AI-based automatic guided vehicle scheduling system and method

Abstract

The invention relates to an AI-based automatic guided vehicle scheduling system and method, and belongs to the technical field of vehicle scheduling. The system comprises a task management module, an environment perception module, a multi-agent reinforcement learning scheduling center, a path planning and conflict prediction module and a communication module. According to the invention, through multi-agent reinforcement learning, the system can learn a complex scheduling strategy from historical data, is adaptive to a dynamically changing environment, and realizes long-term overall benefit maximization instead of simple rule response; a dynamic weight mechanism is introduced, a plurality of factors such as task emergency degree, path efficiency, equipment state and traffic jam can be considered at the same time, and the weight is self-adjusted according to an actual effect, so that fine balance among multiple targets is realized; according to the method, conflict prediction is fused into a path planning stage, passive solution is changed into active avoiding, system deadlock and congestion risks are remarkably reduced, and operation fluency and reliability are improved.

Description

Technical Field

[0001] This invention belongs to the field of vehicle scheduling technology, specifically relating to an AI-based automated guided vehicle scheduling system and method. Background Technology

[0002] Automated Guided Vehicles (AGVs) play a crucial role in intelligent warehousing and flexible manufacturing systems. The core task of an AGV scheduling system is to allocate transportation tasks to multiple AGVs and plan conflict-free, high-efficiency logistics routes to maximize overall system efficiency (such as throughput and equipment utilization) and minimize operating costs (such as total travel distance and task waiting time).

[0003] Existing AGV scheduling methods mainly suffer from the following problems: 1. Static and Local Optimization: Traditional scheduling algorithms (such as first-come, first-served and shortest path first) are static or based on fixed rules. They cannot adapt to dynamically changing operating environments (such as new task insertion, AGV failure, and temporary path congestion). These algorithms typically only consider a single optimization objective (such as the shortest path) without taking into account multi-dimensional factors such as task urgency, AGV battery power, and traffic congestion. This leads to overall system inefficiency.

[0004] 2. Delayed Conflict Resolution: While common path planning methods (such as the A* algorithm and time window method) can plan initial paths, deadlocks or congestion are prone to occur when multiple AGVs are working together. A passive conflict resolution mechanism of "planning-detecting-resolving" is typically used, which has poor real-time performance and the solution may not be optimal, affecting the smoothness of the system.

[0005] 3. Poor coordination: Multiple AGVs are scheduled as independent entities, lacking an effective coordination mechanism. When the task of one AGV changes, it cannot quickly and intelligently influence the decisions of other AGVs, resulting in low overall system coordination efficiency.

[0006] Therefore, there is an urgent need for an AGV intelligent scheduling solution that can adapt to dynamic environments, comprehensively consider multi-dimensional factors, and achieve forward-looking collaborative scheduling. Summary of the Invention (a) Technical problems to be solved The technical problem to be solved by this invention is how to provide an AI-based automated guided vehicle (AGV) scheduling system and method to solve the problems of static and local optimization, delayed conflict resolution, and poor coordination in existing AGV scheduling methods.

[0007] (II) Technical Solution To address the aforementioned technical problems, this invention proposes an AI-based automated guided vehicle (AGV) scheduling system, which includes: a task management module, an environmental perception module, a multi-agent reinforcement learning scheduling center, a path planning and conflict prediction module, and a communication module. Task management module: Used to receive, parse and store transportation task requests from upper-layer business systems, and generate a queue of tasks to be scheduled; Environmental perception module: Through ground markings, sensor networks, and AGV onboard sensors deployed in the work area, it can acquire real-time information on the AGV's position, speed, battery status, task execution status, and static and dynamic obstacles on the map; Multi-agent reinforcement learning scheduling center: This scheduling center constructs an agent for each AGV, and all agents share a central decision network. This network takes global state information as input and outputs an action policy for each AGV. The system is trained through reinforcement learning so that the joint action of all AGVs can maximize a comprehensive reward function. Path planning and conflict prediction module: Receives decisions from the scheduling center and performs refined path planning for each AGV with assigned tasks; this module not only calculates the shortest path, but also combines the planned paths of all AGVs, uses a spatiotemporal graph for forward-looking conflict detection, and makes fine adjustments during the planning stage to avoid conflicts. Communication module: Responsible for stable, low-latency communication between the scheduling system and each AGV, as well as between AGVs, for transmitting instructions, status information, and lightweight coordination signals.

