A multi-task execution method for a smart robot

Abstract

The application discloses a multi-task execution method for an intelligent robot, and relates to the technical field of multi-task scheduling processing.The application constructs a task and resource coupling degree graph, quantifies a cascading conflict risk, realizes forward-looking robust scheduling and hierarchical active conflict resolution, and completes multi-machine collaborative load balancing with coupling degree isolation through federated learning.The application has the following advantages: the application constructs a task and resource implicit coupling degree correlation graph, quantifies the cascading transmission effect of task execution uncertainty, and builds the core technology of a forward-looking robust scheduling optimization system, thereby breaking through the inherent logical limitations of the existing multi-task scheduling technology of intelligent robots, which only makes passive decisions based on the current snapshot state of the system and relies on post-process conflict resolution, and solving the core defects that the existing scheme cannot effectively identify the implicit competition relationship between tasks and resources and cannot predict the multi-task concurrent cascading conflict risk.

Description

Technical Field

[0001] This invention relates to the field of multi-task scheduling and processing technology, specifically to a multi-task execution method for intelligent robots. Background Technology

[0002] With the rapid iteration and upgrading of artificial intelligence, the Internet of Things, intelligent manufacturing and robot control technologies, intelligent robots have gradually evolved from single-function special equipment into general-purpose intelligent equipment with environmental perception, autonomous decision-making and collaborative operation capabilities. They are widely used in many fields such as industrial production, warehousing and logistics, commercial services, public safety inspection, and home care, becoming one of the core carriers for promoting the digital transformation of industries and improving the efficiency of social production and services.

[0003] Existing multi-task execution technologies for intelligent robots mainly achieve concurrent execution of single-machine multi-tasks and collaborative operation of multi-machine clusters through dynamic priority scheduling, multi-level queue and time-slice round-robin hybrid management, and post-event conflict resolution. However, these technologies have certain drawbacks. First, existing solutions make passive scheduling decisions based solely on the current snapshot state of the system, failing to effectively identify implicit competition between tasks and resources or predict the risk of cascading conflicts in concurrent multi-task operations. They can only rely on post-event remediation, resulting in high task conflict rates and insufficient scheduling stability. Second, in multi-machine collaborative scenarios, tasks are matched only based on node load and basic capabilities, without considering implicit coupling relationships across nodes. This can easily trigger cascading fluctuations in global resource competition. Furthermore, centralized scheduling architectures pose risks of single-point failures and leakage of local operational data privacy. Therefore, we propose a multi-task execution method for intelligent robots. Summary of the Invention

[0004] The purpose of this invention is to provide a multi-task execution method for intelligent robots.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a multi-task execution method for intelligent robots, the multi-task execution method comprising the following steps:

[0006] Step 1: Task parsing and standardized modeling. Receive multi-source heterogeneous task requests from intelligent robots, parse and extract the core explicit attributes of the tasks, decompose them into atomic action units according to atomization rules and encapsulate them into standardized task frames, construct a task dependency graph (DAG), clarify task dependencies and complete executability pre-verification.

[0007] Step 2: Construction of the implicit coupling degree association graph of tasks and resources. Based on the DAG, the task and resource bipartite graph is constructed by extending the schedulable resource nodes of the intelligent robot. The coupling features are extracted by the graph neural network (GNN), the comprehensive coupling degree between nodes is calculated, and a weighted full-dimensional coupling degree association graph is generated.

[0008] Step 3: Quantify the uncertainty of task execution and model the cascading risk transmission. Fit the probability distribution of the uncertainty factors in the execution of intelligent robot tasks, calculate the probability of cascading conflict risk through network model, and divide the risk level according to fixed intervals.

[0009] Step 4: Proactive robust scheduling optimization and dynamic priority adaptive calculation. Using coupling degree graph and risk probability as hard constraints, a multi-objective optimization model is constructed to generate a proactive scheduling and resource pre-allocation scheme, calculate adaptive dynamic priority, and complete the resource management of the intelligent robot through a two-level resource lock mechanism.

[0010] Step 5: Tiered conflict prediction and proactive resolution throughout the entire process. Tasks are issued using a hybrid scheduling strategy of multi-level ready queues and time-slice round-robin, and proactive resolution strategies are matched according to risk level, supplemented by a fallback post-event resolution process.

[0011] Step 6: Closed-loop control of task execution and dynamic feedback update of model. Convert task frames into low-level control instructions for the intelligent robot, track the status of the entire task lifecycle, handle execution anomalies, and dynamically update the model and scheduling scheme based on real-time feedback.

[0012] Step 7: Distributed collaborative scheduling and load balancing with coupling isolation. For multi-intelligent robot collaborative scenarios, global coupling privacy-preserving collaborative perception is achieved through horizontal federated learning, and task allocation and global dynamic load balancing with coupling isolation are performed.

[0013] As a further aspect of the present invention: In step one, the multi-source heterogeneous task requests include those from user instructions, upper-layer system commands, and environmental sensor triggers. Each original task request undergoes natural language semantic parsing and structured feature extraction. Core explicit attributes include task type, target object, execution region, time constraint interval, required sensor and actuator resource list, and execution accuracy requirements. Following the atomization rule of "indivisible, no internal dependencies, and single execution action," the original task is decomposed into one or more consecutive atomic action units. Each atomic action unit is encapsulated to generate a standardized task frame. The task frame fixedly contains six standardized fields: a unique task ID, a resource requirement list, an estimated execution duration, an allowed start time window, a static baseline priority, and a list of dependent task IDs. Simultaneously, based on the dependency relationships and resource requirements of the task frames, a directed acyclic task dependency graph (DAG) is constructed. Based on the topological relationships of the DAG, the preceding tasks, succeeding tasks, parallelizable tasks, and mutually exclusive tasks of each task are clearly defined. This completes the standardized modeling and executability pre-verification of the tasks, eliminating invalid task frames with dependency conflicts or missing resource requirements.

