Cluster task allocation and cooperative control method for multiple sorting robots
By constructing a digital twin model and a hierarchical reinforcement learning network, combined with a market auction mechanism and distributed model predictive control, the problem of task allocation and collaborative control of multiple sorting robots in a large warehouse environment was solved, achieving efficient and stable sorting operations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- SHENZHEN XIAONAN INTELLIGENT MFG CO LTD
- Filing Date
- 2026-03-11
- Publication Date
- 2026-06-05
Smart Images
Figure CN122151669A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of cluster task allocation technology, specifically involving a method for cluster task allocation and collaborative control of multiple sorting robots. Background Technology
[0002] With the rapid development of e-commerce, new retail, and smart logistics, the demand for automation and intelligent upgrading of warehousing and sorting processes is becoming increasingly urgent. Currently, existing technologies have developed methods for task allocation and collaborative control of multiple sorting robots. However, these methods are mostly based on traditional centralized scheduling or simple distributed planning strategies. When facing the dynamic and complex operating environment of large warehouses, they still face many technical bottlenecks and cannot meet the requirements of high-density, high-efficiency, and high-precision sorting operations. The specific problems are as follows: Existing task allocation schemes mostly adopt a static pre-allocation mode, which does not fully integrate the robot's real-time status information and heterogeneous capability parameters, and does not achieve hierarchical collaboration between global scheduling and local execution. This easily leads to uneven distribution of task flow and unbalanced robot load within the warehouse area. At the same time, the lack of an effective task arbitration mechanism makes it impossible to match tasks based on their urgency and robot execution efficiency in multiple dimensions, making it difficult to balance the global optimization of cluster operations with the operational efficiency of individual robots, resulting in low overall sorting task completion efficiency.
[0003] The multi-constraint fusion modeling of sorting robots is incomplete, failing to fully integrate key information such as kinematic constraints, energy consumption models, communication latency, and spatial pose. Furthermore, the pose data fusion estimation method collected by multi-source sensors has limited accuracy and cannot accurately reflect the actual motion state of the robot. In addition, the modeling and updating of dynamic obstacles and task target points in the warehouse environment are lagging behind, and there is a lack of efficient environmental characterization methods, resulting in insufficient digital support for subsequent path planning and collaborative control.
[0004] Therefore, there is an urgent need for a cluster task allocation and collaborative control method for multiple sorting robots to solve the above problems. Summary of the Invention
[0005] The purpose of this invention is to provide a cluster task allocation and collaborative control method for multiple sorting robots, which solves the technical problems of insufficient accuracy of digital representation of robots and environment, and poor real-time performance and reliability of path planning and collaborative obstacle avoidance in the prior art.
[0006] To achieve the above objectives, the present invention adopts the following technical solution: A method for cluster task allocation and collaborative control of multiple sorting robots, including: Step 1: Construct a digital twin model of the heterogeneous robot cluster, and collect the status information and capability parameters of each robot in real time through multi-source sensors; Step 2: Establish a digital representation that includes kinematic constraints, load capacity, energy consumption model, communication latency, and spatial pose, and synchronously model obstacles and mission objectives in the dynamic environment. Step 3: Based on digital representation, a hierarchical reinforcement learning network is used to perform coarse global task allocation and fine local task adjustment. The sorting tasks that arrive in real time are decomposed into sub-tasks that can be executed in parallel. Through a market auction mechanism combined with the dynamic priority and load balancing of robots, multiple rounds of bidding arbitration are carried out to dynamically match the real-time capabilities and positions of heterogeneous robots and generate an initial task allocation scheme that takes into account both global optimization and individual benefits. Step 4: Based on the initial task allocation scheme, a distributed model predictive control method is used to generate local paths for each robot; Step 5: Adjust the path in real time through the cooperative obstacle avoidance protocol and deadlock prediction mechanism among robots. At the same time, perform rolling optimization and error compensation based on the real-time deviation between the physical entity and the virtual model in the digital twin environment. Use twin data to drive the model predictive controller to correct subsequent instructions and output a high-precision, adaptive final task execution sequence and cooperative control instructions.
[0007] Furthermore, the status information and capability parameters specifically include: The robot's real-time position coordinates, running speed and acceleration, current orientation angle, remaining battery power, estimated remaining runtime, maximum load capacity, current load weight, signal strength and transmission delay of the communication module, real-time point cloud data of the surrounding environment collected by the vehicle-mounted sensors, and fault diagnosis status information of the robot itself.
