Multi-category robot distribution center intelligent scheduling and operation and maintenance method, system and device

Abstract

The application provides a multi-category robot distribution place intelligent scheduling and operation and maintenance method, system and equipment, relates to the robot control technical field, and the method comprises the following steps: acquiring multi-category robot state data in an entity distribution place, defining a robot group and a unified infrastructure state vector according to the multi-category robot state data, and constructing a digital twin model. According to residual power, a health index, a task value, a task risk, a berth queue state and energy cost, a rolling horizon optimization model is established to solve and generate a cooperative scheduling strategy, the strategy is simulated and verified in the digital twin model, path conflicts, energy shortages, berth conflicts or safety risks are checked, the robot is driven to perform a preset operation according to the verified strategy, and the digital twin model and the optimization model are closed-loop corrected according to real-time feedback state data, and the execution result is written back to the digital twin model. The application improves the long-term operation efficiency and resource cooperation efficiency of the multi-category robot cluster.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and more specifically, to a method, system, and equipment for intelligent scheduling and operation and maintenance of multi-type robot hubs. Background Technology

[0002] Currently, various types of robots, including drones, quadruped robots, wheeled autonomous mobile robots, and humanoid robots, are widely used in scenarios such as power line inspection, logistics distribution, emergency rescue, and smart agriculture. To support the long-term autonomous operation of heterogeneous robots in complex environments, robot hubs, such as robot houses, charging stations, and mobile stations, are becoming a hot research topic in the industry, serving as core infrastructure for energy replenishment, task handover, data offloading, and health maintenance. In actual operations, various types of robots often need to perform long-duration tasks involving multiple sub-steps. For example, when conducting infrared inspections of power distribution cabinets in area A, a quadruped robot is dispatched to a narrow area for verification after an anomaly is detected, and then a mobile operation robot is called in for further handling. Therefore, a systematic and coordinated scheduling of multi-robot task allocation, return to port for recharging, reloading and maintenance, and redeployment within the hub is needed to enable heterogeneous robots to continuously complete complex cross-regional, multi-stage, and multi-interaction tasks under unattended conditions.

[0003] In related technologies, traditional scheduling systems typically employ priority queue-based task allocation methods or path planning-based obstacle avoidance algorithms. These systems manage robot return to port and recharging through fixed thresholds, such as mandatory return when battery levels drop below 30% or queuing rules based on First-Come, First-Served (FCFS). However, in long-term operation scenarios involving multi-type robot hubs, various operating conditions and resource constraints often change dynamically simultaneously, exhibiting complex coupling relationships. The aforementioned single-dimensional scheduling rules, focusing only on one aspect of task allocation or obstacle avoidance, struggle to comprehensively address the multiple interdependent demands arising during operation, and fail to effectively integrate the diverse characteristics of heterogeneous robots and their collaborative requirements during task execution. Therefore, in actual operation, problems such as uneven resource utilization, failure to respond promptly to critical needs, blockage of emergency situations by routine processes, and repeated use of faulty equipment easily arise, raising planning and execution feasibility issues. Summary of the Invention

[0004] The problem addressed by this invention is how to improve the long-term operational efficiency and resource coordination efficiency of multi-category robot clusters in a distribution environment.

[0005] To address the aforementioned problems, this invention provides a method, system, and equipment for intelligent scheduling and operation and maintenance of multi-category robot hubs.

[0006] In a first aspect, the intelligent scheduling and operation and maintenance method for multi-category robot distribution centers of the present invention includes: Acquire status data of multiple types of robots in physical distribution centers, including robot status, task status, berth status, and energy status. Based on the robot state, the task state, the berth state, and the energy state, define a unified state vector for the robot swarm and a unified state vector for the infrastructure. Based on the unified state vector of the robot swarm and the unified state vector of the infrastructure, a digital twin model of the physical distribution center is constructed. Based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue state extracted from the berth state, and the energy cost extracted from the energy state, a rolling time-domain optimization model is established, and a collaborative scheduling strategy is generated by solving the rolling time-domain optimization model. The cooperative scheduling strategy is simulated and verified in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, insufficient energy, berth conflicts or safety risks, to obtain a cooperative scheduling strategy that has passed the verification. The verified collaborative scheduling strategy drives multiple types of robots in the physical distribution center to perform preset operations. During the execution of the preset operations, the digital twin model and the rolling time-domain optimization model are corrected in a closed loop based on the status data fed back by the robots in real time, and the execution results are written back to the digital twin model.

[0007] Optionally, the step of acquiring status data of multiple types of robots in the physical distribution center includes: robot status, task status, berth status, and energy status, including: The robot status is obtained, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The task status is obtained, which includes task identifier, task type, task priority, task value, expected completion time, associated robot, task progress, and operational risk. The berth status is obtained, which includes berth identifier, berth type, berth queue status, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, and available tool set; The energy status is obtained, which includes energy type, current energy price, energy storage capacity, charging equipment status, and energy cost.

[0008] Optionally, defining a unified state vector for the robot swarm and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state includes: The parameter order of the preset robot swarm state vector is defined, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The parameter values ​​corresponding to each robot are obtained from the robot states, and the parameter values ​​corresponding to each robot are assigned according to the preset robot swarm state vector to obtain the state vector of that robot. The state vectors of all robots constitute the unified state vector of the robot swarm. The parameters of the preset infrastructure state vector are ordered, including berth type, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, available tool set, and security isolation level. The parameter values ​​corresponding to each berth are obtained from the berth status, and the parameter values ​​corresponding to each berth are assigned according to the preset infrastructure state vector to obtain the state vector of that berth. The state vectors of all berths constitute the unified state vector of the infrastructure.

[0009] Optionally, constructing a digital twin model of the entity distribution center based on the unified state vector of the robot swarm and the unified state vector of the infrastructure includes: Based on the remaining power in the unified state vector of the robot swarm, an energy consumption model for the physical distribution center is established. A health assessment model for the physical distribution center is established based on the health index in the unified state vector of the robot swarm. Based on the berth type, available charging power, and swappable battery specifications in the unified state vector of the infrastructure, a berth service time model for the physical distribution center is established. Based on the task semantic information in the task status, establish a task semantic graph model of the entity distribution center; The energy consumption model, the health assessment model, the berth service time model, and the task semantic graph model are fused together to obtain a digital twin model of the physical distribution center.

[0010] Optionally, the step of establishing a rolling time-domain optimization model based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue state extracted from the berth state, and the energy cost extracted from the energy state includes: Based on the stated task value, construct task benefit items; Based on the remaining electricity, the energy cost, and the predicted energy consumption of the return route, an energy cost item is constructed. Based on the aforementioned health index, construct maintenance risk items; Based on the described operational risks, construct safety risk items; Based on the berth queue status, construct a queuing cost item; The objective function is obtained by weighted summing of the task benefit item, the energy cost item, the maintenance risk item, the security risk item, and the queuing cost item; Based on the objective function and preset constraints, the rolling time-domain optimization model is established.

[0011] Optionally, the step of generating a cooperative scheduling strategy by solving the rolling time-domain optimization model includes: Starting from the current time, the prediction time domain length is set, and the rolling time domain optimization model is discretized within the prediction time domain length to obtain a discretized optimization model; Obtain the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment, and use the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment as the initial state; The initial state is input into the discretized optimization model, and a preset solver is called to solve the problem, thereby obtaining the optimal action sequence of each robot in the prediction time domain; The actions between the current moment and the next scheduling moment are extracted from the optimal action sequence to generate the collaborative scheduling strategy. The collaborative scheduling strategy includes at least one of the following for each robot: return to port instruction, queuing instruction, energy replenishment instruction, equipment change instruction, maintenance instruction, and redeployment instruction.

[0012] Optionally, the simulation verification of the cooperative scheduling strategy in the digital twin model includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, energy shortages, berth conflicts, or safety risks to obtain a verified cooperative scheduling strategy, including: The collaborative scheduling strategy is loaded into the digital twin model, and the simulation start time and simulation time step are set. According to the simulation time step, the return-to-port instruction, the queuing instruction, the recharging instruction, the equipment replacement instruction, the maintenance instruction, and the redeployment instruction in the cooperative scheduling strategy are simulated and executed frame by frame in the digital twin model; During the simulation, it is detected whether a preset abnormal event occurs. If the preset abnormal event is not detected within the simulation time step, it is determined that the cooperative scheduling strategy has been verified and the verified cooperative scheduling strategy is obtained. If the preset abnormal event is detected, the cooperative scheduling strategy verification is marked as unsuccessful, and the process returns to the step of establishing the rolling time-domain optimization model to regenerate the cooperative scheduling strategy.

[0013] Optionally, the step of driving multiple types of robots in the entity distribution center to perform preset operations according to the verified collaborative scheduling strategy, and, during the execution of the preset operations, performing closed-loop correction on the digital twin model and the rolling time-domain optimization model based on the status data fed back by the robots in real time, and writing the execution result back to the digital twin model, includes: The verified collaborative scheduling strategy is sent to the execution controller of the entity distribution center; The execution controller parses the cooperative scheduling strategy to obtain a sequence of control instructions for each robot; Drive the corresponding robot to perform operations according to the sequence of control instructions; During the operation, the robot's real-time status data is collected at a fixed frequency. The status data includes the actual position, actual remaining battery power, actual health index, command execution status, and fault codes. Based on the collected state data, update the state vector of the corresponding robot and the state vector of the corresponding berth in the digital twin model; The execution status and fault code of the instruction are written as the execution result into the execution record of the digital twin model; Calculate the deviation between the state data and the predicted state in the collaborative scheduling strategy, and adjust the energy consumption model parameters, berth service time model parameters, or health assessment model parameters in the rolling time-domain optimization model according to the deviation.

[0014] Secondly, the present invention provides an intelligent scheduling and operation and maintenance system for multi-category robot distribution centers, comprising: The data acquisition unit is used to acquire status data of multiple types of robots in the physical distribution center. The status data includes: robot status, task status, berth status and energy status. A state definition unit is used to define a unified state vector for the robot swarm and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state. The twin model construction unit is used to construct a digital twin model of the physical distribution center based on the unified state vector of the robot swarm and the unified state vector of the infrastructure. The optimization scheduling unit is used to establish a rolling time-domain optimization model based on the remaining power and health index extracted from the robot status, the task value and operational risk extracted from the task status, the berth queue status extracted from the berth status, and the energy cost extracted from the energy status, and to generate a collaborative scheduling strategy by solving the rolling time-domain optimization model. The simulation verification unit is used to perform simulation verification of the cooperative scheduling strategy in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking whether there are path conflicts, insufficient energy, berth conflicts or safety risks, and obtaining a cooperative scheduling strategy that has passed the verification. The execution and correction unit is used to drive multiple types of robots in the entity distribution center to perform preset operations according to the verified collaborative scheduling strategy, and to perform closed-loop correction of the digital twin model and the rolling time-domain optimization model based on the status data fed back by the robots in real time during the execution of the preset operations, and to write the execution results back to the digital twin model.

[0015] Thirdly, an electronic device according to the present invention includes: a processor and a memory, the memory being used to store a computer program; When the computer program is loaded by the processor, it causes the processor to execute the above-described intelligent scheduling and operation and maintenance method for multi-category robot distribution centers.

