ±15mm.

[0138] As a redundant safety mechanism, photoelectric sensors are installed on the opposite end faces of the two AGVs. When the distance between the two vehicles exceeds 1.8m or is less than 1.4m, the sensor outputs a trigger signal, which directly cuts off the power output of the AGV through the hardware circuit, achieving an emergency stop within milliseconds.

[0139] Stress Adaptive Cooperative Control Closed Loop

[0140] During handling, the stress reconstruction and control module operated continuously. Initially, the measured reaction forces at the four support points were: F1=125N, F2=118N, F3=132N, and F4=120N, totaling 495N (cargo weight approximately 50kg). The ideal distributed force was 123.75N. Actual stress reconstruction results showed that the maximum internal stress of the cargo was 28.5MPa, exceeding the allowable stress by 0.8 times (28MPa). The safety index Ssafe=0.186, below the threshold of 0.7.

[0141] The system triggered adaptive stress adjustment. Distributed quadratic programming yielded the following adjustment amounts: the left support point of AGV#5 increased by 1.2mm, the right support point of AGV#5 decreased by 0.8mm, the left support point of AGV#6 decreased by 0.5mm, and the right support point of AGV#6 increased by 0.9mm. The actuator adjusted the lifting height according to the instructions. After adjustment, the reaction forces at the four support points became: 121N, 125N, 124N, and 122N, and the sum of squared deviations decreased from approximately 56.2 (125-123.75)²+(118-123.75)²+...= to approximately 5.8. Stress reconstruction showed that the maximum stress decreased to 21.2MPa, and the safety index increased to 0.394. The system continued fine-tuning, and after three iterations (total time 4.2 seconds), the safety index reached 0.72, meeting the requirements. Throughout the adjustment process, the AGVs maintained a constant speed, and no visible deformation of the goods was observed.

[0142] Dynamic obstacle avoidance and replanning

[0143] When AGV#5 reaches t=24 seconds, the RGB-D camera detects a worker pushing a material cart across the passage 2.5m ahead. The dynamic obstacle recognition unit outputs the category as "person + cart", with a confidence level of 0.92, a relative distance of 2.5m, and a lateral speed of 0.8m / s. The predicted trajectory of the obstacle intersects with the AGV's path at t=26 seconds.

[0144] Since the obstacle was not another AGV, AGV#5 first attempted local path replanning. Local replanning started from the current pose, retaining the path segment from t=28 seconds onwards (i.e., the part after bypassing the obstacle), and only replanning the path segment between t=24 seconds and t=28 seconds. The local search space was limited to a rectangular area 0.8m to the left and right of the current position and 5m in front. The algorithm generated a detour path within 0.05 seconds: offset 0.4m to the left, increasing the detour distance by 2.8m, with an estimated increase in time of 2.3 seconds. AGV#5 traveled along the local detour path, bypassed the obstacle, and returned to the original planned path, without stopping throughout the entire process.

[0145] If local replanning fails (e.g., obstacles occupy a large area), or the AGV deviates from the predetermined path by more than 0.2m, global path replanning is triggered. Global replanning calls the differentiable path planning module of the decision server to search for a complete spatiotemporal path again, starting from the current pose and ending at the target workstation.

[0146] An AGV material handling method for smart workshops, applied to the aforementioned system, such as... Figure 4 As shown, it includes the following steps:

[0147] S1. The digital twin server initializes the three-dimensional virtual model of the workshop and establishes real-time data synchronization with each AGV and workstation equipment.

[0148] S2. The decision server receives material handling tasks, calculates the comprehensive priority for each task, and calls the joint optimization module to generate task allocation schemes and preliminary path schemes. The digital twin server is used to perform high-fidelity forward simulation on candidate schemes and adjust the objective function weights according to the simulation deviation.

[0149] S3. The differentiable path planning module performs a differentiable spatiotemporal A* path search for each AGV on a three-dimensional spatiotemporal grid map, and outputs a spatiotemporal path containing a sequence of path points and the estimated arrival time of each path point.