[0008] Furthermore, a digital representation is established, specifically through the following methods: Based on the collected state information and capability parameter data, the extended Kalman filter algorithm is used to perform fusion estimation of the robot pose, and a three-dimensional digital model including kinematic differential constraints, nonholonomic constraints, and energy consumption curve fitting function is constructed. The occupied grid map is used to synchronously label and update dynamic obstacles and task target points in real time.
[0009] Furthermore, the hierarchical reinforcement learning network described in step three includes a global scheduling layer and a local execution layer, specifically as follows: The global scheduling layer adopts a deep Q-network algorithm to dynamically divide the warehouse area into several sub-regions and coarsely allocate tasks to each sub-region based on the global task flow density and the spatial distribution characteristics of the robot cluster. The local execution layer adopts a multi-agent near-end strategy optimization algorithm to perform secondary fine-tuning and allocation of tasks in each sub-region based on the real-time heterogeneous capabilities of the robots.
[0010] Furthermore, the specific method of the market auction mechanism described in step three is as follows: The global scheduling layer acts as the auctioneer, releasing task package information. Each robot, as a bidder, calculates its bid price and expected completion time based on its own status and submits its bid. The auctioneer adopts a combined auction strategy, comprehensively considering the urgency of the task, the robot load balancing, and the global completion time window. After multiple rounds of bidding and arbitration, the task is allocated to the optimal robot combination, and a binding digital task contract is signed.
[0011] Furthermore, the distributed model predictive control method described in step four is specifically as follows: Each robot constructs an online rolling optimization problem with a fixed prediction time domain as the window. The objective function is to minimize the sum of path tracking error, its own energy consumption, and collision risk functions with other robots. The constraints include kinematic model, velocity and acceleration boundaries, safe distance maintenance constraints, and task deadline constraints. After solving the problem, the robot executes the first control cycle instruction and updates the system based on the new state in the next cycle.
[0012] Furthermore, the cooperative obstacle avoidance protocol described in step five is specifically as follows: Robots communicate with each other via point-to-point, broadcasting their positions, velocity vectors, and predicted trajectories at a fixed frequency. When a robot detects that its relative distance to another robot is less than a preset safety threshold, it dynamically generates a priority ranking based on task priority, remaining battery power, and load status. Robots with lower priority take proactive measures such as deceleration, waiting, and local path replanning. At the same time, a virtual repulsive force is generated using the artificial potential field method to ensure that robots can pass through dense areas without collisions.
[0013] Furthermore, the deadlock prediction mechanism described in step five is specifically as follows: Based on the robot's predicted trajectory and the warehouse road topology, a directed cycle detection algorithm in graph theory is used to analyze the resource occupation relationship between multiple robots in real time. When a deadlock state of circular waiting is predicted, the central coordinator dynamically adjusts the task execution order of some robots and guides them into a temporary waiting area until the resource occupation cycle is resolved.
[0014] Furthermore, the rolling optimization and error compensation in the digital twin environment described in step five specifically includes: In a digital twin model, the theoretical execution states of all robots are simulated in parallel. The actual poses of the physical robots are compared with the theoretical poses in the twin model in real time, and positional and temporal deviations are calculated. When the deviation exceeds a preset error tolerance threshold, a local replanning mechanism is triggered, and the deviation value is fed back as a penalty term to the objective function of the distributed model predictive controller. The state parameters of the digital twin model are updated synchronously, realizing closed-loop dynamic correction between the physical and virtual systems. In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: 1. This invention achieves dynamic allocation of sorting tasks by integrating a hierarchical reinforcement learning network with a market auction mechanism. The global scheduling layer completes the coarse allocation of tasks in the warehouse area with the help of a deep Q-network, and the local execution layer achieves fine adjustment of tasks in sub-areas through a multi-agent proximal strategy optimization algorithm. Combined with multi-round bidding arbitration to dynamically match the real-time capabilities and positions of heterogeneous robots, it not only effectively reduces the solution complexity of task allocation for large-scale clusters, but also achieves robot load balancing, taking into account both the global optimization of cluster operations and the operational efficiency of individual robots, and greatly improves the overall execution efficiency of warehouse sorting tasks. 