[0016] This invention relates to a method, system, and equipment for intelligent scheduling and operation of multi-category robot hubs. By constructing a unified state vector including robot status, task status, berth status, and energy status, and establishing a digital twin model of the physical hub based on this vector, it achieves a globally accurate mapping of the operational status of multi-category robot clusters. This invention can simultaneously perceive individual robot differences, such as health indices, dynamic task characteristics such as value and risk, facility resource occupancy such as berth queues, and multi-source heterogeneous information such as energy supply costs. Global perception ensures that collaborative scheduling has sufficient information basis, avoiding uneven resource utilization due to information isolation. Secondly, a rolling time-domain optimization model is used to generate collaborative scheduling strategies. The model's input includes multiple constraints such as remaining power, health index, task value, operational risk, berth queue status, and energy costs. It transforms the complex multi-robot, multi-stage, and multi-interaction scheduling problem into an optimization problem that can be solved in a finite time domain, dynamically balancing mutually constraining objectives. For example, high-value emergency tasks can appropriately relax the lower limit of power, while low-value routine tasks prioritize energy-saving return to port. Rolling time-domain optimization can respond in real time to various changes that emerge during operation, thereby avoiding planning failures such as critical requirements being blocked by routine processes and abnormal equipment being repeatedly put into use.

[0017] Furthermore, the generated collaborative scheduling strategy is simulated and verified within the digital twin model to check for path conflicts, insufficient energy, berth conflicts, and safety risks. Before the strategy actually drives the physical robots to execute, a feasibility pre-verification is performed in virtual space. The digital twin model maintains real-time state synchronization with the physical distribution center, and simulation verification can expose potential problems at low cost and without risk, such as two robots vying for the same berth or a robot having insufficient remaining power to complete its assigned task, thereby filtering out infeasible scheduling schemes. This fundamentally solves the problem of frequent conflicts, blockages, or failures during the execution phase caused by the lack of pre-verification in traditional methods, significantly improving the reliability and execution success rate of the scheduling strategy. Finally, during the execution of preset operations, the digital twin model and the rolling time-domain optimization model are corrected in a closed loop based on the real-time status data fed back by the robot, and the execution results are written back to the digital twin model. This achieves continuous synchronization between the scheduling strategy and actual operation. Even if unmodeled dynamics such as sensor noise, robot performance degradation, or environmental disturbances occur, the system can correct model deviations through real-time data to maintain the accuracy of scheduling decisions. Furthermore, the scheduling strategy can learn from each execution result and continuously optimize subsequent rolling time-domain optimization parameters. This long-term adaptive capability is particularly important in scenarios where unattended operation is required at the distribution center, and continuous operation is needed for days or even weeks. It can effectively alleviate the performance decline caused by equipment aging and task flow fluctuations, thereby continuously improving the long-term operational efficiency and resource coordination efficiency of multi-type robot clusters. Attached Figure Description

[0018] Figure 1 This is a flowchart illustrating the intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structure of the intelligent scheduling and operation and maintenance system for multi-category robot distribution centers according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention. Detailed Implementation

[0019] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Although some embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the present invention. It should be understood that the accompanying drawings and embodiments of the present invention are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0020] It should be understood that the various steps described in the method embodiments of the present invention may be performed in different orders and / or in parallel. Furthermore, the method embodiments may include additional steps and / or omit the steps shown. The scope of the present invention is not limited in this respect.

[0021] The term "comprising" and its variations as used herein are open-ended, meaning "including but not limited to"; the term "based on" means "at least partially based on"; the term "one embodiment" means "at least one embodiment"; the term "another embodiment" means "at least one additional embodiment"; the term "some embodiments" means "at least some embodiments"; and the term "optionally" means "optional embodiments". Definitions of other terms will be given in the following description. It should be noted that the concepts of "first," "second," etc., mentioned in this invention are used only to distinguish different devices, modules, or units, and are not intended to limit the order of functions performed by these devices, modules, or units or their interdependencies.

[0022] It should be noted that the terms "a" and "a plurality of" used in this invention are illustrative rather than restrictive. Those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0023] It should be noted that the information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, data stored, data displayed, etc.) and signals involved in this application are all authorized by the user or fully authorized by all parties. The collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0024] Combination Figure 1 As shown in the figure, an intelligent scheduling and operation and maintenance method for multi-category robot distribution centers provided by an embodiment of the present invention includes: Acquire status data of multiple types of robots in physical distribution centers. The status data includes: robot status, task status, berth status, and energy status.

[0025] Specifically, the system acquires real-time status data of various robot types within the physical distribution center via a data bus. This status data includes robot status (such as position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault codes and safety level), task status (such as task value, operational risk, and task progress), berth status (such as berth type, compatibility category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, available tool set, and safety isolation level), and energy status (such as energy cost, charging price, and remaining energy storage). This data is collected in real-time by sensors deployed locally at the distribution center, robot self-reporting, and infrastructure management systems (such as BMS), and synchronized to the scheduling system via a unified data bus.

[0026] Based on the robot state, the task state, the berth state, and the energy state, a unified state vector for the robot swarm and a unified state vector for the infrastructure are defined.

[0027] Specifically, based on the obtained robot states, task states, berth states, and energy states, a unified state vector for the robot swarm and a unified state vector for the infrastructure are defined. For the i-th robot, the state vector includes position, speed, remaining battery power, health index, task progress, payload type, sensor availability, communication quality, fault code, and safety level. For the j-th berth, the state vector includes berth type, compatibility category, occupancy status, estimated release time, available charging power, replaceable battery specifications, available tool set, and safety isolation level. Through this unified state vector expression, heterogeneous robots such as UAVs, quadruped robots, wheeled robots, and humanoid robots can be abstracted and compared within the same optimization framework, laying the foundation for subsequent collaborative scheduling.

[0028] A digital twin model of the physical distribution center is constructed based on the unified state vector of the robot swarm and the unified state vector of the infrastructure.

[0029] Specifically, the system constructs a digital twin model of the physical distribution center based on the unified state vectors of the robot swarm and the infrastructure. This digital twin model not only includes geometric and kinematic models, but also integrates energy consumption models, health assessment and fault models, berth service time models, and task semantic graph models.

[0030] The energy consumption model uses a time step formula: Predict changes in electricity consumption; among which, Represents robots At any moment The percentage of remaining battery power. Represents robots At any moment The percentage of remaining battery power. Indicates the time step. Represents robots The power consumption of motion, Represents robots Total battery capacity, Represents robots Additional power consumption during task execution, such as robotic arm operation, sensor data acquisition, and communication. Represents robots of, Represents robots The charging power, Represents robots At any moment The decision variable for whether or not the device is in a charging state.

[0031] The health assessment model integrates multimodal data such as battery internal resistance, temperature rise, joint current, vibration amplitude, tire or foot wear, sensor obstruction, positioning drift, communication packet loss rate, and anomaly logs, according to the formula: Calculate the health index; among which, Represents robots Health index, Indicates the first The weighting coefficients of health-related observations satisfy the following: =1, used for weighted fusion of multimodal health indicators. Represents robots The Measured values ​​of health-related parameters, such as battery internal resistance, joint current, and vibration amplitude. Represents robots The Reference values ​​for health-related observations (ideal values ​​under normal or nominal operating conditions). This represents the normalization function.

[0032] The berth service time model uses the docking probability function: Predicted service duration; among which, Represents robots The probability of successfully docking (connecting to the berth for charging or equipment replacement), This represents the Sigmoid function. , , , , This represents the learnable coefficients in the berth service time model. Represents robots The communication quality quantization value, Represents robots The quality of the positioning target, Represents robots The positioning uncertainty, Represents robots The intensity of environmental disturbance.

[0033] The task semantic graph model transforms natural language tasks generated by large language models into directed task graphs that include spatial, temporal, capability, and energy constraints. Each state change of the entity system is synchronized to the digital twin via a data bus, ensuring real-time consistency between the twin model and the physical world.

[0034] Based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue status extracted from the berth state, and the energy cost extracted from the energy state, a rolling time-domain optimization model is established, and a collaborative scheduling strategy is generated by solving the rolling time-domain optimization model.

[0035] Specifically, based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue state extracted from the berth state, and the energy cost extracted from the energy state, a rolling time-domain optimization model is established, and a collaborative scheduling strategy is generated by solving this model. The objective function of the rolling time-domain optimization model is: The factors considered include mission delays, queuing costs, energy costs, security risks, maintenance risks, and mission benefits. Let represent the overall objective function value of the rolling time-domain optimization model, and let represent the combined cost across all robots and all time steps within the optimization time domain. The index of the robot is represented by the index of the first robot. Taiwan robot, The index represents the discrete time step, and the index represents the first time step within the rolling time domain. Each time step This represents the summation over all robots and all time steps to accumulate the total global cost. Represents robots At time step Task delay costs Represents robots At time step Queuing costs Represents robots At time step Energy costs, Represents robots At time step The cost of security risks, Represents robots At time step Maintenance risk costs, Represents robots At time step Task rewards , , , , , This represents the weighting coefficients for various costs and benefits. This optimization is performed at fixed time intervals or triggered by events such as low battery alarms, emergency task arrivals, robot malfunctions, berth anomalies, and weather changes. The solution outputs the action sequence for each robot within several future time windows, including returning to port, waiting, recharging, refitting, maintenance, or continuing the task. Simultaneously, the system supports dynamic safe battery levels (dynamically adjusting the low battery threshold based on current location, return distance, environmental resistance, etc.), task relay (splitting tasks into completed and relayable segments when battery is insufficient), dynamic alliance formation, and maintenance priority decisions based on health indices.

[0036] The cooperative scheduling strategy is simulated and verified in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, insufficient energy, berth conflicts or safety risks, to obtain a verified cooperative scheduling strategy.

[0037] Specifically, before the strategy is deployed to the physical system, the collaborative scheduling strategy is simulated and verified in a digital twin model. This simulation verification is conducted in a digital twin sandbox, which uses simplified dynamics, discrete event simulation, and resource constraint models to simulate the execution of the collaborative scheduling strategy and check for path conflicts, such as overlapping time windows of a UAV landing and a ground robot passing through the same entrance; insufficient energy, such as whether the predicted remaining power is below the safety margin; berth conflicts, such as whether the berth is occupied at the target time; conflicts or safety risks in transfer mechanisms, such as whether fire, thermal runaway, or personnel safety zone conflicts are triggered. If the simulation fails, the system rolls back the strategy and re-solves the rolling time-domain optimization model until a verified collaborative scheduling strategy is obtained. Only strategies that have passed simulation verification and are conflict-free are allowed to be deployed to the physical system for execution.

[0038] In one embodiment of the present invention, the digital twin sandbox is a dedicated simulation and verification environment within the digital twin model. Essentially, it is a virtual testing platform built upon a unified state vector, simplified dynamics model, discrete event simulation model, and resource constraint model. Its design goal is to quickly evaluate the physical executability of candidate cooperative scheduling strategies with low computational cost before the strategy is issued to the real robot. In this embodiment, the core functions of the digital twin sandbox include: conflict pre-detection, simulating the execution of the scheduling strategy generated by rolling time-domain optimization in a virtual environment to check for path conflicts, such as overlap in time and space between the UAV landing path and the ground robot's return path, berth conflicts (e.g., the berth being occupied at the target time or the transfer mechanism being used by other robots), and insufficient energy (e.g., resource competition issues such as the robot's inability to safely return to the distribution center as predicted by the energy consumption model). Safety risk verification: assessing whether the strategy triggers safety risks, including but not limited to whether it enters a personnel restricted area, whether it causes battery thermal runaway, and whether it violates fire isolation rules. For high-risk operations, such as a malfunctioning robot entering the dock, the sandbox simulates whether the isolation process is effective. Strategy Rollback and Replanning: If simulation verification fails, for example, if a robot's battery is insufficient to complete the assigned task, the sandbox will reject the strategy and trigger a rolling time-domain optimization model to resolve the problem, adjusting the task order, berth allocation, or robot alliance until a validated feasible strategy is generated. Only validated strategies will be deployed to the physical system for execution. Reduced Execution Costs: By exposing problems in advance in the virtual environment, physical damage, task failures, or even safety accidents caused by conflicts, insufficient energy, or safety violations during actual robot operation are avoided, while also reducing the frequency of manual intervention due to strategy errors. In this embodiment, the digital twin sandbox utilizes the real-time synchronized state of the twin model to perform rapid, low-risk pre-simulation of scheduling decisions, ensuring that every operation deployed to the physical robot is feasible in terms of resources, energy, and safety.