[0008] This invention also provides an AI-based automated guided vehicle (AGV) scheduling method, which includes the following steps: S1. System initialization: Load map information, initialize all AGV states, load reinforcement learning model and dynamic weight initial values; S2, State Awareness and Task Acquisition: The environment awareness module collects the global environment status in real time, and the task management module receives new tasks and adds them to the queue; S3. Comprehensive Priority Evaluation: For each task in the task queue and each AGV in available status, the dynamic weight decision module calculates its comprehensive priority score. S4. Multi-agent collaborative decision-making: The multi-agent reinforcement learning scheduling center takes the current global state as input and combines it with the comprehensive priority score calculated in step S3. It then assigns the optimal task to each AGV through the central decision-making network. The decision objective is to maximize the long-term comprehensive reward. S5. Proactive Path Planning and Conflict Resolution: The path planning module plans detailed paths for each AGV performing a task and simulates the future trajectories of all AGVs on a spatiotemporal graph to predict potential conflicts. If a conflict is predicted, the speed, path fine-tuning, or order adjustment of the relevant AGVs' paths will be performed to avoid conflicts at the source. S6. Instruction Issuance and Execution: The scheduling center issues the final task instructions and planned paths to the corresponding AGVs for execution. S7. Learning and Adaptation: The system continuously monitors the scheduling results and uses these feedback signals to update the parameters of the reinforcement learning model, optimize long-term decision-making capabilities, and adjust the weight coefficients in the dynamic weight decision-making module to make short-term task allocation more in line with actual operational needs.

[0009] (III) Beneficial Effects This invention proposes an AI-based automated guided vehicle (AGV) scheduling system and method. Compared with existing technologies, the technical solution proposed in this invention has the following advantages: 1. Adaptability and Intelligence: Through multi-agent reinforcement learning, the system can learn complex scheduling strategies from historical data, adapt to dynamically changing environments, and maximize long-term overall benefits, rather than simply responding to rules.

[0010] 2. Multi-objective balance optimization: A dynamic weighting mechanism is introduced, which can simultaneously consider multiple factors such as task urgency, path efficiency, equipment status, and traffic congestion, and adjust the weights according to the actual results to achieve a fine balance among multiple objectives.

[0011] 3. Proactive Conflict Avoidance: By integrating conflict prediction into the path planning stage, the passive resolution is transformed into proactive avoidance, which significantly reduces the risk of system deadlock and congestion and improves the smoothness and reliability of operation.

[0012] 4. High-efficiency collaboration: Through a central decision-making network that shares global information, efficient collaborative operation among multiple AGVs is achieved, avoiding the inefficiency caused by "each fighting its own battle". Attached Figure Description

[0013] Figure 1 This is a block diagram of the overall architecture of the AGV intelligent scheduling system of the present invention; Figure 2 This is a flowchart of the scheduling method of the present invention; Figure 3 A schematic diagram illustrating the calculation of the comprehensive priority score for the dynamic weight decision module. Detailed Implementation

[0014] To make the objectives, contents, and advantages of the present invention clearer, the specific embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples.

[0015] This invention belongs to the field of automated guided vehicle (AGV) scheduling technology, specifically involving an AI-based intelligent AGV scheduling method.

[0016] The purpose of this invention is to overcome the shortcomings of the prior art and provide an AGV intelligent scheduling system and method that can adapt to dynamic environments, comprehensively consider multi-dimensional factors, and achieve efficient collaboration and forward-looking conflict avoidance.

[0017] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides an AGV intelligent scheduling system, comprising: Task Management Module: Used to receive, parse, and store transportation task requests from upper-layer business systems (such as WMS and MES), and generate a queue of tasks to be scheduled.

[0018] Environmental perception module: Through ground markings, sensor networks, and AGV onboard sensors deployed in the work area, it can acquire real-time information such as the AGV's position, speed, battery status, task execution status, and static and dynamic obstacles on the map.

[0019] Multi-Agent Reinforcement Learning Scheduling Center: This module is the core of the entire scheduling system. It constructs an agent for each AGV, and all agents share a central decision-making network. This network takes global state information (states of all AGVs, states of all tasks, map topology, and traffic density distribution) as input and outputs an action strategy for each AGV (such as going to a task point, performing charging, or waiting in the buffer zone). Through reinforcement learning training, the system enables the joint actions of all AGVs to maximize a comprehensive reward function.