2. This invention employs a distributed model predictive control to plan the robot's local path and designs a cooperative obstacle avoidance protocol and deadlock prediction mechanism. The distributed architecture avoids the computational bottleneck of centralized path planning, ensuring system real-time performance. Based on dynamic passage rules such as task priority and remaining battery power, combined with the artificial potential field method, it achieves active collision-free obstacle avoidance in densely populated robot areas. By using a graph theory directed cycle detection algorithm to predict deadlock risks in advance and guide robots into temporary waiting areas to release resource occupation cycles, it fundamentally avoids robot collisions and work interruptions, significantly improving the stability and continuity of heterogeneous robot cluster operation. 3. This invention constructs a digital twin model of a heterogeneous robot cluster and realizes closed-loop dynamic correction between the physical and virtual systems. By parallel deducing the theoretical execution state of the robot in the digital twin environment, the pose and time deviation between the physical robot and the virtual model are detected and calculated in real time. When the deviation exceeds the threshold, local replanning is triggered and the deviation is fed back to the model predictive controller as a penalty term, thereby realizing adaptive correction of control commands and improving the system's adaptability and robustness to dynamic working environments. Attached Figure Description
[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0016] Figure 1 The diagram illustrates the steps of the cluster task allocation and collaborative control method for multiple sorting robots according to the present invention. Figure 2 A flowchart of the method for collaborative control of multiple sorting robots according to the present invention is shown; Figure 3 The flowchart of the collaborative task execution based on real-time deviation correction of the present invention is shown. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] like Figure 1 , Figure 2 The method for cluster task allocation and collaborative control of multiple sorting robots shown includes the following steps: Step 1: Construct a digital twin model of the heterogeneous robot cluster, and collect the status information and capability parameters of each robot in real time through multi-source sensors; This implementation constructs a three-tier system architecture: physical entity layer, digital twin layer, and control decision layer, applicable to large-scale warehouse sorting scenarios. Specifically: The physical entity layer includes N heterogeneous sorting robots (N≥50), each robot is equipped with LiDAR, industrial camera, UWB positioning module, battery management system and fault diagnosis sensor; the warehouse environment includes sorting table, aisle, temporary waiting area, charging area and other facilities, and all equipment is connected to the network through 5G communication module.
[0019] The digital twin layer is deployed on edge computing nodes and is responsible for receiving real-time data from the physical entity layer, building and maintaining a virtual model synchronized with the physical world, and realizing state mapping, parallel inference and deviation analysis.
[0020] The control decision layer integrates a hierarchical reinforcement learning network, a distributed model predictive controller, and a cooperative scheduling algorithm, and is responsible for task allocation, path planning, and the generation and issuance of cooperative control commands.
[0021] In this embodiment, each sorting robot is equipped with the following sensors and collects data at a period of 100ms, which is then transmitted to the digital twin layer via 5G slice communication: Pose data: Through the fusion of UWB localization and LiDAR SLAM, the robot's three-dimensional position coordinates (X, Y, Z) in the global coordinate system, as well as its running speed (v), acceleration (a), and current orientation angle are collected. It is used to describe the real-time motion state of a robot.
[0022] State of Energy (SOC) data is collected through the battery management system, and the estimated remaining driving range is calculated based on the current power consumption and remaining charge. It is used for power constraints and charging scheduling in subsequent task allocation.
[0023] Load capacity data: Current load weight is collected via a pressure sensor. and read the maximum load capacity from the robot controller. It is used to determine whether a robot is capable of handling sorting tasks with a specific weight.
[0024] Communication quality data: Real-time monitoring of signal strength RSSI and transmission delay via the communication module. It is used to assess the robot's network connectivity status and prioritize the allocation of non-urgent tasks in areas with weak signals.
[0025] Environmental perception data: Point cloud data (PCD format) of the surrounding environment is collected by LiDAR, and image data (JPG format) of the environment is collected by industrial cameras for obstacle detection and dynamic environment modeling.
[0026] Health status data: The fault diagnosis unit collects operating parameters such as motor temperature and drive voltage, and outputs fault codes (0 indicates normal, 1~10 indicate different fault types) to detect and handle abnormal robots in a timely manner.
[0027] Based on the collected data, a geometric model is generated for each robot and imported into the digital twin platform. A mapping relationship between sensor data and the virtual model is established. After denoising and interpolation preprocessing, the real-time collected status information and capability parameters are assigned to the virtual model, achieving bidirectional synchronization between the physical entity and the virtual model.