[0039] The verified collaborative scheduling strategy drives multiple types of robots in the physical distribution center to perform preset operations. During the execution of the preset operations, the digital twin model and the rolling time-domain optimization model are corrected in a closed loop based on the status data fed back by the robots in real time, and the execution results are written back to the digital twin model.

[0040] Specifically, following a validated collaborative scheduling strategy, the system drives multiple types of robots in the physical distribution center to perform preset operations, including automatic docking, refueling, equipment replacement, health checks, data unloading, and redeployment. During these operations, the system performs closed-loop corrections on the digital twin model and the rolling time-domain optimization model based on real-time robot status data, and writes the execution results back to the digital twin model. Specifically, the system compares predicted energy consumption with actual energy consumption to calibrate energy consumption model parameters, updates the berth service time model using actual docking times, adds fault samples to the fault knowledge graph and triggers edge-side incremental training, and adjusts the capability graph and relay strategy based on the reasons for task failures. Each time a robot returns to port, it automatically uploads images, point clouds, thermal images, acoustic signatures, operation logs, fault codes, and energy consumption records. The system cleans, slices, and labels the data, adding abnormal data to the fault knowledge graph and adding valid samples to the model incremental training queue. After the model update is complete, the system selectively distributes model versions based on robot type and task requirements, while retaining rollback capabilities. Through this continuous closed loop of "perception-decision-simulation-execution-feedback-optimization", the system achieves long-term adaptive operation and continuous improvement of resource collaboration efficiency in multi-category robot gathering and distribution centers under unattended conditions.

[0041] This embodiment of the intelligent scheduling and operation and maintenance method for multi-category robot hubs constructs a unified state vector including robot status, task status, berth status, and energy status. Based on this, a digital twin model of the physical hub is established, achieving a globally accurate mapping of the operational status of multi-category robot clusters. This embodiment can simultaneously perceive individual robot differences, such as health indices, dynamic task characteristics such as value and risk, facility resource occupancy such as berth queues, and multi-source heterogeneous information such as energy supply costs. Global perception ensures that collaborative scheduling has sufficient information basis, avoiding uneven resource utilization due to information isolation. Secondly, a rolling time-domain optimization model is used to generate collaborative scheduling strategies. The model input includes multiple constraints such as remaining power, health index, task value, operational risk, berth queue status, and energy costs. It transforms the complex multi-robot, multi-stage, and multi-interaction scheduling problem into an optimization problem that can be solved in a finite time domain. It can dynamically balance mutually restrictive objectives. For example, the power limit can be appropriately relaxed for high-value emergency tasks, while energy-saving return to port is prioritized for low-value routine tasks. Rolling time-domain optimization can respond in real time to various changes that emerge during operation, thereby avoiding planning failures such as critical requirements being blocked by routine processes and abnormal equipment being repeatedly put into use.

[0042] Furthermore, the generated collaborative scheduling strategy is simulated and verified within the digital twin model to check for path conflicts, insufficient energy, berth conflicts, and safety risks. Before the strategy actually drives the physical robots to execute, a feasibility pre-verification is performed in virtual space. The digital twin model maintains real-time state synchronization with the physical distribution center, and simulation verification can expose potential problems at low cost and without risk, such as two robots vying for the same berth or a robot having insufficient remaining power to complete its assigned task, thereby filtering out infeasible scheduling schemes. This fundamentally solves the problem of frequent conflicts, blockages, or failures during the execution phase caused by the lack of pre-verification in traditional methods, significantly improving the reliability and execution success rate of the scheduling strategy. Finally, during the execution of preset operations, the digital twin model and the rolling time-domain optimization model are corrected in a closed loop based on the real-time status data fed back by the robot, and the execution results are written back to the digital twin model. This achieves continuous synchronization between the scheduling strategy and actual operation. Even if unmodeled dynamics such as sensor noise, robot performance degradation, or environmental disturbances occur, the system can correct model deviations through real-time data to maintain the accuracy of scheduling decisions. Furthermore, the scheduling strategy can learn from each execution result and continuously optimize subsequent rolling time-domain optimization parameters. This long-term adaptive capability is particularly important in scenarios where unattended operation is required at the distribution center, and continuous operation is needed for days or even weeks. It can effectively alleviate the performance decline caused by equipment aging and task flow fluctuations, thereby continuously improving the long-term operational efficiency and resource coordination efficiency of multi-type robot clusters.

[0043] Optionally, the step of acquiring status data of multiple types of robots in the physical distribution center includes: robot status, task status, berth status, and energy status, including: The robot status is obtained, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The task status is obtained, which includes task identifier, task type, task priority, task value, expected completion time, associated robot, task progress, and operational risk. The berth status is obtained, which includes berth identifier, berth type, berth queue status, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, and available tool set; The energy status is obtained, which includes energy type, current energy price, energy storage capacity, charging equipment status, and energy cost.

[0044] Specifically, the robot status is acquired in real time via a data bus deployed in the physical distribution center. Each robot is equipped with an onboard communication module that reports its own status data to the edge server of the distribution center at a fixed frequency. This data includes: position and speed obtained by fusing GPS, IMU, and odometer data; remaining battery power collected by the battery management system; a health index calculated using a health index model based on multimodal data such as battery internal resistance, temperature rise, motor current, joint vibration, tire or foot wear, sensor occlusion, positioning drift, communication packet loss rate, and anomaly logs; current task progress read from the task manager; payload type identified through the payload interface, such as visible light camera, infrared thermal imager, robotic arm, spraying device, etc.; sensor availability obtained through a self-test program; communication quality obtained through wireless link quality assessment; and fault codes generated by the fault self-diagnosis module and a safety level preset according to the task scenario. All robot status data is aggregated via the data bus, stored in a real-time database, and synchronized to the digital twin model.

[0045] Simultaneously, task status is retrieved from the task scheduling queue. Tasks can originate from user-submitted task instructions via a graphical interface or natural language interface, or be issued by higher-level business systems, such as power inspection platforms or logistics management systems. Each task is assigned a unique identifier, task type (e.g., infrared inspection, material handling, emergency response, data collection), priority (assessed based on business urgency and regional importance), task value (quantified based on expected benefits, time sensitivity, historical anomaly probability), expected completion time, currently associated robot identifier, task progress (e.g., not started, in progress, paused, completed, or failed), and operational risk level (e.g., entering a dangerous area, operating high-voltage equipment, nighttime operation). Task status is dynamically updated as scheduling progresses; for example, when a task is assigned to a robot, the associated robot field is filled with its specific number; when the robot completes a task segment, the task progress is updated. Changes in task status are also synchronized in real-time to the digital twin model via the data bus, allowing the rolling time-domain optimization model to extract task value and operational risk. Berth status is retrieved from the distribution center infrastructure controller. Each berth, including charging berths, battery swapping berths, equipment replacement berths, and maintenance isolation berths, is equipped with a local controller that monitors and reports: berth identification, berth type (e.g., fast charging berth, slow charging berth, battery swapping bay, tool replacement station, fault isolation bay), berth queue status (number of robots currently in the queue and their order), compatible robot categories (e.g., only supporting a certain type of drone or quadruped robot), current occupancy status (idle, occupied, or under maintenance), estimated release time, and if occupied, predicts the remaining service time based on the berth service time model, available charging power (maximum supported output power), replaceable battery specifications (battery model and available quantity), and available tool set (e.g., screwdrivers, grippers, cleaning brushes). This data is periodically collected by the edge controller at the distribution center or triggered by berth status change events, such as robot docking, docking, or battery swapping completion, and pushed to the scheduling system via the data bus.

[0046] Energy status is obtained from the energy management system at the distribution center. Energy status includes: energy type (e.g., grid power, photovoltaic energy storage, battery swapping stations, diesel generators); current energy price (combined with peak / valley electricity prices or real-time electricity price information); energy storage capacity (the remaining percentage of energy storage battery capacity); charging equipment status (online, faulty, or in use status of each charging port); and comprehensive energy cost, which can be dynamically calculated based on current electricity price, charging efficiency, and remaining energy storage capacity. Energy status is read from charging piles, energy storage inverters, battery management systems, and other devices via industrial protocols such as Modbus, CAN, or OPC UA, and synchronized to the digital twin model via the data bus, serving as input for the energy cost in the rolling time-domain optimization model. If the distribution center is equipped with a multi-energy complementary system, it will also acquire information such as predicted photovoltaic power generation and energy storage discharge depth to support short-term energy price forecasting.

[0047] In this optional embodiment, by simultaneously sensing individual robot differences, such as health index, payload type, sensor availability, task dynamic characteristics such as value, risk, expected completion time, facility resource occupancy such as berth queue status, expected release time, tool set, and energy supply constraints such as electricity price, energy storage capacity, and charging equipment status, a unified access system for multi-source heterogeneous data is constructed. This allows subsequent collaborative scheduling to move beyond mechanical decisions based on fixed thresholds. Instead, it enables a comprehensive consideration of the robot's actual health level, the actual urgency of the task, the precise service window of the berth, and real-time fluctuations in energy costs. This avoids resource imbalances caused by information fragmentation at the source, such as high-health robots being idle while sick robots are repeatedly dispatched, low-value tasks occupying high-value berths, and urgent tasks being delayed due to congestion in the ordinary refueling queue. Simultaneously, all status data is synchronized in real-time to the digital twin model via a data bus, ensuring dynamic consistency between the twin and the physical world. This provides a reliable data foundation for subsequent simulation verification and closed-loop correction, significantly improving the long-term operational efficiency and resource coordination efficiency of multi-type robot clusters under unattended conditions.

[0048] Optionally, defining a unified state vector for the robot swarm and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state includes: The parameter order of the preset robot swarm state vector is defined, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The parameter values ​​corresponding to each robot are obtained from the robot states, and the parameter values ​​corresponding to each robot are assigned according to the preset robot swarm state vector to obtain the state vector of that robot. The state vectors of all robots constitute the unified state vector of the robot swarm. The parameters of the preset infrastructure state vector are ordered, including berth type, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, available tool set, and security isolation level. The parameter values ​​corresponding to each berth are obtained from the berth status, and the parameter values ​​corresponding to each berth are assigned according to the preset infrastructure state vector to obtain the state vector of that berth. The state vectors of all berths constitute the unified state vector of the infrastructure.

[0049] Specifically, in this embodiment, the parameter order of the robot swarm state vectors can be predefined in memory or a database. This order is fixed as follows: position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. This preset order ensures that all robot state vectors have the same dimensions and field arrangement, providing a unified data structure for subsequent batch matrix operations and optimization solutions. Based on the robot status data acquired in the previous step, the system extracts the corresponding parameter values ​​for each robot: position coordinates and velocity vectors are extracted from the fusion results of GPS, IMU, and odometry reported by the robots; the remaining battery percentage is read from the battery management system; the latest calculated health index is obtained from the health assessment module; the current task progress, such as the percentage completed or the current subtask number, is read from the task manager; the type of the currently mounted payload is identified through the payload interface and converted into a preset code, for example, 1 represents a visible light camera, 2 represents an infrared thermal imager, and 3 represents a robotic arm; sensor availability is summarized from the self-test results, for example, using bitmasks to represent the online status of each sensor; the overall signal-to-noise ratio and packet loss rate scores are obtained from the communication quality assessment module; the latest fault code is read from the fault self-diagnosis log, which is 0 if there is no fault; and the system also includes preset safety levels based on the task and safety policy, such as 1 for low risk, 2 for medium risk, and 3 for high risk. For missing or abnormal data, the system fills in the missing data with default values ​​or valid values ​​from the previous time step. After extracting the parameter values, the system assigns values ​​to each robot according to a pre-defined parameter order, forming a fixed-dimensional numerical vector that serves as the robot's state vector. Arranging all robot state vectors in order of robot number or timestamp creates a unified state vector for the robot swarm. This vector set can be directly input into subsequent digital twin models and rolling temporal optimization models.