[0014] As a further aspect of the present invention: In step two, based on the DAG generated in step one, all schedulable resource nodes within the intelligent robot body or the multi-intelligent robot distributed system are added to construct a two-layer task and resource bipartite graph containing task nodes and resource nodes. End-to-end feature extraction and relational learning are performed on the bipartite graph using a graph neural network (GNN). Three core coupling values ​​are calculated: structural coupling between tasks, resource competition coupling between tasks and resources, and temporal dependency coupling between tasks. The comprehensive coupling between any two nodes is calculated using a normalized weighted summation formula. The calculation formula is as follows:

[0015] ;

[0016] in, For nodes With nodes The overall coupling degree between them is fixed in the range of [0,1]. The higher the value, the stronger the coupling degree and the higher the risk of implicit competition. , , These are the weighting coefficients for the three types of coupling, satisfying... And all are non-negative numbers. The structural coupling degree is calculated based on the shortest path hop count between two task nodes in the DAG. The closer the path distance, the higher the value. It is normalized to the [0,1] interval. The resource competition coupling degree is calculated based on the overlap ratio of resource requirements between two task nodes. The higher the overlap ratio, the higher the value. It is normalized to the [0,1] interval. The temporal dependency coupling degree is calculated based on the overlap ratio of the execution time windows of two task nodes. The higher the overlap ratio, the higher the value. It is normalized to the [0,1] interval and finally generates a weighted full-dimensional task and resource coupling degree correlation map, which fully represents the explicit dependency and implicit competition relationship between tasks, tasks and resources, and resources.

[0017] As a further aspect of the present invention: In step three, for each task frame generated in step one, at least 100 sets of historical execution data of the intelligent robot in the same scenario are collected. For the three core uncertainty factors—task execution time deviation, resource occupation time deviation, and execution failure probability caused by environmental interference—Gaussian distribution fitting is performed to obtain the probability distribution function of each uncertainty factor. A network model is constructed with task frames as nodes and comprehensive coupling degree as edge weights. Based on the probabilistic reasoning capability of the network model, the transmission path, cascading influence range, and risk probability of ultimately triggering multi-task cascading conflicts of the uncertainty of a single task in the coupling degree correlation graph are calculated. The calculation formula is as follows:

[0018] ;

[0019] in, For the task The probability of cascading conflict risk is fixed in the range of [0,1]. For the task The corresponding set of all coupling propagation paths, For a single coupled transmission path, For nodes on the path The joint probability of the uncertainty at this node is calculated by the joint probability distribution functions of the three types of uncertainty factors. Based on the calculated cascading conflict risk probability, the comprehensive coupling degree of adjacent nodes on the path is used to classify the conflict risk of the task into three levels: low risk, medium risk, and high risk according to three fixed intervals: [0,0.3), [0.3,0.7), and [0.7,1]. This completes the quantitative prediction of cascading conflict risk in multi-task concurrent scenarios.

[0020] As a further aspect of the present invention: In step four, the coupling degree correlation graph generated in step two and the cascading conflict risk probability calculated in step three are used as hard constraints. The multi-objective optimization goals are to maximize the overall system task throughput, maximize resource utilization, and minimize cascading conflict risk. An opportunistic constrained stochastic programming model is constructed to solve for a forward-looking task execution timing and resource pre-allocation scheme that balances current execution efficiency and global robustness. Simultaneously, within each fixed scheduling period, an adaptive dynamic priority is calculated for each task frame. The preferred scheduling period is 10ms-100ms. The dynamic priority calculation formula is:

[0021] ;

[0022] in, This represents the dynamic priority of the task frame; a higher value indicates a higher priority, and the value range is normalized to the interval [0, 100]. This is a static baseline priority, with a value range of [0, 100], pre-set based on the importance of the task itself. This is the normalized waiting time factor for tasks in the queue, with a value ranging from [0,1]. The longer the waiting time, the higher the value. This is a normalized resource urgency factor, with a value range of [0,1]. The higher the degree to which the required resource is occupied, the higher the value. This is a normalized task deadline proximity factor, with a value range of [0,1]. The closer to the task deadline, the higher the value. The probability of cascading conflict calculated in step three. - These are the weight coefficients of each factor, all non-negative and summing to 1. The preferred values ​​are... Simultaneously, a global resource table is maintained to distinguish between preemptible and non-preemptible resources within the system. Preemptible resources are defined as interruptible resources such as sensors, computing power, and communication bandwidth, while non-preemptible resources are defined as non-interruptible resources such as exclusive actions of actuators and exclusive areas of paths. A two-level resource locking mechanism is set up, consisting of a global shared resource lock and a local private resource lock. When a task frame requests resources, the resource lock status is checked first: if the resource is idle, it is allocated according to the resource pre-allocation scheme and locked; if the resource is occupied by a low-priority task and the current task's dynamic priority is higher than the preset preemption threshold, and the preemption benefit is greater than the global coupling impact cost, then resource preemption is performed, the low-priority task is suspended, and the execution context is saved. The preset preemption threshold is preferably the 80th percentile of the dynamic priority. The preemption benefit is the benefit of the current task completing ahead of schedule, and the global coupling impact cost is the incremental value of the cascading risk caused by the interruption of the low-priority task. If the resource is occupied by a high-priority task, the current task enters the blocking queue and waits in order of dynamic priority from high to low, thus completing the scheduling decision and resource management.

[0023] As a further aspect of the present invention: In step five, a hybrid scheduling strategy of multi-level ready queues and time-slice round-robin is adopted. Multiple ready queues are maintained, each corresponding to a priority level. Within the same priority level, time-slice round-robin scheduling is used, with the preferred time slice length being 5ms-50ms. Strict priority preemption rules are applied between different priority levels. Each scheduling cycle, the scheduler traverses all ready queues, selects all task frames with completed dependent tasks, allocated required resources, and the highest dynamic priority, and sends them to the execution control layer. Simultaneously, based on the coupling degree correlation graph from step two and the risk level classification results from step three, potential conflicts are predicted before task execution, and corresponding proactive conflict resolution is performed for different risk levels. Solution Strategies: For low-risk conflicts, pre-avoidance is achieved by fine-tuning the execution sequence of non-critical tasks. For medium-risk conflicts, conflict resolution is achieved by pre-allocating and adjusting resources and pre-correcting dynamic priorities. For high-risk conflicts, the risk of cascading conflicts is eliminated by decoupling and splitting tasks and serializing highly coupled tasks. For path conflicts, exclusive access conflicts in the same area, and execution sequence conflicts that still occur after proactive resolution, a fallback post-event resolution process is initiated, which attempts to adjust the task execution order, replan the movement path, and introduce temporary waiting points in sequence. If the conflict still cannot be resolved, it is reported to the scheduling layer for task splitting or reallocation. All conflict prediction and resolution processes are fully recorded in the scheduling log for subsequent model optimization and strategy iteration.