[0028] Step 2: Establish a digital representation that includes kinematic constraints, load capacity, energy consumption model, communication latency, and spatial pose, and synchronously model obstacles and mission objectives in the dynamic environment. Building upon the digital twin model, it is necessary to further establish digital representations of the robot and the environment, transforming the collected data into mathematical expressions that can be used for optimization and control.
[0029] The extended Kalman filter (EKF) is used to fuse multi-source pose data. The state equation and observation equation are as follows: Equations of state: ; Observation equation: Wherein, the state vector Includes three-dimensional position Orientation Angle Velocity v and acceleration a; control vector The observation vector Z represents the sensor's measured values; A, B, and H are the state transition matrix, control input matrix, and observation matrix, respectively; process noise. and observation noise They follow a zero-mean Gaussian distribution, and their covariance matrices are respectively set as follows. , .
[0030] Robot kinematic constraints are the foundation of path planning: differential constraints , , ( (angular velocity); nonholonomic constraints This reflects the robot's inability to move laterally.
[0031] Energy consumption models are crucial for cost assessment in task allocation and path planning. The model is obtained using the least squares method for fitting: , where P is the real-time power (W). , , These are the fitting coefficients (calibrated using actual robot measurement data, corresponding to the speed influence factor, load influence factor, and basic no-load power, respectively).
[0032] An occupancy grid map was used to label dynamic obstacles and task target points. The grid resolution was 0.1m×0.1m, and the grid status was divided into "idle (0)", "occupied (1)" and "unknown (2)". Based on the LiDAR point cloud and visual semantic segmentation results, the grid status was updated every 50ms to provide the robot with real-time environmental perception information.
[0033] For each sorting task, mark the following attribute: weight Accuracy requirements Priority (Levels 1-5, with Level 1 being the highest), Deadline These attributes will be used for bid calculation and priority sorting in task allocation.
[0034] Step 3: Based on digital representation, a hierarchical reinforcement learning network is used to perform coarse global task allocation and fine local task adjustment. The sorting tasks that arrive in real time are decomposed into sub-tasks that can be executed in parallel. Through a market auction mechanism combined with the dynamic priority and load balancing of robots, multiple rounds of bidding arbitration are carried out to dynamically match the real-time capabilities and positions of heterogeneous robots and generate an initial task allocation scheme that takes into account both global optimization and individual benefits. After completing the digital representation, a solution combining a hierarchical reinforcement learning network and a market auction mechanism is adopted to generate an initial task allocation scheme for the task allocation problem of large-scale heterogeneous robot clusters. Hierarchical reinforcement learning network parameter configuration Global scheduling layer (DQN network): Input parameters are task flow density (single / m²), robot cluster spatial distribution variance, and sub-region idle computing power (FLOPS); hidden layers are two fully connected layers with 128 and 64 neurons respectively, using ReLU activation function; output layer is the sub-region task allocation ratio, using Softmax activation function. Training parameters: learning rate. =0.001, discount factor =0.95, experience replay pool capacity 1e6, target network update frequency 100 steps. The global scheduling layer dynamically divides the warehouse area into several sub-regions based on the macro-operation status, completes coarse task allocation, and effectively reduces the problem-solving scale.
[0035] Local Execution Layer (MAPPO Network): The number of agents is the same as the number of robots in the sub-region; the input parameters are the heterogeneous capability vectors of the robots (including...). SOC, current position), task attribute vector ( The sub-region environment state vector includes obstacle density and free grid ratio; the policy network consists of one 3×3 convolutional layer and two fully connected layers, and the value network has the same structure as the policy network. Training parameters: learning rate. =0.0003, Advantage estimation discount factor =0.9, batch size 64. The local execution layer performs secondary fine-tuning and allocation of tasks within each sub-region based on the robot's real-time heterogeneous capabilities, ensuring the operational efficiency of individual robots.
[0036] Market auction mechanism execution process: A market auction mechanism is introduced for fine-tuning. The global scheduling layer acts as the auctioneer, releasing a batch of task packages every 500ms. Each task package includes the number of sub-tasks and the details of each task. Coordinates of the target point.