[0050] Similarly, the parameter order of the infrastructure state vector is predefined, and this order is fixed as follows: berth type, compatible robot category, current occupancy status, estimated release time, available charging power, swappable battery specifications, available tool set, and security isolation level. The system extracts corresponding parameter values ​​for each berth in the distribution center from the previously acquired berth status data: berth type is determined based on the berth hardware configuration, for example, 1 represents a fast charging berth, 2 represents a slow charging berth, 3 represents a battery swapping compartment, 4 represents a tool replacement station, and 5 represents a fault isolation compartment; compatible robot category is represented by a bitmask or enumeration list indicating the robot models or categories that the berth can serve, for example, simultaneously compatible with drones and quadruped robots; current occupancy status is represented by 0 for idle, 1 for occupied, and 2 for under maintenance; estimated release time is a timestamp, set to 0 or infinity if currently idle; available charging power is measured in kilowatts or watts; replaceable battery specifications include battery model code and the number of currently available batteries, which can be encoded as a string or structure; available tool set is represented by a list or bitmask indicating the types of tools equipped on the berth, such as screwdrivers, grippers, cleaning brushes, etc.; safety isolation level indicates whether the berth has fault isolation function and its protection level, for example, 0 for no isolation, 1 for ordinary isolation, and 2 for fireproof and explosion-proof isolation. The system assigns these parameter values ​​one by one to the corresponding positions in the berth state vector according to a preset order, obtaining the state vector for each berth. If a berth lacks a certain parameter, such as a regular charging berth not having a replaceable battery specification, it is filled with a null value or 0. After repeating the above operation for all berths, the state vectors of all berths are combined in order of berth number to form a unified state vector for the infrastructure. This vector set reflects the real-time availability and service capabilities of all berth resources within the distribution center, and together with the unified state vector of the robot swarm, it constitutes the core input of the digital twin model, enabling heterogeneous robots and heterogeneous berths to be uniformly optimized at the same level of abstraction.

[0051] In this optional embodiment, by pre-setting a fixed parameter order and extracting and filling parameter values ​​for each robot and each berth according to a unified specification, heterogeneous robots of different types and with scattered data structures, such as drones, quadruped robots, wheeled AMRs, and humanoid robots, as well as heterogeneous berths, such as fast-charging berths, slow-charging berths, battery swapping compartments, tool changing stations, and fault isolation compartments, are transformed into a set of numerical vectors with consistent dimensions and standardized arrangement. This fundamentally eliminates the differences in data format between heterogeneous entities. This standardized expression enables the subsequent digital twin model to process multi-source inputs in a unified mathematical form, without the need to write special analytical logic for different robots or berths, significantly reducing the complexity of the model. At the same time, the standardized vector structure supports efficient matrix operations and batch processing, providing computational feasibility for the rolling time-domain optimization model to quickly solve the cooperative scheduling strategy within a fixed time window. Furthermore, the unified state vector ensures the alignment and comparability of the robot swarm state and the infrastructure state on the time axis, enabling the system to clearly characterize cross-category matching relationships such as a drone with low remaining battery needing to occupy a certain type of fast charging berth. This provides a clear state input interface for advanced functions such as task relay, dynamic alliance formation, and berth queue optimization, thereby significantly improving the scheduling accuracy and resource coordination efficiency of multi-category robot gathering and distribution centers under unattended conditions.

[0052] Optionally, constructing a digital twin model of the entity distribution center based on the unified state vector of the robot swarm and the unified state vector of the infrastructure includes: Based on the remaining power in the unified state vector of the robot swarm, an energy consumption model for the physical distribution center is established. A health assessment model for the physical distribution center is established based on the health index in the unified state vector of the robot swarm. Based on the berth type, available charging power, and swappable battery specifications in the unified state vector of the infrastructure, a berth service time model for the physical distribution center is established. Based on the task semantic information in the task status, establish a task semantic graph model of the entity distribution center; The energy consumption model, the health assessment model, the berth service time model, and the task semantic graph model are fused together to obtain a digital twin model of the physical distribution center.

[0053] Specifically, based on the remaining battery power parameter in the unified state vector of the robot swarm, an energy consumption model for the physical distribution center is established. This model uses an integral form with discrete time steps to calculate the robot's battery power change in each time step: the robot's current remaining battery power minus the battery consumption caused by motion power consumption and task power consumption, plus the battery replenishment after charging power is converted to charging efficiency during charging. Motion power consumption is obtained through table lookup or regression models based on robot type and motion state, such as drone flight speed, quadruped robot gait, and wheeled robot acceleration; task power consumption is dynamically estimated based on payload type, such as visible light camera, infrared thermal imager, robotic arm operation, and task intensity. Charging efficiency is obtained by fitting historical charging curves provided by the battery management system. This energy consumption model is used in the digital twin to predict battery power changes during future task execution and return processes, and to determine whether there is a risk of insufficient energy.

[0054] A health assessment model for the physical distribution center is established based on the health index parameters in the unified state vector of the robot swarm. This model integrates multimodal sensor data, including battery internal resistance, temperature rise, joint current, vibration amplitude, tire or foot wear, sensor occlusion ratio, positioning drift, communication packet loss rate, and the frequency of fault codes in the anomaly log. For each type of health observation, the system normalizes and compares the measured value with a preset reference value. The normalization function can use minimum-maximum normalization or sigmoid transformation, mapping the difference to the interval between 0 and 1. Then, the normalization results are weighted and summed according to preset weight coefficients. Finally, the health index is obtained by subtracting the weighted sum from 1. The closer the health index is to 1, the healthier the robot is; the closer it is to 0, the more urgent maintenance is required. This health assessment model outputs the health score of each robot in real time in the digital twin, mapping it to four states: dispatchable, restricted dispatch, pending inspection, and isolated, providing maintenance risk cost input for rolling optimization.

[0055] Based on the berth type, available charging power, and swappable battery specifications in the unified state vector of the infrastructure, a berth service time model for physical distribution centers is established. The core of this model is a docking success probability function, which adopts a sigmoid form. Input variables include robot communication quality (signal-to-noise ratio and packet loss rate), target positioning quality (visual marker sharpness and QR code recognition score), positioning uncertainty (standard deviation of position estimation), and environmental disturbance intensity such as wind speed, rain / snow, ground friction changes, and illumination changes. Model coefficients are obtained by fitting historical docking data and are used to characterize the influence of each factor on the docking success rate. The docking success probability is multiplied by a baseline service time (e.g., charging time is determined by battery capacity and charging power, battery swapping time by robotic arm movement speed, and tool replacement time by tool type) to obtain the expected service time. This is further calibrated using real-time feedback of the expected release time. In a digital twin, this model is used to predict the remaining service time of each berth in its current occupancy state, serving as a basis for queuing optimization and berth allocation.

[0056] Based on the task semantic information in the task status, a task semantic graph model of the entity distribution area is established. When a user inputs complex task instructions through a natural language interface, such as performing infrared inspection of the power distribution cabinet in area A and dispatching a quadruped robot to a narrow area for verification when an anomaly is detected, and then calling a mobile operation robot for handling, the natural language is decomposed into a structured task graph by calling a large language model or a visual language action model. The task graph consists of nodes and directed edges. Nodes include task nodes such as infrared inspection, verification, and handling; resource nodes such as required robot type, payload, and tools; spatial nodes such as power distribution cabinet in area A and narrow area; and time constraint nodes such as expected completion time and sequential dependencies. Directed edges represent pre- and post-constraints, collaborative relationships, and risk propagation paths. The generated task graph then undergoes capability constraint verification to check whether a robot has the required payload and obstacle-crossing ability; energy constraint verification to estimate whether the energy consumption of each sub-task exceeds the robot's remaining power; safety constraint verification to check whether a restricted area has been entered or there is a risk of collision; and permission constraint verification to check whether operation permission has been obtained. Only task graphs that pass all verifications will be stored in the digital twin model and used to guide subsequent task decomposition, dynamic alliance formation, and relay decisions.

[0057] Finally, the energy consumption model, health assessment model, berth service time model, and task semantic graph model are fused to obtain a digital twin model of the physical distribution center. The fusion process is implemented via a data bus: all models receive input from a unified state vector and real-time updated state data, and the models interact through a shared database. For example, the remaining power predicted by the energy consumption model is passed to the task semantic graph model for energy constraint verification; the health index output by the health assessment model is passed to the berth service time model to determine whether priority allocation to isolated berths is necessary; and the release time predicted by the berth service time model is fed back to the energy consumption model to calculate the standby energy consumption of the robot while waiting to recharge. The digital twin model stores the parameters and intermediate results of the four sub-models using a unified data structure and provides a standardized query interface for the rolling time-domain optimization model. Every state change of the physical system—whether it's a robot position update, power decrease, health indicator change, berth occupancy status switch, or tool inventory change—is synchronized in real-time to the corresponding sub-model in the digital twin model via the data bus, ensuring that the twin remains consistent with the physical world. This fused digital twin model forms the unified virtual mirror foundation for subsequent rolling optimization and simulation verification.

[0058] In one embodiment of the present invention, the energy consumption model refers to a mathematical model used to predict the change of remaining battery power over time during robot movement, task execution, and charging. Based on the principle of energy conservation, this model divides the robot's power consumption into two parts: movement power consumption and task power consumption. Power replenishment is divided into the portion of charging power after efficiency conversion. Specifically, it is constructed using a recursive formula with discrete time steps. Within each time step, the current remaining battery power is calculated by subtracting the product of movement power consumption and the time step (divided by the total battery capacity), then subtracting the product of task power consumption and the time step (divided by the total battery capacity), and finally adding the product of charging power, charging efficiency, charging state variables, and the time step (divided by the total battery capacity).

[0059] For example, a quadruped robot walking on flat ground at a speed of 0.5 meters per second consumes 50 watts of power during movement; simultaneously, carrying an infrared thermal imager to perform temperature measurement consumes 20 watts of power; with a total battery capacity of 200 watt-hours and a current remaining charge of 60%, and a time step of 1 second, the battery charge decreases by approximately 0.0039 per ten thousand within one step. If the robot returns to its berth to recharge, with a charging power of 300 watts and a charging efficiency of 90%, the battery charge increases by approximately 0.375 per ten thousand per second. Using this model, the system can predict in the digital twin whether the robot's battery charge will fall below the safety margin five minutes in advance, thereby triggering a return-to-port or relay decision.

[0060] A health assessment model is a mathematical model used to quantify the overall health status of a robot. This model integrates multiple sensor indicators and calculates a health index between 0 and 1 through weighted normalization. The specific construction method is as follows: First, select health observation indicators, including battery internal resistance, battery temperature rise rate, joint motor current fluctuation, structural vibration amplitude, tire or foot wear, visual sensor image occlusion ratio, positioning module drift error, communication link packet loss rate, and system log anomaly frequency. Set reference values ​​for each indicator; for example, the reference value for the battery internal resistance of a new robot is 10 milliohms, and the reference value for temperature rise is 0.5 degrees Celsius per minute. Map the difference between the measured values ​​and the reference values ​​to the 0-1 interval using a normalization function; commonly used normalization functions are Min-Max normalization or the Sigmoid function. Then, assign weights to each indicator, with the sum of all weights being 1. The final health index equals 1 minus the sum of the normalized values ​​of each indicator multiplied by their weights.