[0150] S4. Each AGV broadcasts its own status information in real time through a distributed collaborative sensing network, including the path point sequence of the planned path, the reserved time window, and the uncertainty distribution parameters of the travel time; each AGV updates its local distributed reservation table based on the received information; when the probability conflict prediction module calculates that the probability of conflict between the vehicle and the neighboring AGV exceeds a preset threshold, a speed coordination command is triggered.

[0151] S5. The on-board execution controller of each AGV travels according to the spatiotemporal path generated in step S3, and continuously performs path tracking, dynamic obstacle avoidance and replanning during the travel process; for the task of transporting weak rigid materials, the stress reconstruction and control module updates the reconstruction result of the internal stress distribution of the goods in real time in a closed loop, and outputs the lifting height adjustment amount. The on-board execution controller dynamically adjusts the transport execution mechanism of each AGV according to the lifting height adjustment amount, so that the maximum stress of the goods never exceeds the preset allowable stress threshold.

[0152] S6. After the task is completed, the AGV reports the execution data to the decision server. The digital twin server updates the virtual model synchronously and compares the actual running data with the simulation results to correct the model parameters. At the same time, the state, actions taken and rewards obtained based on this scheduling decision are stored as experience samples in the experience pool for online fine-tuning of the multi-agent reinforcement learning strategy.

[0153] Furthermore, in the above technical solution, the probability conflict prediction in S4 specifically includes: using Gaussian process regression to model the time distribution of the vehicle passing through each path point, calculating the overlap area of ​​the time distribution of the vehicle and neighboring AGVs occupying the same grid cell, and triggering a speed coordination command when the overlap area is greater than 0.3 times the width of the single-vehicle time window. The command content includes: specifying the passing order of the AGVs and the adjusted passing time window; the closed-loop control frequency of the stress reconstruction and control module in S5 is not less than 100Hz; when the reconstructed stress shows that the maximum stress of the cargo in a certain road segment is continuously higher than 0.7 times the allowable stress threshold (corresponding to Ssafe<0.3), stress-driven local path replanning is triggered to actively select a gentler path or reduce the driving speed.

[0154] Example 5: Complete Handling Task Execution Process

[0155] The following example of a complete material handling task illustrates the workflow of this system.

[0156] Task Issuance: The MES system issues task T001 to the decision server: Move a piece of photovoltaic glass (2.0m × 1.0m, 50kg) from loading point Pload (coordinates 35.6, 18.2) to target station Punload (coordinates 42.1, 22.5). The delivery deadline is 10 minutes, and the task is marked as "weakly rigid".

[0157] Step S1 (Twin Model Initialization): The digital twin server has pre-loaded the workshop model. When the task is issued, the twin model is in real-time synchronization. The current positions of each AGV are as follows: AGV#5 is located at (24.3, 12.7), with 92% battery and is idle; AGV#6 is located at (25.1, 13.2), with 90% battery and is idle; the remaining AGVs each have tasks. The system automatically loads the corresponding finite element reduced-order model from the material model library according to the material ID and retrieves the stress response data of the material from the offline stress database.

[0158] Step S2 (Task Allocation and Joint Optimization): The decision server calculates the task priority P=0.56. The joint optimization module assigns the task to AGV#5 (master) and AGV#6 (slave). The preliminary path scheme generated by the optimization module is: AGV#5 and AGV#6 travel from their current positions to the loading point, maintaining a distance of 1.6m, and then synchronously transport the materials to the target workstation. The total cost calculated by the joint optimization objective function is 156.3. The digital twin server simulates this scheme at a 10x speedup. During the simulation, the AGV motion and stress simulation are decoupled, and the stress index is directly obtained from the table. The actual simulation result is a total cost of 162.8, with a deviation of 4.2% (<5%), and the scheme is approved.

[0159] Step S3 (Differentiable Path Planning): The differentiable path planning module generates an accurate spatiotemporal path for AGV#5. The path contains 47 spatiotemporal nodes, each containing coordinates (X, Y) and an estimated arrival time T. For example, node 1: (24.3, 12.7, 0.0s); node 2: (24.7, 12.8, 0.4s); ...; node 47: (42.1, 22.5, 78.0s). The path makes 3 turns. The path of AGV#6 maintains a relative offset from that of AGV#5.