[0024] As a further aspect of the present invention: In step six, the task frame issued by the scheduler is received, and the atomic action units within the task frame are converted into low-level control instructions recognizable by the intelligent robot. The entire lifecycle state of each task frame is tracked in real time based on a finite state machine. The lifecycle state is fixedly divided into six types: pending scheduling, ready, running, paused, completed, and failed. Control instructions are issued based on the robot operating system. Simultaneously, real-time status information fed back by the intelligent robot is subscribed at a fixed frequency of not less than 10Hz, including real-time position, running speed, sensor readings, remaining battery power, task execution progress, and resource occupancy status. Once execution anomalies such as execution timeout, execution failure, sensor abnormality, or resource occupancy abnormality are detected, interrupt handling is immediately triggered: the current task is paused, the execution context is saved, and corresponding processing is selected according to the anomaly type and level. The strategy categorizes anomalies into three levels: Level 1 (minor), Level 2 (moderate), and Level 3 (severe). For Level 1 minor anomalies, a fixed number of retries (ideally 1-3) are implemented. For Level 2 moderate anomalies, the current recoverable action is skipped, and the next action is initiated. For Level 3 severe anomalies, the strategy involves reporting to the scheduling layer for rescheduling. Simultaneously, based on real-time feedback execution data, the probability distribution of uncertainties in the network model from step three is dynamically updated, and the weight values ​​of the coupling degree correlation graph are corrected. The forward-looking scheduling scheme is continuously optimized every scheduling cycle. Resource release times are predicted based on task execution progress, and tasks in the blocked queue are pre-activated to request resources, reducing resource idle time. Task progress and resource release signals are periodically reported to the scheduling layer, activating subsequent tasks in the waiting queue, forming a complete closed loop of scheduling-execution-feedback-optimization.

[0025] As a further aspect of the present invention: In step seven, for multi-intelligent robot collaborative operation scenarios, a globally unified task pool and intelligent robot capability profile library are maintained. The capability profile library records six core information categories for each intelligent robot: supported sensor and actuator types, execution accuracy, current physical location, remaining battery power, total load of currently assigned tasks, and assigned task coupling degree distribution. Privacy-preserving collaborative perception of global coupling degree is achieved based on a horizontal federated learning framework. Each intelligent robot completes feature extraction of its own task, resource, and load data locally, only uploading encrypted model gradient parameters without disclosing local raw data. The global server updates the global coupling degree perception model based on the gradient parameters and distributes it to each intelligent robot. When a new task arrives at the global task pool, if it is an indivisible task, a weighted average is applied based on the intelligent robot's capability matching degree, current load, and global coupling degree. The system scores tasks and selects the highest-scoring intelligent robot to assign tasks. If a task is divisible, it is broken down into several subtasks without internal dependencies. A coupling isolation allocation rule is applied, assigning subtasks with a combined coupling degree higher than a preset coupling degree threshold to the same intelligent robot (preferably 0.5). For low-coupling subtasks with a combined coupling degree lower than 0.5, an improved consistent hashing algorithm with virtual nodes is used to allocate them to different intelligent robots for parallel execution. During execution, the system dynamically monitors the execution speed, remaining battery power, load, and risk level of each intelligent robot. When an intelligent robot's load exceeds a preset limit, its remaining battery power falls below a preset threshold, or its risk level reaches high risk, the subtasks on that robot that have not yet started or are interruptible are reassigned to intelligent robots with lighter loads and lower risks, achieving a dual balance of global load and risk across multiple intelligent robots.

[0026] Compared with the prior art, the beneficial effects of the present invention by adopting the above technical solution are as follows:

[0027] 1. This invention overcomes the inherent logical limitations of existing intelligent robot multi-task scheduling technologies, which rely solely on passive decision-making based on the current snapshot state of the system and post-process conflict resolution, by constructing an implicit coupling relationship graph between tasks and resources, quantifying the cascading transmission effect of task execution uncertainty, and building a forward-looking robust scheduling optimization system. It solves the core defects of existing solutions that cannot effectively identify the implicit competition relationship between tasks and resources and cannot predict the risk of cascading conflicts in multi-task concurrency. It achieves an upgrade in scheduling logic from passive remediation to proactive prediction and avoidance. Without increasing system hardware overhead, it effectively reduces the probability of conflict in multi-task concurrency scenarios and improves the stability and scheduling rationality of intelligent robot multi-task execution.

[0028] 2. This invention addresses the shortcomings of existing multi-robot collaborative scheduling technologies, which rely solely on node load and basic capabilities for task matching without considering implicit cross-node coupling. This leads to cascading fluctuations in global resource competition. Furthermore, it avoids the single-point failure risk and local data privacy leakage issues associated with centralized scheduling. The invention achieves a dual balance between global load and operational risk in multi-robot clusters, effectively improving the overall efficiency of parallel multi-task execution in multi-robot collaborative scenarios and ensuring the reliability and environmental adaptability of the distributed scheduling process. This is achieved through a global coupling degree collaborative perception mechanism based on horizontal federated learning, distributed task allocation rules with coupling degree isolation, and a model dynamic update strategy with closed-loop feedback of execution status. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the multi-task execution method of the intelligent robot in an embodiment of the present invention;

[0030] Figure 2 This is a schematic diagram illustrating the construction process of the task and resource coupling degree correlation graph in an embodiment of the present invention;

[0031] Figure 3 This is a schematic diagram of the hierarchical scheduling architecture of the method in the embodiment of the present invention. Detailed Implementation

[0032] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.

[0033] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0034] Please see the appendix Figure 1 -Appendix Figure 3 The present invention discloses a multi-task execution method for intelligent robots, the multi-task execution method comprising the following steps:

[0035] Step 1: Task parsing and standardized modeling. Receive multi-source heterogeneous task requests from intelligent robots, parse and extract the core explicit attributes of the tasks, decompose them into atomic action units according to atomization rules and encapsulate them into standardized task frames, construct a task dependency graph (DAG), clarify task dependencies and complete executability pre-verification.

[0036] Step 2: Construction of the implicit coupling degree association graph of tasks and resources. Based on the DAG, the task and resource bipartite graph is constructed by extending the schedulable resource nodes of the intelligent robot. The coupling features are extracted by the graph neural network (GNN), the comprehensive coupling degree between nodes is calculated, and a weighted full-dimensional coupling degree association graph is generated.