[0020] The overall reward function is designed as a weighted sum, including both global and local rewards. The formula is as follows: R total = a *R global + b *R local .

[0021] In the above formula, R total R is the reward value. global For global rewards, a R is the weighted value of the global reward. local For local rewards, b This is the weighted value of the local reward. a and b Used to balance global and local effects.

[0022] The global reward is calculated by the scheduling center and distributed equally to all managed AGVs in the system. It includes task completion rewards and system efficiency penalties. The task completion reward means that whenever any AGV completes a task, all AGVs receive a positive reward. The system efficiency penalty means that at each point in time, a small penalty is imposed on the total waiting time of all unfinished tasks in the system.

[0023] Local rewards are applied to individual AGVs, based on their own behavior and state, and are designed to guide them to make correct local decisions. These include task-related rewards, path and efficiency rewards, safety and collaboration rewards, and state management rewards.

[0024] Dynamic weighted decision module: Embedded within the scheduling center, it calculates the comprehensive priority score of each task to be assigned to each available AGV in real time. This score is calculated based on the dynamic weights of multiple factors, primarily including: Task urgency: The specified time limit for completing the task.

[0025] Path efficiency: Estimated travel time / distance from the AGV to the task pickup point.

[0026] AGV Status: Whether the AGV's current battery level is sufficient to complete the task.

[0027] System load balancing: Prevent some AGVs from being overloaded while others remain idle.

[0028] Regional traffic density: The current level of AGV congestion on the task path.

[0029] This module uses an online learning mechanism to dynamically adjust the weight coefficients of the above factors based on historical scheduling results (such as task completion rate and average waiting time), enabling the scheduling strategy to adapt to different work scenarios and peak periods.

[0030] The path planning and conflict prediction module receives decisions from the scheduling center and performs refined path planning for each AGV with assigned tasks. This module not only calculates the shortest path but also combines the planned paths of all AGVs to perform proactive conflict detection using a spatiotemporal graph (such as head-on conflicts, rear-end collisions, and node deadlocks). It also makes fine-tuning during the planning phase to avoid conflicts, rather than resolving them afterward.

[0031] Communication module: Responsible for stable, low-latency communication between the scheduling system and each AGV, as well as between AGVs, for transmitting instructions, status information, and lightweight coordination signals.

[0032] Secondly, the present invention provides an intelligent scheduling method for AGVs based on the above system, comprising the following steps: S1. System initialization: Load map information, initialize all AGV states, and load reinforcement learning models and initial dynamic weight values.

[0033] S2. Status Awareness and Task Acquisition: The environment awareness module collects the global environment status in real time, and the task management module receives new tasks and adds them to the queue.

[0034] S3. Comprehensive Priority Evaluation: For each task in the task queue and each AGV in available status, the dynamic weight decision module calculates its comprehensive priority score.

[0035] In this invention, a weighted multi-factor scoring model is used to calculate the overall priority score. The formula is as follows: Task priority score = α * T 紧急 +β*T 效率 +γ*T 系统 +δ*T 任务 -ξ*T 成本 .

[0036] α, β, γ, δ, and ξ are weighting coefficients. Their sum is 1, and their specific values ​​reflect the preferences of the scheduling strategy.

[0037] T 紧急 This is a task urgency factor, determined by task waiting time, task type priority, and deadline urgency.

[0038] T 效率 The handling efficiency factor is determined by the AGV task distance and path overlap.

[0039] T 系统 Optimize the overall system operation from a macro perspective. This is determined by factors such as station waiting penalties, AGV load balancing, and traffic congestion avoidance.

[0040] T 任务 Determined by the task length.

[0041] T 成本 As a negative indicator, it is determined by energy consumption cost and path cost.

[0042] S4. Multi-agent collaborative decision-making: The multi-agent reinforcement learning scheduling center uses the current global state as input and combines it with the comprehensive priority score calculated in step S3 to assign the optimal task to each AGV through the central decision-making network. The decision objective is to maximize the long-term comprehensive reward.

[0043] S5. Proactive Path Planning and Conflict Resolution: The path planning module plans detailed paths for each AGV performing a task and simulates the future trajectories of all AGVs on a spatiotemporal graph to predict potential conflicts. If a conflict is predicted, the speed, path fine-tuning, or sequence adjustment of the relevant AGVs is performed to avoid conflicts at their source.

[0044] S6. Instruction Issuance and Execution: The scheduling center issues the final task instructions and planned paths to the corresponding AGVs for execution.