[0160] Step S4 (Distributed Collaborative Perception): AGV#5 and AGV#6 begin executing their tasks. Each AGV broadcasts a status message every 50ms. AGV#5 broadcasts the 47 nodes of its planned path (using a fragmented broadcast mechanism, sent in two fragments) and the reserved time window for each node (the time window width is 0.2s, i.e., the estimated arrival time ±0.1s). Other AGVs (such as AGV#1) update their local distributed reservation tables after receiving the message. When AGV#1 detects that its planned path conflicts with AGV#5's path at intersection (30.5, 15.2) at t=32s, AGV#1 reports a conflict warning to the decision server.

[0161] Step S5 (Adaptive Scheduling and Execution): After receiving the conflict warning, the reinforcement learning scheduling module of the decision server outputs a speed coordination instruction: AGV#1 decelerates to 0.5m / s, waits 2 seconds to pass, and AGV#5 passes at its original speed. AGV#1 and AGV#5 execute this instruction, and the conflict is resolved.

[0162] Step S6 (Stress Control and Obstacle Avoidance): During the handling process, the stress reconstruction and control module continues to operate. When traversing a slight slope, the stress reconstruction shows that the maximum stress has risen to 24.5 MPa, triggering stress-driven deceleration. The speed drops from 1.2 m / s to 0.8 m / s, and the stress falls back to 20.1 MPa. Subsequently, when encountering a worker pushing a cart, the local path replanning successfully avoids the obstacle.

[0163] Step S7 (Task Completion and Data Feedback): AGV#5 and AGV#6 arrive at the target workstation and unload simultaneously. Task completed, total time 93.2 seconds, energy consumption 0.32kWh. Execution data is reported to the decision server and digital twin server. The digital twin server compares the actual path with the planned path and updates the simulation model parameters (such as traffic interference frequency). Simultaneously, the state-action-reward samples from this scheduling process are stored in the experience pool for subsequent online fine-tuning of the policy network.

[0164] An electronic device includes at least one processor, a memory communicatively connected to the at least one processor, and a computer program stored in the memory and executable by the at least one processor; the processor may be a central processing unit (CPU), a graphics processing unit (GPU), a field-programmable gate array (FPGA), or an application-specific integrated circuit (ASIC). The memory may be a read-only memory (ROM), a random access memory (RAM), flash memory, or a solid-state drive (SSD). When the processor executes the computer program, it implements the aforementioned AGV material handling method for smart workshops. This electronic device can serve as an onboard controller for AGVs, or as a workshop edge server or a cloud server.

[0165] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described AGV material handling method for smart workshops.

[0166] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An AGV material handling system for smart workshops, characterized in that, Specifically, it includes: Multiple automated guided vehicles (AGVs), each equipped with a wireless communication module, an environmental perception sensor, a positioning and navigation module, and a material handling actuator, wherein the wireless communication module supports direct data exchange between the AGVs; The distributed collaborative sensing network is dynamically formed by each AGV broadcasting its own status information in real time through the wireless communication module. The status information includes: current pose, velocity vector, path point sequence of the planned path and reserved time window for each path point, and uncertainty distribution parameters of travel time. Each AGV maintains a distributed reservation table locally based on the status information of neighboring AGVs, recording the occupancy of each grid unit in the workshop by neighboring AGVs at different time periods. The probability conflict prediction module is deployed on each AGV or edge server. It is used to calculate the time distribution overlap area of ​​the vehicle and the neighboring AGV occupying the same grid cell as the conflict probability based on the uncertainty distribution parameters of the travel time using Gaussian process regression. When the conflict probability exceeds a preset threshold, a speed coordination command or a path replanning request is triggered. A digital twin server is used to build and update a three-dimensional virtual model that is synchronized with the physical workshop in real time. The virtual model contains the real-time status of each AGV, workstation, material, and passageway. A decision server, communicatively connected to the digital twin server, includes: The joint optimization module receives material handling tasks, constructs an objective function that includes task allocation variables and path planning variables, and the objective function includes a stress safety cost term, which is calculated based on the estimated stress safety index of the cargo obtained from the pre-simulation of candidate handling paths by the digital twin server. The joint optimization module sends the generated candidate schemes to the digital twin server for high-fidelity forward simulation verification, and adjusts the weight coefficients of the objective function according to the deviation between the simulation results and the calculated value of the objective function. The differentiable path planning module constructs a differentiable computation graph from the spatiotemporal A* path search algorithm. The grid occupancy cost parameter of the differentiable computation graph is used as a learnable parameter and is updated end-to-end through a multi-agent reinforcement learning algorithm. The differentiable path planning module searches for a collision-free spatiotemporal path for each AGV from its current position through the loading point to the target workstation on a three-dimensional spatiotemporal grid map. The stress reconstruction and control module uses a finite element reduced-order model and Kalman filtering, with the data collected by the force / torque sensors installed on each AGV handling actuator as boundary conditions, to reconstruct the stress distribution inside the transported goods in real time; and by solving a distributed quadratic programming problem, it outputs the lifting height adjustment amount of each AGV handling actuator, so that the reaction force at each support point approaches the preset ideal distribution force, and constrains the maximum stress of the goods not to exceed the preset allowable stress threshold. The vehicle-mounted execution controller, deployed in each AGV, is used to receive the spatiotemporal path output by the differentiable path planning module and the lifting height adjustment output by the stress reconstruction and control module, and to perform path tracking, obstacle avoidance and collaborative transport actions.