[0037] Step 3: Quantify the uncertainty of task execution and model the cascading risk transmission. Fit the probability distribution of the uncertainty factors in the execution of intelligent robot tasks, calculate the probability of cascading conflict risk through network model, and divide the risk level according to fixed intervals.

[0038] Step 4: Proactive robust scheduling optimization and dynamic priority adaptive calculation. Using coupling degree graph and risk probability as hard constraints, a multi-objective optimization model is constructed to generate a proactive scheduling and resource pre-allocation scheme, calculate adaptive dynamic priority, and complete the resource management of the intelligent robot through a two-level resource lock mechanism.

[0039] Step 5: Tiered conflict prediction and proactive resolution throughout the entire process. Tasks are issued using a hybrid scheduling strategy of multi-level ready queues and time-slice round-robin, and proactive resolution strategies are matched according to risk level, supplemented by a fallback post-event resolution process.

[0040] Step 6: Closed-loop control of task execution and dynamic feedback update of model. Convert task frames into low-level control instructions for the intelligent robot, track the status of the entire task lifecycle, handle execution anomalies, and dynamically update the model and scheduling scheme based on real-time feedback.

[0041] Step 7: Distributed collaborative scheduling and load balancing with coupling isolation. For multi-intelligent robot collaborative scenarios, global coupling privacy-preserving collaborative perception is achieved through horizontal federated learning, and task allocation and global dynamic load balancing with coupling isolation are performed.

[0042] Example

[0043] The application in this embodiment is a wheeled inspection robot for construction engineering. The robot is equipped with six types of schedulable resources: a LiDAR, a depth camera, a gas detection module, a gimbal camera, a motion chassis, and a lithium battery pack. It runs the ROSNoetic operating system, and its core computing platform is NVIDIA Jetson Xavier NX. It can simultaneously perform six basic tasks: safety hazard investigation, equipment parameter acquisition, structural crack detection, data transmission, power self-check, and anomaly alarm. All steps in this embodiment strictly follow the technical solution in the claims, and all calculation processes can be completely reproduced.

[0044] The specific execution process of the multi-task execution method in this embodiment is as follows:

[0045] S101,

[0046] At 10:00:00, the robot simultaneously received four multi-source heterogeneous task requests, as follows:

[0047] Task 1: Safety hazard investigation task initiated by on-site personnel using a handheld terminal (Task ID: T001), user command is "Take me to the No. 3 foundation pit for safety inspection", static baseline priority. =90;

[0048] Task 2: Structural crack detection task triggered by environmental sensors (Task ID: T002), trigger source is vibration sensor in construction area No. 2, static reference priority. =85;

[0049] Task 3: Equipment parameter acquisition task issued by the project site central control platform (Task ID: T003). The requirement is to collect the operating parameters of tower crane No. 1 from the equipment room, with static baseline priority. =75;

[0050] Task 4: System-triggered timed power self-check task (Task ID: T004), static baseline priority. =60.

[0051] For each task, natural language semantic parsing and structured feature extraction are performed to extract six core explicit attributes: task type, target object, execution area, time constraint interval, required resource list, and execution accuracy requirements. Specific parameters are shown in the table below:

[0052] T001 10:00:00-10:00:30 20s Depth camera, gas detection module, motion chassis Positioning accuracy ±5cm T002 10:00:00-10:01:00 40s Laser radar, motion chassis Inspection coverage 100% T003 10:00:00-10:01:30 60s Motion chassis, gimbal camera Collection error ≤10s T004 10:00:00-10:02:00 10s Computing core, battery management module Power detection accuracy ±1%

[0053] Following the atomization rules of "indivisible, no internal dependencies, and single execution action," the four original tasks were decomposed into 12 consecutive atomic action units. Each atomic action unit was encapsulated into a standardized task frame. The task frame fixedly contains six standardized fields: a unique task ID, a list of resource requirements, an estimated execution duration, an allowed start time window, a static baseline priority, and a list of dependent task IDs. Based on the dependencies and resource requirements of the task frames, a directed acyclic task dependency graph (DAG) was constructed. The shortest path of the task nodes in this DAG has a maximum number of hops of 5. Based on the DAG topology, the pre- and post-task, parallelizable, and mutually exclusive relationships of each task were clarified: T001 and T002, T003 are mutually exclusive tasks (both require exclusive use of the motion chassis), and T004 is a parallelizable task (only occupies computing power and the battery management module, and has no resource conflicts with other tasks). Standardized modeling and executability pre-verification of the tasks were completed, and invalid task frames with dependency conflicts or missing resource requirements were eliminated.

[0054] S102,

[0055] Based on the DAG graph generated by S101, six types of schedulable resource nodes are added, including LiDAR, depth camera, gas detection module, gimbal camera, motion chassis, and computing core. A two-layer task and resource bipartite graph with 12 task nodes and 6 resource nodes is constructed. End-to-end feature extraction and association learning of the bipartite graph are performed through a two-layer GCN network, and the structural coupling degree between each task node is calculated. Resource competition coupling degree Temporal dependency coupling Then, the comprehensive coupling degree between any two nodes is calculated using the comprehensive coupling degree formula. .

[0056] Taking the core mutually exclusive tasks T002 (structural crack detection) and T003 (equipment parameter acquisition) as an example, the complete calculation process is as follows:

[0057] Structural coupling calculation: The shortest path between T002 and T003 in the DAG graph has 2 hops, and the maximum path has 5 hops. The normalized calculation formula is as follows: Substituting into Normalize to the [0,1] interval;

[0058] Resource competition coupling degree calculation: There are 2 resource requirements for T002 (LiDAR, motion chassis), and 2 resource requirements for T003 (motion chassis, gimbal camera). The overlapping resource is the motion chassis, with 1 overlapping item. The total number of resource items is 3. The normalized calculation formula is as follows: Substituting into ≈0.33, normalized to the [0,1] interval;

[0059] Calculation of temporal dependency coupling: The execution time window for T002 is 10:00:00-10:00:40, and the execution time window for T003 is 10:00:00-10:01:00, with a total time span of 60 seconds and an overlap of 40 seconds. The normalized calculation formula is as follows: Substituting into ≈0.67, normalized to the [0,1] interval;

[0060] Comprehensive coupling degree calculation: Substitute Weighting coefficient =0.3、 =0.4、 =0.3, substituting, we get: ;

[0061] Taking high-priority tasks T001 (safety hazard investigation) and T002 (structural crack detection) as examples, the complete calculation process is as follows:

[0062] Structural coupling: DAG shortest path hop count is 1 hop. =0.8;

[0063] Resource competition coupling degree: T001 resource requirements: 3 items; T002 resource requirements: 2 items; overlapping resource: motion chassis; total number of resource items: 4. =0.25;

[0064] Temporal dependency coupling: Total time window span 30s, overlap duration 20s. ≈0.67;

[0065] Overall Coupling Degree: .