[0061] For example, the measured internal resistance of a drone's battery is 15 milliohms, which is 5 milliohms higher than the reference value, and after normalization, it becomes 0.2; the normalized motor current fluctuation is 0.1; the normalized positioning drift is 0.05; and other indicators are normal (0). Assuming the weights are set as follows: battery internal resistance 0.3, motor current 0.2, positioning drift 0.1, and the remaining indicators 0.4, the weighted sum is 0.3 x 0.2 + 0.2 x 0.1 + 0.1 x 0.05 = 0.085, and the health index is 1 minus 0.085 = 0.915. An index higher than 0.9 indicates that the robot can be dispatched normally, while an index lower than 0.6 triggers a maintenance pending inspection state.

[0062] The berth service time model is a mathematical model used to predict the total time required for a robot to complete recharging or reloading and release from the berth. The model consists of two parts: a successful docking probability function and a baseline service time calculation. The successful docking probability function uses a sigmoid form, and the input variables include the signal-to-noise ratio of communication quality between the robot and the berth, the sharpness of the visual target recognition, the standard deviation of the robot's positioning uncertainty, and the intensity of environmental disturbances such as wind speed or ground vibration. The model coefficients are obtained by logistic regression using historical successful and failed docking data. The baseline service time is determined based on the berth type: for fast-charging berths, the baseline service time is the remaining battery capacity divided by the charging power and then multiplied by a safety factor; for battery swapping berths, the baseline service time is the total action time for the robotic arm to grasp, disassemble, and install the battery; for tool changing stations, the baseline service time is the execution time for changing tools. The final expected service time equals the successful docking probability multiplied by the baseline service time, plus a reserved buffer time.

[0063] For example, in a battery swapping berth, based on the docking probability function fitted from historical data, the current robot's communication quality is 0.9, the target quality is 0.8, the positioning uncertainty is 0.05 meters, and the environmental disturbance intensity is 0.2. Substituting these values ​​into the Sigmoid function, the calculated docking success probability is 0.95. The robot's battery needs replacing, and the baseline service time for the robotic arm's battery swap is 30 seconds. Therefore, the expected service time is 0.95 multiplied by 30 seconds, equaling 28.5 seconds. Adding a 2-second buffer, the berth is expected to be released after 30.5 seconds. If the docking success probability is only 0.6, the expected service time is 18 seconds, but the risk of failure is high. In this case, the system will reserve more buffer time or reallocate the berth.

[0064] A task semantic graph model is a formalized representation of task instructions described in natural language, converted into a directed graph structure. This directed graph contains task nodes, resource nodes, spatial nodes, temporal constraint nodes, and edges representing dependencies. The specific construction method is as follows: First, semantic parsing of user input is performed using a large language model or a visual language action model to extract information such as action sequences, required robot capabilities and payloads, work location, preconditions, and post-influences. Then, this information is transformed into graph nodes: each atomic action, such as "infrared inspection," is a task node; "drone" and "infrared camera" are resource nodes; "distribution cabinet A area" is a spatial node; and "start within 5 minutes" is a temporal constraint node. Directed edges represent sequential dependencies, for example, "After the drone completes the infrared inspection, if an anomaly is found, a quadruped robot will be dispatched for verification." The task semantic graph model also includes a verification function, checking whether each task node in the graph can find at least one robot with the required payload and movement capabilities in the capability map of the current distribution area, and whether the total energy consumption of all sub-tasks is within the robot's remaining battery power. For example, a user inputs, "Conduct fire reconnaissance on the warehouse on the east side of the factory area. If smoke is detected, dispatch a firefighting robot to extinguish the fire." The task semantic graph model decomposes this into two parallel task nodes: drone reconnaissance and a firefighting robot on standby. A conditional trigger edge connects the reconnaissance node and the firefighting node, with the trigger condition being that the reconnaissance results contain a smoke detection tag. Resource constraints are also bound to the reconnaissance node: a vertical take-off and landing drone needs to be equipped with visible light and thermal imaging cameras; resource constraints are bound to the firefighting node: a tracked firefighting robot needs to be equipped with fire extinguishing bombs. The system then checks whether there are robots at the current distribution center that meet the above constraints. If so, the task graph passes the verification and is stored in the digital twin model.

[0065] The digital twin model of a physical distribution center refers to a unified virtual mirror image composed of the aforementioned energy consumption model, health assessment model, berth service time model, task semantic graph model, and the distribution center's geometric model, kinematic model, data bus, and real-time database. It is not a single algorithm or data structure, but a multi-model integrated software entity running on the distribution center's edge server or cloud platform. This digital twin model has the following characteristics: First, the digital twin model of the physical distribution center synchronizes its status in real time with the physical equipment of the physical distribution center, including robots, berth controllers, charging piles, tool bays, and sensor networks, via the data bus; Second, the digital twin model of the physical distribution center encapsulates the input and output interfaces of the above four core sub-models. For example, externally, the energy consumption model can be called to input the robot ID and future motion trajectory, and output a predicted remaining battery power curve; Third, the digital twin model of the physical distribution center provides a sandbox simulation environment that can load candidate scheduling strategies, simulate execution in virtual space, and detect conflicts; Fourth, it maintains a unified state database, storing the unified state vectors of the robot swarm and the infrastructure. For example, a distribution center deploys five drones, three quadruped robots, and six berths. The digital twin model stores the real-time location, battery level, and health index of each robot; the occupancy status and estimated release time of each berth; and simultaneously runs an energy consumption model to predict the battery depletion of each robot in real time, a health assessment model to update the health index every five minutes, a berth service time model to dynamically update the remaining service time of each berth, and a task semantic graph model to store the currently executing and pending tasks. When the scheduler generates a policy requiring a drone with 35% battery to perform an inspection task estimated to consume 20% of its battery before returning to port to recharge, the energy consumption model in the digital twin will simulate and predict that the drone's battery level will drop to 10% upon return, below the safety margin of 15%, thus triggering a sandbox verification failure. The scheduler needs to readjust the policy, for example, by changing it to return to port to recharge before executing the task. In this way, the digital twin model ensures the feasibility of scheduling decisions in terms of energy, health, berth resources, and safety constraints.

[0066] In this optional embodiment, the core energy dynamics, health decay, berth service efficiency, and task semantic constraints of robot operation are unified into the same virtual image. This allows the system to simultaneously predict whether the battery can support a return trip, whether the robot's health is suitable for dispatch, when the berth will be available, and whether the task graph is executable before scheduling decisions are made. Furthermore, sandbox simulation can detect insufficient energy, operation with defects, berth conflicts, or task logic flaws in advance, thereby completely avoiding problems such as low-battery robot shutdowns, repeated dispatches of faulty robots, berth resource deadlocks, and unexecutable task plans generated by large models caused by the lack of a single-dimensional model in traditional methods. Simultaneously, four... The sub-models interact and perform rolling calibration in real time through a data bus. For example, the power prediction results of the energy consumption model are directly used to verify the energy constraints of the task semantic graph model. The health index of the health assessment model affects whether to prioritize the allocation of isolated berths in the berth service time model. The release time predicted by the berth service time model is fed back to the energy consumption model to estimate the waiting energy consumption. This twin architecture with multiple coupled models makes scheduling decisions no longer dependent on fixed thresholds or human experience, but based on high-fidelity simulation of the global physical processes of the distribution center. This significantly improves the long-term operational reliability, resource utilization efficiency, and task success rate of multi-type robot clusters under unattended conditions.

[0067] Optionally, the step of establishing a rolling time-domain optimization model based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue state extracted from the berth state, and the energy cost extracted from the energy state includes: Based on the stated task value, construct task benefit items; Based on the remaining electricity, the energy cost, and the predicted energy consumption of the return route, an energy cost item is constructed. Based on the aforementioned health index, construct maintenance risk items; Based on the described operational risks, construct safety risk items; Based on the berth queue status, construct a queuing cost item; The objective function is obtained by weighted summing of the task benefit item, the energy cost item, the maintenance risk item, the security risk item, and the queuing cost item; Based on the objective function and preset constraints, the rolling time-domain optimization model is established.

[0068] Specifically, a task reward term is constructed based on the task value extracted from the task status. The task value is determined by factors including business urgency, regional importance, time sensitivity, historical anomaly probability, and manually set weights. The system calculates the reward for each robot at each time step for the task it is currently performing. For example, the reward is positive when a high-value inspection task is completed and increases linearly with task progress; if the task is delayed or canceled, the reward decreases or even becomes negative. The task reward term appears with a negative sign in the objective function, making maximizing the total task reward equivalent to minimizing its negative value.

[0069] An energy cost item is constructed based on the remaining power extracted from the robot's status, the energy cost extracted from the energy status, and the predicted energy consumption of the return route. The energy cost includes the electricity consumed during charging. The system calculates the cost of charging the robot by combining the current peak / valley electricity price or the real-time electricity price for energy storage. Simultaneously, for robots performing tasks, the system predicts the total power consumption from their current location to completing the sub-task and returning to the distribution center based on the energy consumption model. If the remaining power is insufficient to support a safe return, a significant penalty cost is imposed. Furthermore, this embodiment also considers charging efficiency losses, converting the energy wasted due to heat dissipation during charging into a cost. The energy cost item encourages robots to charge during periods of low electricity prices and avoid risky operations with low power levels.

[0070] A maintenance risk item is constructed based on a health index extracted from the robot's status. This health index, output in real-time by a health assessment model, ranges from 0 to 1. The health index is mapped to maintenance risk cost: a lower health index indicates potential problems such as battery degradation, joint wear, and sensor drift. Continuing to assign tasks may lead to mid-task failures, thus increasing maintenance risk cost. In practice, the system uses an exponential or piecewise linear function. When the health index falls below a preset threshold, the maintenance risk cost increases sharply, potentially even forcing the robot to be set to "unassignable" or "isolated." The maintenance risk item is also used to decide whether to schedule maintenance in advance: when the maintenance risk cost exceeds the task's benefits, the system prioritizes sending the robot back to port for inspection.

[0071] Safety risk items are constructed based on operational risks extracted from task status. Operational risks include robots entering hazardous areas such as near high-voltage power towers or chemical plant leak areas, operating hazardous equipment such as opening and closing high-voltage switches, operating at night or in inclement weather, and interacting with densely populated areas. Each task is labeled with its risk level by a semantic decomposition model or manually when it is created. The system quantifies the risk level into a safety risk cost. High-risk tasks have high cost values, causing the optimization model to tend to assign such tasks to robots with higher safety protection levels or to perform them during low-risk periods. If a strategy causes the robot to violate safety boundaries, such as entering a no-fly zone or exceeding the safe distance for personnel, the safety risk item will be set to a maximum value, causing the strategy to be automatically eliminated in the optimization process.

[0072] A queuing cost term is constructed based on the berth queue status extracted from the berth status. The berth queue status includes the number of waiting robots in front of each berth, the estimated waiting time for each robot, and the service time distribution of the berth. The system accumulates the waiting time of each robot in the berth queue at each time step and converts the waiting time into queuing cost. The queuing cost can be designed as a linear or convex function of the waiting time to avoid some robots being starved for a long time. At the same time, the system considers the task priority of different robots: high-value or urgent tasks have higher waiting cost coefficients, thus gaining the right to jump the queue in queuing optimization. The queuing cost term is also used to balance berth load and prevent all robots from rushing to the same fast-charging berth while other berths are idle.