[0037] The robot, acting as the bidder, calculates its bid price based on its own condition: ; in, This indicates the estimated completion time, determined based on the distance between the current location and the target point, as well as the average speed. This represents the estimated energy consumption of the path. The total energy consumption of the entire task path is determined based on the real-time power P. The total energy consumption of the entire task path is then integrated over the entire path from the current position to the task target point. , This represents the current load factor, which is equal to the ratio of the current load weight to the maximum weight the robot is designed to carry. The weights for completion time, energy consumption, and load rate are respectively. The multi-round bidding and arbitration process is as follows: In the first round, the robot submits a bid containing Price and The auctioneer selects bids that meet the requirements. Valid bids; Round 2, according to Priority sorting, matching high-priority tasks first; Round 3, based on load balancing. arbitration( (The average load rate of all robots in the sub-area) Select the robot combination with the smallest Price and the load balance Bal meets the preset requirements (e.g., Bal>0.8).
[0038] Upon successful bidding, the system signs a binding digital task contract. This contract is a structured, executable data file encapsulated in JSON format, uniquely bound to the robot ID and task ID. It includes basic information, performance constraints, breach of contract determination, and penalty rules. The signing of the digital task contract is based on the market auction mechanism of this invention. The specific process is as follows: After the control decision-making level completes the rough allocation and fine adjustment of tasks, it issues the digital task contract data file to the winning robot. The robot (physical entity layer) verifies whether its own status (battery, load, communication) meets the performance constraints, and returns a digital signature confirmation after the verification is successful. Once the control decision-making level receives the signature, it marks the contract status as signed, and the signing process is complete.
[0039] The digital twin layer collects robot performance data in real time. When the judgment conditions for any type of breach are met, the contract status is immediately marked as breach, and the breach type, breach time, and breach quantification value (such as accuracy deviation of 0.003m, timeout of 25s, etc.) are recorded.
[0040] The penalty for breach of contract is deeply integrated with the hierarchical reinforcement learning auction mechanism and distributed model prediction controller of this invention. The penalty measures are quantified and directly affect the task allocation and control instruction generation in the next round. If the task times out, the penalty measures are a 20% reduction in auction weight, and the next round can only accept tasks with a priority of ≤3. The energy consumption compensation coefficient is increased by 0.1, and the penalty is effective for 1 hour.
[0041] After the penalty period ends, the bidding weight and controller parameters will automatically revert to their default values. If a defaulting robot completes three or more non-defaulting tasks during the penalty period, the penalty can be lifted 50% early (verified by the digital twin layer and reported to the control decision layer). All default records and penalty enforcement information are stored on blockchain nodes, serving as the basis for long-term robot credit rating.
[0042] Step 4: Based on the initial task allocation scheme, a distributed model predictive control method is used to generate local paths for each robot; After task allocation, a dedicated motion path is planned for each robot, and each robot constructs a rolling optimization problem with a fixed prediction time domain online: prediction time domain (Corresponding to 1 second), control time domain The objective function comprehensively considers path tracking accuracy, energy consumption, and obstacle avoidance safety, such as: ; in This is the path tracking error, which is the sum of the Euclidean distances between the current position and the reference path. To predict time-domain energy consumption, the energy consumption in the prediction time domain is determined by the real-time power P. The energy consumption in the prediction time domain is then integrated, and the integration interval is the prediction window for rolling optimization, thus obtaining the predicted time-domain energy consumption. Collision risk is the sum of the reciprocals of the current distances between the robot and other robots. The weights for tracking accuracy, energy consumption, and obstacle avoidance safety are respectively determined to meet the following requirements. .
[0043] The constraints set in this embodiment include: Kinematic constraints (Preset) , ); Acceleration constraints (Preset) ); Safety distance constraint: Distance from other robots or obstacles ≥ 0.3m; Task constraints: Actual execution time ; The quadratic programming problem described above is solved online using the interior-point method, with a solution time of ≤50ms. Each control cycle (100ms) executes only the speed and steering angle commands from the first control cycle; the next cycle re-solves the problem based on the updated robot state and environmental information, achieving rolling optimization. This distributed architecture avoids the computational bottleneck of centralized planning, ensuring system real-time performance. Simultaneously, a prediction mechanism anticipates future conflict risks, providing support for subsequent collaborative obstacle avoidance.