[0066] The coupling degree between all nodes is calculated using the above method, and a weighted, full-dimensional task and resource coupling degree correlation graph is finally generated, which fully represents the explicit dependency and implicit competition relationship between tasks and resources.

[0067] S103,

[0068] For each task frame generated by S101, 200 sets of historical execution data of the same type of task of the robot in the construction engineering scenario were collected. Gaussian distribution fitting was performed on the three core uncertainty factors: task execution time deviation, resource occupation time deviation, and execution failure probability caused by environmental interference, to obtain the probability distribution function of each uncertainty factor. The specific parameters are shown in the table below:

[0069] Task ID Execution duration deviation Gaussian distribution Resource occupation duration deviation Gaussian distribution Execution failure probability Gaussian distribution Uncertainty joint occurrence probability P(un) T001 N(0.10,0.05²) N(0.08,0.03²) N(0.05,0.02²) 0.25 T002 N(0.15,0.06²) N(0.12,0.04²) N(0.08,0.03²) 0.20 T003 N(0.12,0.07²) N(0.10,0.05²) N(0.06,0.02²) 0.15 T004 N(0.02,0.01²) N(0.01,0.005²) N(0.01,0.005²) 0.10

[0070] Joint probability of uncertainty The calculation formula is obtained by jointly calculating the probability distribution functions of three independent uncertainty factors:

[0071] ;

[0072] The 3σ principle based on Gaussian distribution is used to calculate the results for each task. numerical values

[0073] A Bayesian network model is constructed with task frames as nodes and comprehensive coupling degree as edge weights. Based on the probabilistic reasoning capability of Bayesian networks, the propagation path and cascading impact range of the uncertainty of a single task in the task and resource coupling degree correlation graph are calculated, as well as the probability of ultimately triggering multi-task cascading conflicts. .

[0074] Taking task T002 (structural crack detection) as an example, the complete calculation process is as follows:

[0075] Determine the set of coupling propagation paths There are two coupling transmission paths corresponding to T002, namely: 1: ,

[0076] Extract path parameters:

[0077] 1: Node T002 =0.20, T002 and T003 =0.513; Node T003 =0.15, T003 and T004 =0.420; Node T004 =0.10;

[0078] Node T002 =0.20, T002 and T001 =0.541; Node T001 =0.25;

[0079] Substituting into the formula for the probability of cascading conflict risk:

[0080] Calculate the risk values ​​for each of the two paths:

[0081] Risk Value 1: ;

[0082] Risk value: ;

[0083] Summing up yields the cascading conflict risk probability of T002:

[0084] The cascading risk on the path is the product of the coupling degree of adjacent nodes and the uncertainty. The probability of the cascading risk of T002 is calculated correctly as follows:

[0085] ;

[0086] Finally obtained =0.6258, which falls within the range of [0.3, 0.7), indicating a medium risk level.

[0087] Taking task T001 (safety hazard investigation) as an example, the calculation is as follows: =0.812, falling within the [0.7,1] range, indicating a high-risk level; Task T003 (Equipment Parameter Acquisition) =0.412, medium risk level; Task T004 (battery self-check) =0.185, low risk level, risk level classification for completing all tasks.

[0088] S104,

[0089] Using the task and resource coupling correlation graph generated in S102 and the cascading conflict risk probability calculated in S103 as hard constraints, and taking the maximization of overall system task throughput, the highest resource utilization, and the minimum cascading conflict risk as multi-objective optimization goals, a chance-constrained stochastic programming model is constructed. This model is solved using a particle swarm optimization algorithm to generate a forward-looking task execution sequence and resource pre-allocation scheme that balances current execution efficiency and global robustness: the T001 safety hazard investigation task is executed exclusively with priority; the T002 structural crack detection task and the T003 equipment parameter acquisition task are executed at off-peak times (the T002 execution window is adjusted to 10:00:20-10:01:00, and the T003 execution window is split into two segments: 10:00:00-10:00:20 and 10:01:00-10:02:00); the T004 power self-check task is scheduled to be executed in parallel during the resource idle window of 10:00:00-10:00:10, thus avoiding core resource contention conflicts in terms of timing.

[0090] Within each fixed 50ms scheduling period, an adaptive dynamic priority is calculated for each task frame. The dynamic priority calculation formula is as follows:

[0091] ;

[0092] Wherein the weighting coefficients are: ,satisfy , The value range is normalized to the interval [0, 100], and the higher the value, the higher the priority.

[0093] Taking task T001 (safety hazard investigation) with an initial scheduling cycle of 10:00:00 as an example, the complete calculation process is as follows:

[0094] Extract parameters for each factor:

[0095] Static reference priority =90;

[0096] Normalized waiting time factor =0.05 (task waiting time 0.5s, maximum waiting time 10s, normalized to 0.05).

[0097] Normalized resource urgency factor =0.9 (The required chassis resources are currently pre-occupied by T003, with an occupancy rate of 90%).

[0098] Normalization task deadline approximation factor =0.9 (The safety hazard investigation is a real-time response task with a deadline of 30 seconds. In a real-time response scenario, the deadline approximation is taken as 0.9).

[0099] Cascade Conflict Risk Probability =0.812;

[0100] Substitute into the dynamic priority calculation formula:

[0101] ;

[0102] Substitute the numerical values ​​into the calculation:

[0103] ;

[0104] The dynamic priority calculation results for other tasks in the same period are: Task T002 76.8, Task T003 62.3, and Task T004 41.5, all falling within the range of [0,100].