2. The AGV material handling system for smart workshops according to claim 1, characterized in that: In the distributed collaborative sensing network, the status information broadcast by each AGV also includes reservation requests and reservation confirmations. When any AGV needs to update its planned path, it first broadcasts a reservation request to neighboring AGVs through a wireless communication module. The reservation request includes the identifier of the grid cell to be occupied and a time window. The AGV that receives the request determines whether there is a conflict based on its local distributed reservation table. If there is no conflict, it returns a reservation confirmation. The requesting AGV can only occupy the time window after receiving confirmations from all relevant neighboring AGVs.

3. The AGV material handling system for smart workshops according to claim 1, characterized in that: In the probability conflict prediction module, the uncertainty distribution parameters of the travel time are modeled using Gaussian process regression. The input is the historical operation data of the AGV, and the output is the mean and variance of the actual time to pass each path point. The conflict probability is defined as the overlapping integral of the probability density function of the arrival time distribution of the two AGVs on the time axis.

4. The AGV material handling system for smart workshops according to claim 1, characterized in that: The objective function of the joint optimization module is expressed as: J = min [w1 ·∑T wait +w2·∑T travel +w3·∑C conflict +w4·∑E energy +w5·∑C stress ] Where C stress For the stress safety cost term, C stress =ζ·max(0,S threshold -S safe_pred ), S safe_pred To obtain the estimated minimum stress safety index for cargo through simulation of candidate paths using a digital twin server, S threshold ζ is the stress penalty threshold for the scheduling layer and the weight coefficient. The time speedup ratio of the high-fidelity forward simulation is not less than 10:

1. When the deviation between the simulation result and the calculated value of the objective function exceeds 5%, the weight coefficients w1 to w5 and ζ are automatically adjusted using the Bayesian optimization method. Specifically, with the goal of minimizing the mean square error between the simulation output and the calculated value of the objective function, a finite number of iterations are performed in the weight coefficient space until the deviation is less than 5% or the maximum number of iterations is reached (5). Then, the joint optimization objective function is re-solved using the adjusted weight coefficients.

5. The AGV material handling system for smart workshops according to claim 1, characterized in that: In the differentiable path planning module, the grid cell size of the 3D spatiotemporal grid map is 0.4 to 0.6 times the width of the AGV body, and the time discretization step size is 0.05 to 0.2 seconds. The evaluation function of the spatiotemporal A* algorithm is f(n,t)=g(n,t)+λ·h(n,t), where g(n,t) includes the movement distance cost, time cost, waiting cost, and turning cost, and λ is a dynamic weight factor. The grid occupancy cost parameter is used as the input to the differentiable computation graph and is updated through the policy gradient of multi-agent reinforcement learning, enabling the path planning algorithm to adaptively learn avoidance strategies.