[0044] Step 5: Adjust the path in real time through the cooperative obstacle avoidance protocol and deadlock prediction mechanism among robots. At the same time, perform rolling optimization and error compensation based on the real-time deviation between the physical entity and the virtual model in the digital twin environment. Use twin data to drive the model predictive controller to correct subsequent instructions and output a high-precision, adaptive final task execution sequence and cooperative control instructions.
[0045] like Figure 3 As shown, through collaborative obstacle avoidance protocols, deadlock prediction mechanisms, and digital twin rolling optimization, the path is adjusted in real time and operational deviations are corrected, ensuring task execution accuracy and system robustness. The robot broadcasts its own status (position) via 5G point-to-point communication at 200ms intervals. velocity vector The predicted trajectory point for the next 1 second is set with the following safety threshold in this embodiment: safe distance from static obstacles. Safe distance from dynamic obstacles (and other robots) Set the priority ranking rules for obstacle avoidance and passage rights: ; in, This is a task priority mapping value (levels 1-5 are mapped to 5-1, with higher values indicating higher priority). The normalized remaining battery capacity (value between 0 and 1). For load rate, For the corresponding weights; Obstacle avoidance actions are generated based on the artificial potential field method: robots with lower priority are subject to virtual repulsive forces. ,in Here, d represents the repulsive force coefficient, and d represents the actual distance. The virtual gravity is the distance safety threshold. , The gravitational coefficient is determined by the resultant force. The robot is guided to generate deceleration (in this embodiment, the deceleration amplitude is set to ≤0.3m / s²), waiting, or partial path replanning instructions to ensure that the robot can pass through densely populated work areas without collisions.
[0046] A resource occupancy graph is constructed based on graph theory, with nodes representing robots and critical pathways (intersections, sorting stations, etc.). Directed edges represent the robot-resource occupancy relationship. The Tarjan algorithm is used to detect directed cycles in real time, with a detection frequency of 50ms / time. When a circular waiting pattern (a precursor to deadlock) is detected, the central coordinator marks the robot at risk and dynamically adjusts the task execution order of low-priority robots, guiding them into pre-set temporary waiting areas (1.5m × 1.0m) on both sides of the pathway until the resource occupancy cycle is resolved. This pre-emptive prediction mechanism effectively avoids deadlock and improves system stability.
[0047] The digital twin model extrapolates all theoretical execution states of the robot at 100ms intervals and calculates operational deviations in real time. Positional deviation: ; Time deviation: ; in This represents the actual coordinates of the physical robot in the global coordinate system (acquired by UWB+SLAM fusion localization). Represents the theoretical coordinates of the virtual robot in the digital twin model (generated by model derivation). This represents the actual execution time of the physical robot from task initiation to the current moment (collected by the system timer). This represents the theoretical time from task initiation to the corresponding moment, as inferred from a virtual scene by the digital twin model.
[0048] This embodiment sets an error tolerance threshold: , When the deviation exceeds the threshold, a local replanning mechanism is triggered, using the deviation value as a penalty. Add the objective function of the distributed model predictive controller, where This represents the penalty coefficient, and simultaneously updates the kinematic coefficients, energy consumption coefficients, and other state parameters of the digital twin model to achieve closed-loop correction between the physical and virtual systems. The final output is a high-precision task execution sequence and collaborative control instructions, with an execution accuracy of ≤±1mm, meeting the requirements for fine sorting operations.
[0049] The weights, thresholds, prediction time domains, and other parameters in the above steps are set based on the task execution accuracy requirements (such as instruction execution accuracy ≤ ±1mm) and system real-time requirements of large-scale warehouse sorting operations. They can also be flexibly fine-tuned according to the actual warehouse operation scenario.
[0050] The above formulas are all dimensionless calculations, and the preset parameters in the formulas should be set by those skilled in the art according to the actual situation.
[0051] To verify the feasibility, effectiveness, and superiority of the task allocation and collaborative control method for multiple sorting robots described in this invention, a virtual-real fusion simulation environment was built using the MATLAB / Simulink+V-REP co-simulation platform. Comparative experiments were designed based on the actual operational characteristics of large-scale warehouse sorting, and the performance of the scheme was verified from multiple dimensions such as task execution, operational stability, and precision control. During the simulation, all parameters were kept consistent with the parameter configuration of the specific implementation of this invention.