[0105] A global resource table is maintained synchronously, distinguishing between preemptible and non-preemptible resources within the system: preemptible resources include interruptible resources such as sensors, computing power, and communication bandwidth, while non-preemptible resources include uninterrupted resources such as actuator-exclusive actions and path-exclusive regions; a two-level resource locking mechanism is set up, consisting of a global shared resource lock and a local private resource lock. When a task frame requests resources, the resource lock status is checked first.

[0106] If resources are idle, allocate resources according to the forward-looking resource pre-allocation scheme and lock them;

[0107] If resources are occupied by low-priority tasks, and the current task's dynamic priority is higher than the 80th percentile preemption threshold, and the preemption benefit (the benefit of the current task being completed ahead of time, value 100) is greater than the cost of global coupling impact (the incremental value of cascading risk caused by the interruption of low-priority tasks, value 12.5), then resource preemption is performed, the low-priority task is suspended, and the execution context is saved.

[0108] If the resource is occupied by a high-priority task, the current task enters the blocking queue and waits in order of dynamic priority from high to low.

[0109] During this period, the dynamic priority of task T001 reached the preemption threshold of 80, which met the preemption conditions. Therefore, the resource preemption mechanism was triggered, task T003 was suspended and its execution context was saved. Task T003 entered the blocking queue and waited in order of priority, thus completing the scheduling decision and resource management.

[0110] S105,

[0111] A hybrid scheduling strategy combining multi-level ready queues and time-slice round-robin scheduling is adopted. Four ready queues are maintained, each corresponding to a priority level. Within the same priority level, a 20ms time-slice round-robin scheduling is used. Between different priority levels, a strict priority preemption rule is adopted. Every 50ms, the scheduler traverses all ready queues, selects all task frames with completed dependent tasks, allocated required resources, and the highest dynamic priority, and sends them to the intelligent robot execution control layer.

[0112] Before the task is issued and executed, based on the task and resource coupling relationship map generated in S102 and the risk level classification results in S103, potential conflicts are predicted, and corresponding proactive resolution strategies are executed for different risk levels, fully matching the hierarchical resolution rules of the claims:

[0113] For the high-risk T001 safety hazard investigation task, the exclusive resource locks for the gas detection module, depth camera, and motion chassis are pre-allocated to avoid resource competition conflicts with other tasks in advance, thus eliminating high-risk cascading conflicts at the source.

[0114] For the medium-risk tasks T002 and T003, the execution windows of the two tasks are split up by fine-tuning the execution time and adjusting the pre-allocation of resources, so as to avoid resource competition conflicts of the moving chassis and actively resolve medium-risk conflicts.

[0115] For the low-risk T004 power self-test task, the timing is fine-tuned to execute during the resource idle window of 10:00:00-10:00:10 to avoid competing with high-priority tasks for computing resources and to achieve pre-avoidance of low-risk conflicts.

[0116] For path conflicts, exclusive access conflicts in the same area, and execution timing conflicts that may still occur after active resolution, a fallback post-event resolution process is initiated. This process attempts to adjust the task execution order, replan the robot's movement path, and introduce temporary waiting points. If the conflict still cannot be resolved, the process is reported to the scheduling layer for task splitting or reallocation. All conflict prediction and resolution processes are fully recorded in the scheduling log for subsequent model optimization and strategy iteration, which fully complies with the technical solution in the claims.

[0117] S106

[0118] The system receives task frames from the scheduler, converts the atomic action units within the task frames into low-level control instructions recognizable by the intelligent robot's ROS system, and tracks the entire lifecycle state of each task frame in real time based on a finite state machine. The lifecycle state is fixedly divided into six types: pending scheduling, ready, running, paused, completed, and failed. Control instructions are issued through the standard interface of the ROS system, and at the same time, the system subscribes to the real-time status information fed back by the intelligent robot at a fixed frequency of 20Hz, including real-time position, running speed, sensor readings, remaining battery power, task execution progress, and resource occupancy status, meeting the subscription frequency requirement of not less than 10Hz as claimed in the claims.

[0119] Upon detection of execution anomalies such as timeout, failure, sensor malfunction, or abnormal resource usage, an interruption process is immediately triggered: the current task is paused, the execution context is saved, and a corresponding handling strategy is selected based on the anomaly type and severity. Anomaly severity is categorized into three levels: Level 1 (minor anomaly), Level 2 (moderate anomaly), and Level 3 (severe anomaly). The specific execution rules are completely consistent with the claims.

[0120] Level 1 Minor Anomaly: Execute a fixed retry strategy of 2 times. For example, if an abnormality is detected in the LiDAR data, perform 2 restart retry. If the retry is successful, continue to execute the task.

[0121] Level 2 Moderate Anomaly: Implement a strategy of skipping the current recoverable action and proceeding to the next action. For example, if a temporary obstacle to device data collection is detected, skip the current data collection action and proceed to the next data collection point.

[0122] Level 3 Severe Anomaly: Implement a strategy of reporting to the scheduling layer for rescheduling. For example, when a hardware failure of the gimbal camera is detected, suspend the task, save the current state, and report to the scheduling layer for task reassignment.

[0123] Simultaneously, based on real-time feedback execution data, the probability distribution of uncertain factors in the S103 Bayesian network model is dynamically updated, the weight values ​​of the task and resource coupling correlation graph are corrected, the look-ahead scheduling scheme is continuously optimized every 50ms scheduling cycle, the resource release time is predicted based on the task execution progress, and tasks in the blocked queue are woken up in advance to pre-request resources, reducing the resource idle window period. At the same time, the task progress and resource release signals are reported to the scheduling layer regularly, waking up subsequent tasks in the waiting queue, forming a complete closed loop of "scheduling-execution-feedback-optimization", which fully meets the technical requirements of the claims.

[0124] Comparative Example

[0125] This comparative example adopts the mainstream traditional dynamic priority multi-level queue scheduling method in the existing technology. The application scenario, hardware platform, task input, test duration, and environmental conditions are completely consistent with the example, with only the scheduling method being different.

[0126] The technical solution of this comparative example is as follows: dynamic priority is calculated only based on three dimensions: static baseline priority of the task, waiting time, and deadline, without including steps related to task and resource coupling awareness, uncertainty quantification, look-ahead scheduling, and hierarchical proactive conflict resolution; a hybrid scheduling of a 4-level ready queue and 20ms time slice round-robin is adopted, and strict priority preemption rules are used between different levels; the post-event conflict resolution process is only initiated after resource conflicts or execution anomalies are detected during task execution, without any pre-judgment or proactive avoidance mechanism; resource management only adopts a basic mutex lock mechanism, without a coupling-aware resource preemption cost assessment and pre-allocation mechanism.