[0073] Furthermore, the objective function is obtained by weighted summation of the aforementioned task benefit items, energy cost items, maintenance risk items, safety risk items, and queuing cost items. In this embodiment, delay cost can be calculated based on the difference between the expected and actual task completion times, and processed separately from queuing cost to allow for more refined optimization. The weights of each item can be dynamically adjusted according to the operational scenario. For example, in a power safety scenario, the safety risk weight can be increased; in a logistics scenario, the throughput weight can be increased while the queuing cost weight can be decreased; and in an emergency rescue scenario, the task benefit weight can be increased while the energy cost weight can be decreased. Weight adjustments can be configured manually or automatically learned by the system based on historical operational indicators.

[0074] A rolling time-domain optimization model is established based on the objective function and preset constraints. The preset constraints include: a power constraint, requiring the robot's remaining power at any given time to not fall below a dynamic safety threshold, which is determined by the robot's current position, return distance, environmental resistance, load weight, positioning uncertainty, and task urgency, rather than a fixed value; a berth occupancy mutual exclusion constraint, ensuring that only one robot can occupy the same berth at any given time; a task timing constraint, for example, requiring inspection to be completed before verification; a robot capability constraint, ensuring that only robots with the required load and obstacle-crossing capabilities can be assigned to the corresponding tasks; and a safety isolation constraint, requiring faulty robots to enter isolated berths and not share berths or tools with normal robots. The rolling time-domain optimization model employs a model predictive control framework, starting from the current time, predicting the state evolution over a finite future time domain, such as the next 30 minutes, and solving for the optimal action sequence. However, only the decision at the first time step is executed, and the solution is recalculated at the next sampling time, thus achieving rolling optimization. The model executes at fixed time intervals, such as 30 seconds, or is triggered by events such as low battery alarms, emergency mission arrivals, robot malfunctions, berth anomalies, and weather changes. The solver can employ mixed-integer linear programming or heuristic search algorithms to output the sequence of actions for each robot within several future time windows, including returning to port, waiting, recharging, changing equipment, maintenance, or continuing to perform tasks.

[0075] In one embodiment of the present invention, the rolling time-domain optimization model is a discrete-time dynamic optimization model based on a model predictive control framework. It is used to make real-time decisions on the actions of each robot in a multi-category robot cluster. At each decision moment, starting from the current real state, the system uses the energy consumption model, health assessment model, berth service time model, and task semantic graph model in the digital twin to predict the evolution of the system state within a finite time window. Then, within this time window, an optimization problem with multiple constraints is solved to obtain the optimal action sequence for each robot at each future time step, but only the first action in the sequence is executed. When the next decision moment arrives, the system re-collects the latest state data and solves the problem again in a rolling manner. The rolling time-domain optimization model belongs to the mixed-integer programming model. Its structure includes five elements: First, state variables, namely the unified state vector of the robot swarm and the unified state vector of the infrastructure, covering each robot's position, remaining battery power, health index, current task progress, as well as the occupancy status, estimated release time, and available tool set of each berth; Second, control variables, which are the actions each robot can be assigned at each time step, such as continuing to execute a task, moving to a berth, queuing at a berth, starting or stopping charging, performing equipment changes, entering maintenance isolation, or being assigned a new task; Third, the prediction model, provided by four sub-models in the digital twin. Given the current state and control variables, the energy consumption model outputs the remaining battery power at future time steps, the health assessment model outputs the trend of health index changes, and the berth service time model outputs the berth occupancy time... The probability distribution between them; fourth, the objective function, which comprehensively weighs the costs of task delays, queuing, charging, operational safety, maintenance, and task benefits, and obtains a total cost through weighted summation. The optimization objective is to minimize this total cost; fifth, the constraints, including dynamic safety power constraints (the robot's remaining power at any time must not be lower than a threshold calculated in real time based on return distance, environmental resistance, and task urgency), berth mutual exclusion constraints (only one robot can serve the same berth at a time), task timing constraints (the sequential dependencies between subtasks), capability matching constraints (the robot must have the required payload and obstacle-crossing capabilities), safety isolation constraints (robots with low health indices can only enter isolated berths), and resource availability constraints (the battery swapping compartment must have matching batteries, and the tool station must have the specified tools), etc. In actual solution, this embodiment uses a mixed integer programming solver or a heuristic optimization algorithm. For a typical distribution center, such as ten robots, five berths, and ten time steps in the prediction time domain, each time step being thirty seconds, the number of decision variables in this problem is within several hundred. Modern computing hardware can complete the solution within seconds or even hundreds of milliseconds. The system has pre-set priority rules, such as emergency task preemption and low-battery priority return to port, to generate an initial feasible solution. Then, it optimizes the solution through local search or branch and bound methods to ensure real-time performance.In the action sequence output by the solution, only the decision of the first time step is sent to the physical robot for execution. After execution, at the next decision moment, such as after thirty seconds or triggered asynchronously by events such as low battery alarm, emergency task insertion, robot failure, berth abnormality, or sudden weather change, the system rereads the unified state vector and performs rolling optimization again.

[0076] For example, at a certain distribution center, there are currently three robots: a drone with 20% battery power returning to port, a quadruped robot with 80% battery power idle, and a wheeled robot with 50% battery power performing an inspection task. An emergency task arrives, requiring a robot to be dispatched to a certain area to take photos within 30 seconds. The rolling time-domain optimization model compares multiple strategies in the prediction time domain: dispatch the quadruped robot to perform the task, the drone returns to port to recharge, and the wheeled robot continues its inspection; or dispatch the wheeled robot to interrupt its current task to perform the emergency task, and the quadruped robot takes over the inspection. The energy consumption model predicts the battery power change under each strategy, the berth service time model predicts the berth occupancy, the objective function weighs the high benefits of the emergency task against the delay cost caused by interrupting the original task, and the constraints ensure that the battery power of all robots does not fall below the dynamic safety threshold. After the model solves, it outputs the current step action, for example, dispatching the quadruped robot to perform the emergency task, and the system immediately issues the execution order. Thirty seconds later, the quadruped robot is already on its mission, the drone enters the berth to start charging, and the system rolls again to continue optimizing based on the latest status. Through this rolling optimization, the model can continuously make near-optimal decisions that satisfy all physical and safety constraints in a dynamically changing environment.

[0077] It should also be noted that, based on the above description and combined with well-known model predictive control theory and mixed-integer programming methods, those skilled in the art can implement this rolling time-domain optimization model on the edge controller of the distribution center or on the cloud server, which will not be elaborated further here. The essential difference between this model and fixed threshold rules or single-dimensional scheduling methods is that the model in this embodiment simultaneously considers constraints and objectives in five dimensions: energy, health, task, berth, and safety, and makes forward rolling predictions on the time axis, thereby possessing foresight and adaptive capabilities.

[0078] In this optional embodiment, several conflicting and dynamically changing core objectives in the operation of multi-category robot hubs—namely, maximizing task benefits, minimizing energy consumption and charging costs, controlling equipment health and maintenance risks, avoiding operational safety risks, and improving queuing efficiency—are mathematically quantified and weighed through a unified objective function. This allows scheduling decisions to move beyond simple rules such as fixed thresholds or first-come, first-served, and can dynamically adjust the weights of various items based on real-time status. For example, when an emergency task arrives, the weight of the task benefit item is automatically increased to give it priority scheduling; when the robot's health index declines, the maintenance risk item rises rapidly, triggering proactive isolation or early maintenance; and during periods of low electricity prices, the weight of the robot can be adjusted accordingly. The energy cost item is reduced to encourage slow charging; at the same time, the energy cost item introduces predicted energy consumption on the return route, so that the low battery threshold is no longer a fixed value but is dynamically calculated based on the return distance and environmental resistance, fundamentally avoiding the risk of robots stopping halfway due to insufficient power; the maintenance risk item uses a health index to quantify the probability of equipment aging and failure, effectively preventing the repeated assignment of defective robots, which would lead to task failure or secondary damage; the safety risk item incorporates the hazard level of the working environment into the optimization, so that high-risk tasks are prioritized for robots with stronger safety protection or low-risk periods; the queuing cost item avoids congestion and starvation caused by multiple robots competing for the same berth by modeling the berth queue status and waiting time. These cost and benefit items are weighted and summed to form an optimizable objective function. Combined with constraints on power supply, berth mutual exclusion, task timing, capability, and safety isolation, the constructed rolling time-domain optimization model can proactively solve for the action sequence with the minimum overall cost at each decision moment. This achieves synergistic optimization across five dimensions: energy, health, safety, task value, and berth resources, significantly improving the long-term operational efficiency, resource utilization, and task success rate of multi-category robot clusters in unattended distribution centers.

[0079] Optionally, the step of generating a cooperative scheduling strategy by solving the rolling time-domain optimization model includes: Starting from the current time, the prediction time domain length is set, and the rolling time domain optimization model is discretized within the prediction time domain length to obtain a discretized optimization model; Obtain the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment, and use the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment as the initial state; The initial state is input into the discretized optimization model, and a preset solver is called to solve the problem, thereby obtaining the optimal action sequence of each robot in the prediction time domain; The actions between the current moment and the next scheduling moment are extracted from the optimal action sequence to generate the collaborative scheduling strategy. The collaborative scheduling strategy includes at least one of the following for each robot: return to port instruction, queuing instruction, recharging instruction, equipment change instruction, maintenance instruction, and redeployment instruction.

[0080] Specifically, starting from the current decision-making moment, a prediction time domain length is set. This length can be dynamically configured according to the operational scenario of the hub, for example, typically set to 10 time steps, each time step being 30 seconds, to predict the system evolution within the next 5 minutes. The rolling time domain optimization model is discretized within this prediction time domain, that is, the continuous time axis is divided into time steps of equal length, and the decision variables are treated as constants at each time step, thus obtaining a discretized optimization model. This model is mathematically expressed as a mixed integer programming form, containing discrete action selection variables and continuous charging time variables. The system reads the unified state vector of the robot swarm at the current moment from the digital twin model, including the position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level of each robot, as well as the unified state vector of the infrastructure at the current moment, including the type, compatibility category, current occupancy status, estimated release time, available charging power, swappable battery specifications, available tool set, and safety isolation level of each berth. The above two sets of vectors are used as the initial state of the discretized optimization model. The system inputs the initial state into the discretized optimization model and calls a preset solver to solve it. The preset solver can be a commercial optimization solver such as Gurobi or CPLEX, an open-source solver such as OR-Tools, or a heuristic search algorithm customized for the scale of the distribution center. Under the premise of satisfying dynamic safety power constraints, berth mutual exclusion constraints, task timing constraints, capability matching constraints, and safety isolation constraints, the solver minimizes the objective function and outputs the optimal action sequence for each robot at each time step in the prediction time domain. This sequence records the actions that each robot should perform at each future time step, such as continuing to perform the current task, going to a berth with a certain number, queuing at the berth, starting charging, stopping charging and performing equipment change, entering a maintenance isolation berth, or being assigned a new task. The system extracts actions from the optimal action sequence obtained from the solution, between the current moment and the next scheduling moment. The next scheduling moment can be a fixed interval, such as 30 seconds later, or the most recent event trigger moment, such as a low battery alarm moment. The extracted actions constitute a collaborative scheduling strategy, which includes return-to-port instructions for each robot, such as U3 immediately returning to berth 1; queuing instructions, such as quadruped robot R2 waiting in the queue at berth 2; recharging instructions, such as wheeled robot A3 connecting to fast charging pile 3 for 15 minutes; changing instructions, such as humanoid robot H1 changing its gripper at tool station 4; maintenance instructions, such as U5 with a health index below 0.6 entering isolation berth 5 for self-check; and re-dispatch instructions, such as mobile operation robot M7 going to area B to perform a disposal task after completing charging. The system sends these instructions to the corresponding robot and berth controllers via the data bus, driving the physical equipment to execute them.