[0052] The bottom simulation layer (V-REP) builds a large-scale 100m×80m warehouse 3D simulation scene, restores physical facilities such as sorting stations, operation channels, temporary waiting areas, and charging areas, constructs 3D models of 50 heterogeneous sorting robots, and configures simulation modules for sensors such as LiDAR, UWB positioning, and industrial cameras to simulate sensor data acquisition (cycle 100ms) and 5G slice communication (end-to-end latency ≤5ms), realizing full-element simulation of the physical entity layer; Algorithm Operation Layer (MATLAB / Simulink): Encapsulates the hierarchical reinforcement learning network (DQN+MAPPO), market auction mechanism, distributed model predictive control, cooperative obstacle avoidance protocol, deadlock prediction mechanism and digital twin closed-loop correction algorithm of this invention, builds the algorithm simulation model of the control decision layer, and realizes real-time data interaction with V-REP (data transmission frequency 10Hz). Digital Twin Simulation Layer: A 1:1 virtual twin model of the V-REP physical scene is built in Simulink to realize real-time mapping and parallel simulation of robot state, environmental information and task execution process, and support simulation verification of deviation calculation, rolling optimization and error compensation.
[0053] Heterogeneous robot configuration: 50 robots are divided into 3 categories, corresponding to light load high speed (maximum load 5kg, maximum speed 1.5m / s), medium load medium speed (maximum load 15kg, maximum speed 1.2m / s), and heavy load low speed (maximum load 30kg, maximum speed 0.8m / s). Each type of robot is randomly distributed in the warehouse area, with the initial state of charge (SOC) of the battery ranging from 50% to 90%. Task flow settings: Generate dynamic random sorting task flow, with a total of 1000 tasks, task priority levels 1 to 5 (level 1 accounts for 10%, level 2 to 3 accounts for 60%, level 4 to 5 accounts for 30%), task weight 0.5 to 25 kg, task deadline 5 to 30 min, and task target points evenly distributed on each sorting station. Environment setup: A 0.1m×0.1m occupancy grid map is used to model the environment, and 10 dynamic moving obstacles (simulating warehouse workers and AGVs) are set. The obstacle movement speed is 0.3~0.8m / s, and the grid status update cycle is 50ms. Comparison scheme: Two sets of comparative experiments were set up, namely the traditional centralized PID scheduling scheme (control group 1) and the simple distributed non-cooperative planning scheme (control group 2). The scheme of this invention is the experimental group. The three sets of experiments used the same robot, task flow and environmental parameters.
[0054] Based on the operational requirements of warehouse sorting robot clusters, five core indicators were selected: task execution efficiency, load balancing, operational stability, execution accuracy, and energy efficiency. According to the simulation test results of three sets of schemes, it was verified whether the performance indicators of the method described in this invention meet the requirements of high timeliness, high accuracy, and high robustness in warehouse sorting operations.
[0055] In this implementation, the control decision layer incorporates the twin bias as a state feedback quantity into the optimization problem of the distributed model predictive controller: when the bias exceeds the corresponding threshold, the penalty weight in the cost function is dynamically updated, and the safety distance constraint with nearby robots / obstacles is tightened simultaneously, so that the controller prioritizes reducing the risk of conflict and correcting the bias in the next prediction window. At the same time, the bias penalty of this cycle is written into the bidding cost, so that robots with large continuous biases will automatically reduce their winning probability in subsequent bidding, thereby realizing a closed-loop adaptive collaboration of twin bias-control optimization-task reallocation.
[0056] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
[0057] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.
Claims
1. A method for cluster task allocation and collaborative control of multiple sorting robots, characterized in that, include: Step 1: Construct a digital twin model of the heterogeneous robot cluster, and collect the status information and capability parameters of each robot in real time through multi-source sensors; Step 2: Establish a digital representation that includes kinematic constraints, load capacity, energy consumption model, communication latency, and spatial pose, and synchronously model obstacles and mission objectives in the dynamic environment. Step 3: Based on digital representation, a hierarchical reinforcement learning network is used to perform coarse global task allocation and fine local task adjustment. The sorting tasks that arrive in real time are decomposed into sub-tasks that can be executed in parallel. Through a market auction mechanism combined with the dynamic priority and load balancing of robots, multiple rounds of bidding arbitration are carried out to dynamically match the real-time capabilities and positions of heterogeneous robots and generate an initial task allocation scheme that takes into account both global optimization and individual benefits. Step 4: Based on the initial task allocation scheme, a distributed model predictive control method is used to generate local paths for each robot; Step 5: Adjust the path in real time through the cooperative obstacle avoidance protocol and deadlock prediction mechanism among robots. At the same time, perform rolling optimization and error compensation based on the real-time deviation between the physical entity and the virtual model in the digital twin environment. Use twin data to drive the model predictive controller to correct subsequent instructions and output a high-precision, adaptive final task execution sequence and cooperative control instructions.
2. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, Status information and capability parameters, specifically including: The robot's real-time position coordinates, running speed and acceleration, current orientation angle, remaining battery power, estimated remaining runtime, maximum load capacity, current load weight, signal strength and transmission delay of the communication module, real-time point cloud data of the surrounding environment collected by the vehicle-mounted sensors, and fault diagnosis status information of the robot itself.
3. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The specific methods for establishing digital representations are as follows: Based on the collected state information and capability parameter data, the extended Kalman filter algorithm is used to perform fusion estimation of the robot pose, and a three-dimensional digital model including kinematic differential constraints, nonholonomic constraints, and energy consumption curve fitting function is constructed. The occupied grid map is used to synchronously label and update dynamic obstacles and task target points in real time.
4. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The hierarchical reinforcement learning network described in step three includes a global scheduling layer and a local execution layer. The specific method is as follows: The global scheduling layer adopts a deep Q-network algorithm to dynamically divide the warehouse area into several sub-regions and coarsely allocate tasks to each sub-region based on the global task flow density and the spatial distribution characteristics of the robot cluster. The local execution layer adopts a multi-agent near-end strategy optimization algorithm to perform secondary fine-tuning and allocation of tasks in each sub-region based on the real-time heterogeneous capabilities of the robots.
5. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The specific method of the market auction mechanism described in step three is as follows: The global scheduling layer acts as the auctioneer, releasing task package information. Each robot, as a bidder, calculates its bid price and expected completion time based on its own status and submits its bid. The auctioneer adopts a combined auction strategy, comprehensively considering the urgency of the task, the robot load balancing, and the global completion time window. After multiple rounds of bidding and arbitration, the task is allocated to the optimal robot combination, and a binding digital task contract is signed.
6. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The distributed model predictive control method described in step four is as follows: Each robot constructs an online rolling optimization problem with a fixed prediction time domain as the window. The objective function is to minimize the sum of path tracking error, its own energy consumption, and collision risk functions with other robots. The constraints include kinematic model, velocity and acceleration boundaries, safe distance maintenance constraints, and task deadline constraints. After solving the problem, the robot executes the first control cycle instruction and updates the system based on the new state in the next cycle.
7. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The cooperative obstacle avoidance protocol mentioned in step five is specifically as follows: Robots communicate with each other via point-to-point, broadcasting their positions, velocity vectors, and predicted trajectories at a fixed frequency. When a robot detects that its relative distance to another robot is less than a preset safety threshold, it dynamically generates a priority ranking based on task priority, remaining battery power, and load status. Robots with lower priority take proactive measures such as deceleration, waiting, and local path replanning. At the same time, a virtual repulsive force is generated using the artificial potential field method to ensure that robots can pass through dense areas without collisions.
8. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, The deadlock prediction mechanism described in step five is as follows: Based on the robot's predicted trajectory and the warehouse road topology, a directed cycle detection algorithm in graph theory is used to analyze the resource occupation relationship between multiple robots in real time. When a deadlock state of circular waiting is predicted, the central coordinator dynamically adjusts the task execution order of some robots and guides them into a temporary waiting area until the resource occupation cycle is resolved.
9. The method for cluster task allocation and collaborative control of multiple sorting robots according to claim 1, characterized in that, Step five, which describes performing rolling optimization and error compensation in the digital twin environment, specifically involves: The theoretical execution states of all robots are simulated in parallel within the digital twin model, and the actual poses of the physical robots are compared with the theoretical poses in the twin model in real time to calculate positional and temporal deviations. When the deviation exceeds the preset error tolerance threshold, a local replanning mechanism is triggered, and the deviation value is fed back as a penalty term to the objective function of the distributed model predictive controller, and the state parameters of the digital twin model are updated synchronously to achieve closed-loop dynamic correction between the physical system and the virtual system.