[0045] S7. Learning and Adaptation: The system continuously monitors scheduling results (such as task completion time, AGV idle rate, and number of conflicts) and utilizes these feedback signals: Update the parameters of the reinforcement learning model to optimize its long-term decision-making ability.

[0046] Adjust the weight coefficients in the dynamic weight decision module to make the allocation of short-term tasks more in line with actual operational needs.

[0047] Example 1: In a large warehousing and sorting center, 50 AGVs are deployed. The system of this invention is deployed on a central server. The specific implementation steps are as follows: 1. The task management module receives order sorting tasks from WMS and generates instructions such as "from shelf 1 to packaging station 2".

[0048] 2. The environmental perception module uses the UWB positioning system and AGV on-board laser sensors to monitor the real-time position (e.g., "located in channel 1") and power (e.g., "65% remaining") of all AGVs, and detects that the channel in area 3 has narrowed due to temporary stacking of goods.

[0049] 3. The dynamic weighted decision module calculates the overall score of a new task T (high-priority order). During the current peak inbound and outbound period, the system automatically increases the weight of "task urgency" through online learning. Therefore, even if an idle AGV is closer to the task point but has low battery, the system may still assign task T to an AGV with sufficient battery power, which, although slightly farther away, has a smoother path, because the latter has a higher overall score.

[0050] The multi-agent reinforcement learning scheduling center decides whether AGV-25 will execute task T based on the global state (the positions of all AGVs, the task queue, and the map congestion status) and a comprehensive score. Its internal neural network has been trained to know that during congestion periods, choosing an AGV with a smooth path, although the single path may be slightly longer, can avoid exacerbating congestion and result in higher overall efficiency in the long run.

[0051] The path planning and conflict prediction module plans a path for AGV-25. Through spatiotemporal simulation, it predicts a conflict with AGV-10 at an intersection in 10 seconds. Therefore, the module instructs AGV-25 to slightly reduce its speed to yield to AGV-10, thus smoothly avoiding the conflict.

[0052] After the instruction is issued, the AGV-25 executes the task. The system records the effectiveness of this scheduling (whether task T is completed on time, whether there are any unexpected delays), and uses this information to fine-tune the reinforcement learning model and dynamic weights for continuous optimization.

[0053] Beneficial effects: Compared with the prior art, the technical solution proposed in this invention has the following beneficial effects: 1. Adaptability and Intelligence: Through multi-agent reinforcement learning, the system can learn complex scheduling strategies from historical data, adapt to dynamically changing environments, and maximize long-term overall benefits, rather than simply responding to rules.

[0054] 2. Multi-objective balance optimization: A dynamic weighting mechanism is introduced, which can simultaneously consider multiple factors such as task urgency, path efficiency, equipment status, and traffic congestion, and adjust the weights according to the actual results to achieve a fine balance among multiple objectives.

[0055] 3. Proactive Conflict Avoidance: By integrating conflict prediction into the path planning stage, the passive resolution is transformed into proactive avoidance, which significantly reduces the risk of system deadlock and congestion and improves the smoothness and reliability of operation.

[0056] 4. High-efficiency collaboration: Through a central decision-making network that shares global information, efficient collaborative operation among multiple AGVs is achieved, avoiding the inefficiency caused by "each fighting its own battle".

[0057] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. An AI-based automated guided vehicle (AGV) scheduling system, characterized in that, The system includes: a task management module, an environment perception module, a multi-agent reinforcement learning scheduling center, a path planning and conflict prediction module, and a communication module; Task management module: Used to receive, parse and store transportation task requests from upper-layer business systems, and generate a queue of tasks to be scheduled; Environmental perception module: Through ground markings, sensor networks, and AGV onboard sensors deployed in the work area, it can acquire real-time information on the AGV's position, speed, battery status, task execution status, and static and dynamic obstacles on the map; Multi-agent reinforcement learning scheduling center: This scheduling center constructs an agent for each AGV, and all agents share a central decision network. This network takes global state information as input and outputs an action policy for each AGV. The system is trained through reinforcement learning so that the joint action of all AGVs can maximize a comprehensive reward function. Path planning and conflict prediction module: Receives decisions from the scheduling center and performs refined path planning for each AGV with assigned tasks; this module not only calculates the shortest path, but also combines the planned paths of all AGVs, uses a spatiotemporal graph for forward-looking conflict detection, and makes fine adjustments during the planning stage to avoid conflicts. Communication module: Responsible for stable, low-latency communication between the scheduling system and each AGV, as well as between AGVs, for transmitting instructions, status information, and lightweight coordination signals.

2. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 1, characterized in that, The global status information includes: the status of all AGVs, the status of all tasks, the map topology, and the traffic density distribution.

3. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 1, characterized in that, The action strategy includes: going to a task point, performing charging, and waiting in the buffer zone.

4. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 1, characterized in that, The comprehensive reward function is designed as a weighted sum, including global and local rewards; the formula is as follows: R total = a *R global + b *R local In the above formula, R total R is the reward value. global For global rewards, a R is the weighted value of the global reward. local For local rewards, b This is a weighted value for local rewards. a and b Used to balance global and local effects.

5. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 4, characterized in that, The global reward is calculated by the scheduling center and distributed equally to all managed AGVs in the system. It includes task completion rewards and system efficiency penalties. The task completion reward means that whenever any AGV completes a task, all AGVs receive a positive reward. The system efficiency penalty means that at each point in time, a small penalty is imposed on the total waiting time of all unfinished tasks in the system.

6. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 4, characterized in that, Local rewards are applied to individual AGVs and are based on their own behavior and state. They are designed to guide individuals to make correct local decisions and include task-related rewards, path and efficiency rewards, safety and collaboration rewards, and state management rewards.

7. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 1, characterized in that, The multi-agent reinforcement learning scheduling center also includes a dynamic weight decision module, which calculates the comprehensive priority score of each task to be assigned for each available AGV in real time. This score is calculated based on the dynamic weights of multiple factors.

8. The AI-based automated guided vehicle (AGV) scheduling system as described in claim 7, characterized in that, The comprehensive priority score is calculated based on task urgency, path efficiency, AGV status, system load balancing, and regional traffic density. This module uses an online learning mechanism to dynamically adjust the weight coefficients of the above factors based on historical scheduling results, enabling the scheduling strategy to adapt to different work scenarios and peak periods.

9. An AI-based automated guided vehicle (AGV) scheduling method based on the system according to any one of claims 1-8, characterized in that, The method includes the following steps: S1. System initialization: Load map information, initialize all AGV states, load reinforcement learning model and dynamic weight initial values; S2, State Awareness and Task Acquisition: The environment awareness module collects the global environment status in real time, and the task management module receives new tasks and adds them to the queue; S3. Comprehensive Priority Evaluation: For each task in the task queue and each AGV in available status, the dynamic weight decision module calculates its comprehensive priority score. S4. Multi-agent collaborative decision-making: The multi-agent reinforcement learning scheduling center takes the current global state as input and combines it with the comprehensive priority score calculated in step S3. It then assigns the optimal task to each AGV through the central decision-making network. The decision objective is to maximize the long-term comprehensive reward. S5. Proactive Path Planning and Conflict Resolution: The path planning module plans detailed paths for each AGV performing a task and simulates the future trajectories of all AGVs on a spatiotemporal graph to predict potential conflicts. If a conflict is predicted, the speed, path fine-tuning, or order adjustment of the relevant AGVs' paths will be performed to avoid conflicts at the source. S6. Instruction Issuance and Execution: The scheduling center issues the final task instructions and planned paths to the corresponding AGVs for execution. S7. Learning and Adaptation: The system continuously monitors the scheduling results and uses these feedback signals to update the parameters of the reinforcement learning model, optimize long-term decision-making capabilities, and adjust the weight coefficients in the dynamic weight decision-making module to make short-term task allocation more in line with actual operational needs.

10. The AI-based automated guided vehicle scheduling method as described in claim 9, characterized in that, In S3, the comprehensive priority score is calculated using a weighted multi-factor scoring model; the formula is as follows: Task priority score = α * T 紧急 +β*T 效率 +γ*T 系统 +δ*T 任务 -ξ*T 成本 ; Where α, β, γ, δ, and ξ are weighting coefficients; their sum is 1, and their specific values ​​reflect the preferences of the scheduling strategy. T 紧急 This is a task urgency factor, determined by task waiting time, task type priority, and deadline urgency. T 效率 The handling efficiency factor is determined by the AGV task distance and path overlap. T 系统 Optimize the overall system operation from a macro perspective; determined by station waiting penalties, AGV load balancing, and traffic congestion avoidance. T 任务 Determined by the task length; T 成本 As a negative indicator, it is determined by energy consumption cost and path cost.

Images