6. The AGV material handling system for smart workshops according to claim 1, characterized in that: In the stress reconstruction and control module, the finite element reduced-order model is constructed offline through principal component analysis, compressing the high-dimensional finite element model of the cargo to no more than 50 modal coordinates. In the online stage, six-dimensional force / moment data of each AGV support point are collected at a frequency of no less than 100Hz. The low-dimensional modal coordinates are estimated by Kalman filtering, and then the stress values ​​of no less than 8 key points inside the cargo are reconstructed. The cargo stress safety index Ssafe=1-σmax / σallowable is defined, where σmax is the maximum reconstructed stress of the cargo and σallowable is the allowable stress of the cargo. The system sets three levels of stress control thresholds: scheduling layer penalty threshold Ssafe=0.7, online early warning replanning threshold Ssafe=0.3, and execution layer hard constraint threshold Ssafe=0.

2. The distributed quadratic programming is specifically: the objective function is to minimize the sum of squares of the deviations between the reaction forces of each support point and the ideal distributed forces, and the constraint condition is that the maximum stress of the cargo does not exceed 0.8 times the allowable stress threshold.

7. An AGV material handling method for smart workshops, applied to the system according to any one of claims 1 to 6, characterized in that, Includes the following steps: S1. The digital twin server initializes the three-dimensional virtual model of the workshop and establishes real-time data synchronization with each AGV and workstation equipment. S2. The decision server receives material handling tasks, calculates the comprehensive priority for each task, and calls the joint optimization module to generate task allocation schemes and preliminary path schemes. The digital twin server is used to perform high-fidelity forward simulation on candidate schemes and adjust the objective function weights according to the simulation deviation. S3. The differentiable path planning module performs a differentiable spatiotemporal A* path search for each AGV on a three-dimensional spatiotemporal grid map, and outputs a spatiotemporal path containing a sequence of path points and the estimated arrival time of each path point. S4. Each AGV broadcasts its own status information in real time through a distributed collaborative sensing network, including the path point sequence of the planned path, the reserved time window, and the uncertainty distribution parameters of the travel time; each AGV updates its local distributed reservation table based on the received information; when the probability conflict prediction module calculates that the probability of conflict between the vehicle and the neighboring AGV exceeds a preset threshold, a speed coordination command is triggered. S5. The on-board execution controller of each AGV travels according to the spatiotemporal path generated in step S3, and continuously performs path tracking, dynamic obstacle avoidance and replanning during the travel process; for the task of transporting weak rigid materials, the stress reconstruction and control module updates the reconstruction result of the internal stress distribution of the goods in real time in a closed loop, and outputs the lifting height adjustment amount. The on-board execution controller dynamically adjusts the transport execution mechanism of each AGV according to the lifting height adjustment amount, so that the maximum stress of the goods never exceeds the preset allowable stress threshold. S6. After the task is completed, the AGV reports the execution data to the decision server. The digital twin server updates the virtual model synchronously and compares the actual running data with the simulation results to correct the model parameters. At the same time, the state, actions taken and rewards obtained based on this scheduling decision are stored as experience samples in the experience pool for online fine-tuning of the multi-agent reinforcement learning strategy.

8. The AGV material handling method for smart workshops according to claim 7, characterized in that: The probability conflict prediction described in S4 specifically includes: using Gaussian process regression to model the time distribution of the vehicle passing each path point, calculating the overlap area of ​​the time distribution of the vehicle and neighboring AGVs occupying the same grid cell, and triggering a speed coordination command when the overlap area is greater than 0.3 times the width of the single-vehicle time window. The command content includes: specifying the passing order of the AGVs and the adjusted passing time window; the closed-loop control frequency of the stress reconstruction and control module described in S5 is not less than 100Hz; when the reconstructed stress shows that the maximum stress of the cargo in a certain road segment is continuously higher than 0.7 times the allowable stress threshold, stress-driven local path replanning is triggered to actively select a gentler path or reduce the driving speed.

9. An electronic device, characterized in that, It includes at least one processor, a memory communicatively connected to the at least one processor, and a computer program stored in the memory and executable by the at least one processor; when the processor executes the computer program, it implements the AGV material handling method for a smart workshop as described in any one of claims 7 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the AGV material handling method for smart workshops as described in any one of claims 7 to 8.

Images