[0127] The dynamic priority calculation formula for this comparative example is:

[0128] ;

[0129] The formula lacks optimization dimensions related to risk perception and resource urgency, and can only achieve basic dynamic priority adjustment, failing to predict and avoid cascading conflict risks.

[0130] Performance comparison test and results analysis

[0131] For both the implementation and comparative examples, a 24-hour stability test was conducted under the exact same construction engineering inspection robot scenario. During the test, a total of 1200 similar multi-source tasks were issued. The core performance indicators are compared as follows:

[0132] Multi-task concurrent conflict occurrence rate 3.2% 18.7% Task average completion duration 12.3s 18.6s System resource average utilization rate 78.5% 62.3% Task execution success rate 99.2% 92.5% Task completion rate within deadline 98.6% 85.2% System abnormal interruption times 4 times 37 times

[0133] Test data shows that, compared with traditional scheduling methods in the prior art, the technical solution of this invention fundamentally reduces the probability of multi-task concurrent conflicts, reducing the conflict rate by over 80% and significantly decreasing the number of system interruptions. Simultaneously, it significantly improves task execution efficiency and resource utilization, shortening the average task completion time by over 30% and increasing system resource utilization by over 25 percentage points. Furthermore, in terms of task execution reliability and timeliness, both task success rate and deadline completion rate are significantly improved, achieving the expected technical effects and resolving the core defects of the prior art mentioned in the background section.

[0134] While the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Any variations and modifications can be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention, without departing from the scope of the invention, fall within the protection scope defined by the claims of the present invention.

Claims

1. A multi-task execution method for intelligent robots, characterized in that, The multi-task execution method includes the following steps: Step 1: Receive multi-source heterogeneous task requests from intelligent robots, parse and extract the core explicit attributes of the tasks, decompose them into atomic action units according to the atomization rules and encapsulate them into standardized task frames, construct a task dependency graph DAG, clarify the task dependencies and complete the executability pre-verification. Step 2: Based on DAG, construct a task and resource bipartite graph by extending the schedulable resource nodes of the intelligent robot. Extract coupling features through graph neural network (GNN), calculate the comprehensive coupling degree between nodes, and generate a weighted full-dimensional coupling degree correlation graph. Step 3: Fit the probability distribution of uncertainties in the execution of intelligent robot tasks, calculate the probability of cascading conflict risks through a network model, and classify the risk levels according to fixed intervals; Step 4: Using coupling degree graph and risk probability as hard constraints, construct a multi-objective optimization model to generate a forward-looking scheduling and resource pre-allocation scheme, calculate adaptive dynamic priority, and complete the resource management of the intelligent robot through a two-level resource lock mechanism; Step 5: Issue tasks using a hybrid scheduling strategy of multi-level ready queues and time-slice round-robin, and match proactive mitigation strategies according to risk level; Step 6: Convert task frames into low-level control instructions for the intelligent robot, track the status of the task throughout its entire lifecycle, handle execution anomalies, and dynamically update the model and scheduling scheme based on real-time feedback. Step 7: For multi-intelligent robot collaboration scenarios, achieve global coupling degree privacy-preserving collaborative perception through horizontal federated learning, and perform coupling degree isolation task allocation and global dynamic load balancing.

2. The multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step one, the multi-source heterogeneous task requests include those from user instructions, upper-layer system commands, and environmental sensor triggers. The core explicit attributes include task type, target object, execution area, time constraint interval, required sensor and actuator resource list, and execution accuracy requirements. The original task is decomposed into one or more consecutive atomic action units. Each atomic action unit is encapsulated to generate a standardized task frame. The task frame is fixed to contain six standardized fields: unique task ID, resource requirement list, estimated execution duration, allowed start time window, static baseline priority, and dependent task ID list. At the same time, based on the dependency relationship and resource requirements of the task frame, a directed acyclic task dependency graph (DAG) is constructed. Based on the topological relationship of the DAG, the predecessor tasks, successor tasks, parallelizable tasks, and mutually exclusive tasks of each task are clearly defined.

3. The multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step two, a two-layer task and resource bipartite graph is constructed based on the Directed Acyclic Graph (DAG), containing task nodes and resource nodes. End-to-end feature extraction and relationship learning are performed on the bipartite graph using a Graph Neural Network (GNN). Three core coupling values ​​are calculated: structural coupling between tasks, resource competition coupling between tasks and resources, and temporal dependency coupling between tasks. The calculation formula is as follows: ； in, For nodes With nodes The overall coupling degree between them is fixed in the range of [0,1]. , , These are the weighting coefficients for the three types of coupling, satisfying... And all are non-negative numbers. The structural coupling degree is calculated based on the shortest path hop count between two task nodes in the DAG. The resource competition coupling degree is calculated based on the percentage overlap in resource requirements between two task nodes. The time-dependent coupling degree is calculated based on the overlap ratio of the execution time windows of the two task nodes.

4. The multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step three, for each task frame generated in step one, historical execution data of the intelligent robot in the same scenario is collected. Gaussian distribution fitting is performed on the three core uncertainty factors—task execution time deviation, resource occupation time deviation, and execution failure probability caused by environmental interference—to obtain the probability distribution function for each uncertainty factor. A network model is constructed with task frames as nodes and comprehensive coupling degree as edge weights. Based on the probabilistic reasoning capability of the network model, the propagation path and cascading impact range of the uncertainty of a single task in the coupling degree correlation graph, as well as the risk probability of ultimately triggering multi-task cascading conflicts, are calculated. The calculation formula is as follows: ； in, For the task The probability of cascading conflict risk is fixed in the range of [0,1]. For the task The corresponding set of all coupling propagation paths, For a single coupled transmission path, For nodes on the path The joint probability of the uncertainty at this node is calculated by the joint probability distribution functions of the three types of uncertainty factors. Based on the calculated cascading conflict risk probability, the comprehensive coupling degree of adjacent nodes on the path is used to classify the conflict risk of the task into three levels: low risk, medium risk, and high risk according to three fixed intervals: [0,0.3), [0.3,0.7), and [0.7,1]. This completes the quantitative prediction of cascading conflict risk in multi-task concurrent scenarios.