[0081] In this optional embodiment, the originally complex continuous-time dynamic optimization problem is transformed into a finite-dimensional mixed-integer programming problem, enabling the solver to complete the calculation in a very short time, thus meeting the timeliness requirements of real-time scheduling in the distribution center. Using the unified state vector of the robot swarm and infrastructure at the current moment as the initial state ensures that the optimization decision is based on an accurate mirror of the real physical state, avoiding policy deviations caused by state lag. By executing only the actions in the optimal action sequence between the current moment and the next scheduling moment, rather than all instructions in the entire prediction time domain, a rolling execution mechanism in model predictive control is realized, enabling the system to re-optimize using the latest state after each decision window ends. This allows for rapid response to dynamic events such as low battery alarms, emergency task insertions, and berth malfunctions, avoiding policy rigidity caused by one-time long-cycle planning. The final generated collaborative scheduling strategy covers multiple types of instructions, including returning to port, queuing, recharging, reloading, maintenance, and redeployment, solving the shortcomings of traditional methods in that the scheduling instructions are single and cannot be closed-loop maintained. This significantly improves the execution reliability and task continuity of multi-type robot swarms in unattended distribution centers.

[0082] Optionally, the simulation verification of the cooperative scheduling strategy in the digital twin model includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, energy shortages, berth conflicts, or safety risks to obtain a verified cooperative scheduling strategy, including: The collaborative scheduling strategy is loaded into the digital twin model, and the simulation start time and simulation time step are set. According to the simulation time step, the return-to-port instruction, the queuing instruction, the recharging instruction, the equipment replacement instruction, the maintenance instruction, and the redeployment instruction in the cooperative scheduling strategy are simulated and executed frame by frame in the digital twin model; During the simulation, it is detected whether a preset abnormal event occurs. If the preset abnormal event is not detected within the simulation time step, it is determined that the cooperative scheduling strategy has been verified and the verified cooperative scheduling strategy is obtained. If the preset abnormal event is detected, the cooperative scheduling strategy verification is marked as unsuccessful, and the process returns to the step of establishing the rolling time-domain optimization model to regenerate the cooperative scheduling strategy.

[0083] Specifically, the collaborative scheduling strategy is loaded into the digital twin model. The simulation start time is set to the current real-time, and the simulation time step is set to, for example, 0.5 seconds or 1 second. The total simulation duration is the same as or slightly longer than the predicted time domain length to ensure complete coverage of the strategy execution cycle. The sandbox environment in the digital twin model begins to simulate and execute each instruction in the collaborative scheduling strategy frame by frame according to the set simulation time step. This includes the process of the robot returning to its designated berth from its current position along the planned path as specified by the return-to-port instruction; the behavior of the robot waiting in the berth queue and moving forward sequentially as specified by the queuing instruction; the process of the robot connecting to the charging pile and charging according to a specific power and duration as described by the refueling instruction; the action of the robot entering the tool changing station and completing the load or tool switching action as specified by the changing-up instruction; the process of the robot entering the isolated berth and performing self-checks or data uploads as triggered by the maintenance instruction; and the movement of the robot leaving the distribution center for a new task area after refueling or changing up as instructed by the reassignment instruction. During each frame of the simulation, the digital twin sandbox continuously monitors for pre-set anomalies. These pre-set anomalies include path conflicts (two or more robots' planned paths occupying the same spatial location at the same time), insufficient energy (based on the energy consumption model, a robot's remaining battery power is below the dynamic safety threshold after completing the task assigned in the strategy), berth conflicts (the berth specified in the strategy is still occupied at the target time or there is a battery swapping compartment), insufficient tool inventory leading to swapping failure, and safety risks (the robot's trajectory enters a no-fly zone, personnel safety zone, or triggers fire or thermal runaway alarm conditions). If the system detects no pre-set anomalies throughout the entire simulation time step, the collaborative scheduling strategy is deemed to have passed verification, and the verified collaborative scheduling strategy is prepared to be deployed to the physical system for execution. If any of the above-mentioned preset abnormal events are detected during the simulation, the cooperative scheduling strategy is immediately marked as unqualified, the current simulation is terminated, and the process automatically returns to the step of establishing a rolling time-domain optimization model. The weights or constraints of the optimization model are adjusted according to the type of anomaly found in the simulation verification. For example, if a path conflict is found, a space avoidance penalty term is added; if insufficient energy is found, the dynamic safe power threshold is increased; if a berth conflict is found, the berth allocation order is adjusted. Then, the rolling time-domain optimization model is solved again and a new cooperative scheduling strategy is generated. The simulation verification is then carried out again in the digital twin sandbox. This process is repeated until a strategy that has passed verification is obtained.

[0084] In this optional embodiment, by performing a full-process pre-simulation of the scheduling strategy in a virtual environment with zero physical risk, potential execution obstacles can be identified and eliminated in advance without affecting the operation of real robots. This completely avoids the real costs caused by directly issuing strategies in traditional methods, such as path congestion, robot power outages, berth conflicts, and even safety accidents. At the same time, the high efficiency of simulation verification allows it to complete the feasibility assessment of a single strategy within seconds or milliseconds. For strategies that fail verification, a replanning loop can be automatically triggered. The system will dynamically adjust the target or constraints of the optimization model according to the anomaly type, such as increasing space avoidance penalties or raising the safe power threshold and resolving. The automated iteration mechanism significantly improves the initial success rate of scheduling strategies and reduces the frequency of manual intervention due to strategy infeasibility. In addition, simulation verification covers the entire chain of instructions from returning to port, queuing, recharging, changing equipment, maintenance, and redeployment, ensuring that every step from task execution to returning to port for recharging and then to re-departure is truly feasible in terms of energy, space, resources, and safety. Ultimately, this ensures the long-term stable, safe, and efficient operation of multi-type robot clusters in unattended distribution centers.

[0085] Optionally, the step of driving multiple types of robots in the entity distribution center to perform preset operations according to the verified collaborative scheduling strategy, and, during the execution of the preset operations, performing closed-loop correction on the digital twin model and the rolling time-domain optimization model based on the status data fed back by the robots in real time, and writing the execution result back to the digital twin model, includes: The verified collaborative scheduling strategy is sent to the execution controller of the entity distribution center; The execution controller parses the cooperative scheduling strategy to obtain a sequence of control instructions for each robot; Drive the corresponding robot to perform operations according to the sequence of control instructions; During the operation, the robot's real-time status data is collected at a fixed frequency. The status data includes the actual position, actual remaining battery power, actual health index, command execution status, and fault codes. Based on the collected state data, update the state vector of the corresponding robot and the state vector of the corresponding berth in the digital twin model; The execution status and fault code of the instruction are written as the execution result into the execution record of the digital twin model; Calculate the deviation between the state data and the predicted state in the collaborative scheduling strategy, and adjust the energy consumption model parameters, berth service time model parameters, or health assessment model parameters in the rolling time-domain optimization model according to the deviation.

[0086] Specifically, the collaborative scheduling strategy, validated through digital twin sandbox simulation, is distributed to the execution controller at the physical distribution center. This execution controller, deployed on an edge server or programmable logic controller at the distribution center, is responsible for low-level communication with each robot, berth controller, charging station, and tool bay. Upon receiving the collaborative scheduling strategy, the execution controller parses it, breaking down the return-to-port, queuing, recharging, reloading, maintenance, and redeployment instructions for each robot into a sequence of control instructions recognizable by the robot factory. For example, returning U3 to berth number 1 is parsed as the navigation target point, speed limit, and docking alignment parameters. Following the parsed control instruction sequence, the execution controller sends drive commands to the corresponding robots one by one via wireless communication links such as WiFi, 4G, 5G, or proprietary radio protocols. The robots execute corresponding operations based on the commands, while the berth controller, charging station, and tool bay also receive instructions and cooperate to complete actions such as docking guidance, charging connection, and reloading / grabbing.

[0087] During operation, real-time status data transmitted back by the robot is collected via a data bus at a fixed frequency, such as 1 Hz or 5 Hz. This status data includes the robot's actual position coordinates obtained by fusing GPS, IMU, and odometry; the actual remaining battery power reported by the battery management system; the actual health index calculated from actual sensor indicators required by the health assessment model, such as battery internal resistance, joint current, and vibration amplitude; the execution status of each command, such as success, failure, in progress, or timeout; and fault codes generated by the fault self-diagnosis module. Based on the collected real-time status data, the corresponding robot's state vector in the digital twin model is updated, including position, speed, remaining battery power, health index, task progress, payload type, sensor availability, communication quality, fault codes, and safety level. Simultaneously, the corresponding berth's state vector is updated, including occupancy status, estimated release time, and changes in available charging power, ensuring strict consistency between the digital twin model and the physical entity. The execution status and fault codes of each instruction are recorded as execution results and written to the execution record database of the digital twin model. This record includes a timestamp, robot identifier, instruction source version number, actual execution time, whether an anomaly occurred, and related data fragments before and after the anomaly, providing the original basis for subsequent auditing, review, and model training. The system calculates the deviation between the real-time collected status data and the status predicted during optimization by the energy consumption model, berth service time model, and health assessment model in the collaborative scheduling strategy. Examples include the difference between predicted and actual remaining power, the difference between predicted docking service time and actual occupancy time, and the difference between predicted health index decay and actual health index change. Based on these deviations, the system automatically adjusts the model parameters in the rolling time-domain optimization model: if actual energy consumption is consistently higher than predicted energy consumption, the motion power consumption coefficient or task power consumption coefficient in the energy consumption model is increased; if actual berth service time is generally longer than predicted time, the docking probability function coefficient in the berth service time model is refitted using historical data; if the measured decay rate of the health index is inconsistent with the model, the reference values ​​or weight coefficients of each observation indicator in the health assessment model are adjusted. The adjusted model parameters are saved and used for subsequent rolling time-domain optimization, forming a continuous closed-loop correction.

[0088] In this optional embodiment, the collaborative scheduling strategy is sent to the execution controller at the physical distribution center. After parsing, the controller drives the robot to perform operations. During execution, the system collects real-time status data transmitted back by the robot at a fixed frequency, updating the corresponding robot and berth state vectors in the digital twin model accordingly. Simultaneously, the execution results are written to the execution record. Furthermore, by calculating the deviation between the actual and predicted states, the energy consumption model parameters, berth service time model parameters, or health assessment model parameters in the rolling time-domain optimization model are dynamically adjusted. This constructs a complete closed loop from virtual decision-making to physical execution and data feedback, ensuring that the digital twin model is no longer a static image but a real-time updated model capable of continuous self-calibration based on actual execution data. Real-time status acquisition and twin updates ensure that subsequent scheduling decisions are always based on the latest physical reality, avoiding strategy deviations caused by model drift. Writing execution results to records provides high-value samples for audit backtracking, fault analysis, and model training. The deviation-driven parameter adjustment mechanism enables energy consumption prediction, service time estimation, and health assessment to adaptively adapt to robot aging, environmental changes, and task mode evolution. For example, when the actual energy consumption of a robot is higher due to tire wear, the energy consumption model coefficient is automatically adjusted upward, thereby reserving more power margin in future scheduling. This enables the hub to have adaptive optimization capabilities under long-term operation, significantly improving the scheduling accuracy, task success rate, and resource utilization efficiency of multi-type robot clusters.

[0089] Combination Figure 2 As shown, this embodiment of the invention provides an intelligent scheduling and operation and maintenance system for multi-category robot distribution centers, including: The data acquisition unit is used to acquire status data of multiple types of robots in the physical distribution center. The status data includes: robot status, task status, berth status and energy status. A state definition unit is used to define a unified state vector for the robot swarm and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state. The twin model construction unit is used to construct a digital twin model of the physical distribution center based on the unified state vector of the robot swarm and the unified state vector of the infrastructure. The optimization scheduling unit is used to establish a rolling time-domain optimization model based on the remaining power and health index extracted from the robot status, the task value and operational risk extracted from the task status, the berth queue status extracted from the berth status, and the energy cost extracted from the energy status, and to generate a collaborative scheduling strategy by solving the rolling time-domain optimization model. The simulation verification unit is used to perform simulation verification of the cooperative scheduling strategy in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking whether there are path conflicts, insufficient energy, berth conflicts or safety risks, and obtaining a cooperative scheduling strategy that has passed the verification. The execution and correction unit is used to drive multiple types of robots in the entity distribution center to perform preset operations according to the verified collaborative scheduling strategy, and to perform closed-loop correction of the digital twin model and the rolling time-domain optimization model based on the status data fed back by the robots in real time during the execution of the preset operations, and to write the execution results back to the digital twin model.

[0090] The intelligent scheduling and operation and maintenance system for multi-category robot distribution centers of the present invention has the same advantages over the prior art as the aforementioned intelligent scheduling and operation and maintenance method for multi-category robot distribution centers, and will not be repeated here.

[0091] Combination Figure 3 As shown, an embodiment of the present invention provides an electronic device, including: a processor and a memory, wherein the memory is used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the above-described intelligent scheduling and operation and maintenance method for multi-category robot distribution centers.

[0092] The electronic device of the present invention has the same advantages over the prior art as the aforementioned intelligent scheduling and operation and maintenance method for multi-category robot distribution centers, and will not be repeated here.

[0093] While the present invention has been disclosed above, its scope of protection is not limited thereto. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the present invention, and all such changes and modifications will fall within the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling and operation and maintenance of multi-category robot distribution centers, characterized in that, include: Acquire status data of multiple types of robots in physical distribution centers, including robot status, task status, berth status, and energy status. Based on the robot state, the task state, the berth state, and the energy state, a unified state vector for the robot swarm composed of the multi-category robots and a unified state vector for the infrastructure are defined. Based on the unified state vector of the robot swarm and the unified state vector of the infrastructure, a digital twin model of the physical distribution center is constructed; specifically, this includes: establishing an energy consumption model of the physical distribution center based on the remaining power in the unified state vector of the robot swarm. A health assessment model for the physical distribution center is established based on the health index in the unified state vector of the robot swarm. Based on the berth type, available charging power, and swappable battery specifications in the unified state vector of the infrastructure, a berth service time model for the physical distribution center is established. Based on the task semantic information in the task status, establish a task semantic graph model of the entity distribution center; The energy consumption model, the health assessment model, the berth service time model, and the task semantic graph model are fused together to obtain a digital twin model of the physical distribution center; Based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue state extracted from the berth state, and the energy cost extracted from the energy state, a rolling time-domain optimization model is established, and a collaborative scheduling strategy is generated by solving the rolling time-domain optimization model. The cooperative scheduling strategy is simulated and verified in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, insufficient energy, berth conflicts or safety risks, to obtain a cooperative scheduling strategy that has passed the verification. The verified collaborative scheduling strategy drives multiple types of robots in the physical distribution center to perform preset operations. During the execution of the preset operations, the digital twin model and the rolling time-domain optimization model are corrected in a closed loop based on the status data fed back by the robots in real time, and the execution results are written back to the digital twin model. Specifically, this includes: sending the verified collaborative scheduling strategy to the execution controller of the physical distribution center. The execution controller parses the cooperative scheduling strategy to obtain a sequence of control instructions for each robot; Drive the corresponding robot to perform operations according to the sequence of control instructions; During the operation, the robot's real-time status data is collected at a fixed frequency. The status data includes the actual position, actual remaining battery power, actual health index, command execution status, and fault codes. Based on the collected state data, update the state vector of the corresponding robot and the state vector of the corresponding berth in the digital twin model; The execution status and fault code of the instruction are written as the execution result into the execution record of the digital twin model; Calculate the deviation between the state data and the predicted state in the collaborative scheduling strategy, and adjust the energy consumption model parameters, berth service time model parameters, or health assessment model parameters in the rolling time-domain optimization model according to the deviation.

2. The intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to claim 1, characterized in that, The acquisition of status data for multiple types of robots in the physical distribution center includes: robot status, task status, berth status, and energy status. The robot status is obtained, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The task status is obtained, which includes task identifier, task type, task priority, task value, expected completion time, associated robot, task progress, and operational risk. The berth status is obtained, which includes berth identifier, berth type, berth queue status, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, and available tool set; The energy status is obtained, which includes energy type, current energy price, energy storage capacity, charging equipment status, and energy cost.

3. The intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to claim 1, characterized in that, The step of defining a unified state vector for the robot swarm and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state includes: The parameter order of the preset robot swarm state vector is defined, including position, speed, remaining battery power, health index, current task progress, payload type, sensor availability, communication quality, fault code, and safety level. The parameter values ​​corresponding to each robot are obtained from the robot states, and the parameter values ​​corresponding to each robot are assigned according to the preset robot swarm state vector to obtain the state vector of that robot. The state vectors of all robots constitute the unified state vector of the robot swarm. The parameters of the preset infrastructure state vector are ordered, including berth type, compatible robot category, current occupancy status, estimated release time, available charging power, replaceable battery specifications, available tool set, and security isolation level. The parameter values ​​corresponding to each berth are obtained from the berth status, and the parameter values ​​corresponding to each berth are assigned according to the preset infrastructure state vector to obtain the state vector of that berth. The state vectors of all berths constitute the unified state vector of the infrastructure.

4. The intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to claim 1, characterized in that, The step of establishing a rolling time-domain optimization model based on the remaining power and health index extracted from the robot's state, the task value and operational risk extracted from the task state, the berth queue status extracted from the berth state, and the energy cost extracted from the energy state includes: Based on the stated task value, construct task benefit items; Based on the remaining electricity, the energy cost, and the predicted energy consumption of the return route, an energy cost item is constructed. Based on the aforementioned health index, construct maintenance risk items; Based on the described operational risks, construct safety risk items; Based on the berth queue status, construct a queuing cost item; The objective function is obtained by weighted summing of the task benefit item, the energy cost item, the maintenance risk item, the security risk item, and the queuing cost item; Based on the objective function and preset constraints, the rolling time-domain optimization model is established.

5. The intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to claim 4, characterized in that, The step of generating a cooperative scheduling strategy by solving the rolling time-domain optimization model includes: Starting from the current time, the prediction time domain length is set, and the rolling time domain optimization model is discretized within the prediction time domain length to obtain a discretized optimization model; Obtain the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment, and use the unified state vector of the robot swarm at the current moment and the unified state vector of the infrastructure at the current moment as the initial state; The initial state is input into the discretized optimization model, and a preset solver is called to solve the problem, thereby obtaining the optimal action sequence of each robot in the prediction time domain; The actions between the current moment and the next scheduling moment are extracted from the optimal action sequence to generate the collaborative scheduling strategy. The collaborative scheduling strategy includes at least one of the following for each robot: return to port instruction, queuing instruction, energy replenishment instruction, equipment change instruction, maintenance instruction, and redeployment instruction.

6. The intelligent scheduling and operation and maintenance method for multi-category robot distribution centers according to claim 5, characterized in that, The simulation verification of the cooperative scheduling strategy in the digital twin model includes simulating the execution of the cooperative scheduling strategy and checking for path conflicts, energy shortages, berth conflicts, or safety risks, to obtain a verified cooperative scheduling strategy, including: The collaborative scheduling strategy is loaded into the digital twin model, and the simulation start time and simulation time step are set. According to the simulation time step, the return-to-port instruction, the queuing instruction, the recharging instruction, the equipment replacement instruction, the maintenance instruction, and the redeployment instruction in the cooperative scheduling strategy are simulated and executed frame by frame in the digital twin model; During the simulation, it is detected whether a preset abnormal event occurs. If the preset abnormal event is not detected within the simulation time step, it is determined that the cooperative scheduling strategy has been verified and the verified cooperative scheduling strategy is obtained. If the preset abnormal event is detected, the cooperative scheduling strategy verification is marked as unsuccessful, and the process returns to the step of establishing the rolling time-domain optimization model to regenerate the cooperative scheduling strategy.

7. A multi-category robot distribution center intelligent scheduling and operation and maintenance system, characterized in that, include: The data acquisition unit is used to acquire status data of multiple types of robots in the physical distribution center. The status data includes: robot status, task status, berth status and energy status. A state definition unit is used to define a unified state vector for the robot swarm composed of the multi-category robots and a unified state vector for the infrastructure based on the robot state, the task state, the berth state, and the energy state. A digital twin model construction unit is used to construct a digital twin model of the physical distribution center based on the unified state vector of the robot swarm and the unified state vector of the infrastructure. Specifically, this includes: establishing an energy consumption model of the physical distribution center based on the remaining battery power in the unified state vector of the robot swarm; establishing a health assessment model of the physical distribution center based on the health index in the unified state vector of the robot swarm; establishing a berth service time model of the physical distribution center based on the berth type, available charging power, and swappable battery specifications in the unified state vector of the infrastructure; establishing a task semantic graph model of the physical distribution center based on the task semantic information in the task state; and fusing the energy consumption model, the health assessment model, the berth service time model, and the task semantic graph model to obtain a digital twin model of the physical distribution center. The optimization scheduling unit is used to establish a rolling time-domain optimization model based on the remaining power and health index extracted from the robot status, the task value and operational risk extracted from the task status, the berth queue status extracted from the berth status, and the energy cost extracted from the energy status, and to generate a collaborative scheduling strategy by solving the rolling time-domain optimization model. The simulation verification unit is used to perform simulation verification of the cooperative scheduling strategy in the digital twin model. The simulation verification includes simulating the execution of the cooperative scheduling strategy and checking whether there are path conflicts, insufficient energy, berth conflicts or safety risks, and obtaining a cooperative scheduling strategy that has passed the verification. The execution and correction unit is used to drive multiple types of robots in the physical distribution center to perform preset operations according to the verified collaborative scheduling strategy. During the execution of the preset operations, it performs closed-loop correction of the digital twin model and the rolling time-domain optimization model based on the real-time status data fed back by the robots, and writes the execution results back to the digital twin model. Specifically, it includes: sending the verified collaborative scheduling strategy to the execution controller of the physical distribution center; the execution controller parsing the collaborative scheduling strategy to obtain a control instruction sequence for each robot; and driving the corresponding robot to perform operations according to the control instruction sequence. During operation, real-time status data transmitted back by the robot is collected at a fixed frequency. This status data includes actual position, actual remaining battery power, actual health index, command execution status, and fault codes. Based on the collected status data, the state vectors of the corresponding robot and the corresponding berth in the digital twin model are updated. The command execution status and fault codes are used as execution results and written into the execution record of the digital twin model. The deviation between the status data and the predicted status in the collaborative scheduling strategy is calculated, and the energy consumption model parameters, berth service time model parameters, or health assessment model parameters in the rolling time-domain optimization model are adjusted based on the deviation.

8. An electronic device, characterized in that, include: Processor and memory, the memory being used to store computer programs; When the computer program is loaded by the processor, it causes the processor to execute the intelligent scheduling and operation and maintenance method for multi-category robot distribution centers as described in any one of claims 1-6.

Images