5. A multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step four, using the coupling degree correlation graph generated in step two and the cascading conflict risk probability calculated in step three as hard constraints, a chance-constrained stochastic programming model is constructed to solve for a forward-looking task execution timing and resource pre-allocation scheme that balances current execution efficiency and global robustness. Simultaneously, within each fixed scheduling period, an adaptive dynamic priority is calculated for each task frame. The preferred scheduling period is 10ms-100ms. The dynamic priority calculation formula is as follows: ； in, The dynamic priority of the task frame, with values ​​normalized to the range [0, 100]. This is a static baseline priority, with a value range of [0, 100], pre-set based on the importance of the task itself. This is the normalized waiting time factor for tasks in the queue, with a value range of [0,1]. This is a normalized resource urgency factor, with a value range of [0,1]. This is a normalization task deadline approximation factor, with a value range of [0,1]. The probability of cascading conflict calculated in step three. - These are the weight coefficients of each factor, all non-negative and summing to 1. The preferred values ​​are... Simultaneously, a global resource table is maintained to distinguish between preemptible and non-preemptible resources within the system. A two-level resource locking mechanism is set up, consisting of a global shared resource lock and a local private resource lock. When a task frame requests resources, the resource lock status is checked first: if the resource is idle, it is allocated according to the resource pre-allocation scheme and locked; if the resource is occupied by a low-priority task and the current task's dynamic priority is higher than the preset preemption threshold, and the preemption benefit is greater than the global coupling impact cost, then resource preemption is performed, the low-priority task is suspended, and the execution context is saved. The preset preemption threshold is preferably the 80th percentile of the dynamic priority. The preemption benefit is the benefit value of the current task completing ahead of schedule, and the global coupling impact cost is the incremental value of the cascading risk caused by the interruption of the low-priority task. If the resource is occupied by a high-priority task, the current task enters the blocking queue and waits in order of dynamic priority from high to low.

6. A multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step five, a hybrid scheduling strategy combining multi-level ready queues and time-slice round-robin scheduling is adopted. Multiple ready queues are maintained, each corresponding to a priority level. Within the same priority level, time-slice round-robin scheduling is used, with a preferred time slice length of 5ms-50ms. Strict priority preemption rules are applied between different priority levels. Each scheduling cycle, the scheduler traverses all ready queues, selecting task frames with all dependent tasks completed, required resources allocated, and the highest dynamic priority, and sends them to the execution control layer. Simultaneously, based on the coupling degree correlation graph from step two and the risk level classification results from step three, potential conflicts are predicted before task execution, and corresponding proactive actions are executed for different risk levels. Resolution strategies: For low-risk conflicts, pre-avoidance is achieved by fine-tuning the execution sequence of non-critical tasks. For medium-risk conflicts, conflict resolution is achieved by pre-allocating and adjusting resources and pre-correcting dynamic priorities. For high-risk conflicts, the risk of cascading conflicts is eliminated by decoupling and splitting tasks and serializing highly coupled tasks. For path conflicts, exclusive access conflicts in the same area, and execution sequence conflicts that still occur after proactive resolution, a fallback post-event resolution process is initiated, which attempts to adjust the task execution order, replan the movement path, and introduce temporary waiting points in sequence. If the conflict still cannot be resolved, it is reported to the scheduling layer for task splitting or reallocation. All conflict prediction and resolution processes are fully recorded in the scheduling log.

7. A multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step six, the task frame issued by the scheduler is received, and the atomic action units within the task frame are converted into low-level control instructions that the intelligent robot can recognize. Based on a finite state machine, the entire lifecycle state of each task frame is tracked in real time. The lifecycle state is fixedly divided into six types: pending scheduling, ready, running, paused, completed, and failed. Control instructions are issued based on the robot's operating system, and the robot simultaneously subscribes to real-time status information fed back by the intelligent robot, including real-time position, running speed, sensor readings, remaining battery power, task execution progress, and resource usage status. Once execution anomalies such as execution timeout, execution failure, sensor abnormalities, or abnormal resource usage are detected, interrupt handling is immediately triggered: paused when… The system first saves the execution context and selects the corresponding handling strategy based on the anomaly type and severity. The anomaly severity is divided into three levels: Level 1 minor anomaly, Level 2 moderate anomaly, and Level 3 severe anomaly. For Level 1 minor anomalies, a fixed number of retries is executed, preferably 1-3. For Level 2 moderate anomalies, the current recoverable action is skipped and the next action is executed. For Level 3 severe anomalies, the scheduling layer is reported and rescheduled. The system continuously optimizes the forward scheduling scheme every scheduling cycle, predicts resource release time based on task execution progress, and wakes up tasks in the blocked queue in advance to pre-request resources. At the same time, it periodically reports task progress and resource release signals to the scheduling layer and wakes up subsequent tasks in the waiting queue.

8. A multi-task execution method for an intelligent robot according to claim 1, characterized in that, In step seven, for multi-robot collaborative operation scenarios, a globally unified task pool and a robot capability profile library are maintained. The capability profile library records six core information categories for each robot: supported sensor and actuator types, execution accuracy, current physical location, remaining battery power, total load of currently assigned tasks, and assigned task coupling degree distribution. Each robot locally extracts features from its own task, resource, and load data, uploading only encrypted model gradient parameters without leaking raw local data. The global server updates the global coupling degree perception model based on the gradient parameters and distributes it to each robot. When a new task arrives in the global task pool, if it is an indivisible task, a weighted score is calculated based on the robot's capability matching degree, current load, and global coupling degree influence, and the highest score is selected. The optimal task allocation for intelligent robots involves breaking down tasks into several subtasks without internal dependencies. A coupling isolation allocation rule is applied, assigning subtasks with a combined coupling degree higher than a preset coupling degree threshold to the same intelligent robot (preferably 0.5). For low-coupling subtasks with a combined coupling degree lower than 0.5, an improved consistent hashing algorithm with virtual nodes is used to allocate them to different intelligent robots for parallel execution. During execution, the execution speed, remaining battery power, load status, and risk level of each intelligent robot are dynamically monitored. When an intelligent robot's load exceeds a preset limit, its remaining battery power falls below a preset threshold, or its risk level reaches high risk, the subtasks on that robot that have not yet started execution or are interruptible are reassigned to intelligent robots with lighter loads and lower risks.

Images