A robot management system control method based on big data

Abstract

This invention discloses a big data-based robot management system control method, relating to the fields of artificial intelligence and automated control technology. The method includes: constructing a unified data representation model that integrates four types of heterogeneous data: task semantics, environmental graph, robot dynamics, and communication state; deploying a distributed reinforcement learning agent based on a dual-delay deep deterministic policy gradient algorithm for multi-robot collaborative decision-making; implementing an improved dynamic window method that integrates communication connectivity and social compliance constraints for local path replanning; utilizing a long short-term memory network to predict the state in the next 12 seconds to intervene in conflicts in advance; and designing a standardized control command conversion layer to adapt to various heterogeneous robots. This invention achieves stable collaborative control with a task completion rate ≥97% and a response latency <3 seconds in high-density complex environments, supports hot-swappable access and self-recovery scheduling, and significantly improves system robustness, versatility, and operational efficiency.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and automation control technology, and in particular to a control method for a robot management system based on big data. Background Technology

[0002] With the deep integration of artificial intelligence and big data technologies, big data-based robot management systems have demonstrated significant application value in diverse scenarios such as industrial automation, intelligent services, and special operations. These systems integrate historical robot trajectories, real-time environmental perception data, task semantic descriptions, and resource status information to construct a data-driven decision-making mechanism. This enables intelligent scheduling and closed-loop control of robot behavior, thereby improving overall operational efficiency, system robustness, and task completion quality. Against this technological backdrop, robot management has evolved from the independent control of single devices to a complex systems engineering approach oriented towards multi-machine collaboration, multi-task concurrency, and dynamic environmental adaptation. Its core lies in establishing a unified data fusion framework and a globally optimized control strategy.

[0003] Among them, the big data-based robot management system control method focuses on using massive heterogeneous data to manage robot swarms throughout their entire lifecycle, covering multiple functional dimensions such as task allocation, path planning, state prediction, resource scheduling, and fault response. The core objective of this method is to construct a joint optimization model in a high-dimensional state space, mapping multi-source information such as task requirements, environmental constraints, energy consumption characteristics, and robot capabilities into a sequence of executable control commands, thereby achieving efficient, reliable, and scalable collaborative operations in dynamic and uncertain environments. However, most current technical solutions are still limited to specific hardware platforms or preset task types, failing to form a standardized control paradigm applicable to a general robot ecosystem.

[0004] While some existing solutions incorporate big data analytics, they suffer from multiple structural flaws in system architecture and control logic. First, existing methods generally lack a unified representation capability for multi-dimensional heterogeneous data, failing to effectively integrate key elements such as task priority, environmental semantic graphs, robot dynamics models, and real-time communication status, resulting in a semantic gap between high-level decision-making and low-level execution. Second, in multi-robot collaborative scenarios, existing control strategies often employ static task allocation or local path replanning mechanisms, struggling to cope with dynamic disturbances such as sudden obstacles, task insertions, or robot failures, leading to severe deficiencies in global optimality and system resilience. Furthermore, current systems generally lack state evolution prediction models based on time-series data, failing to identify potential conflicts or resource bottlenecks in advance, resulting in delayed control response and impacting overall operational efficiency. In addition, existing solutions exhibit weak cross-platform compatibility; their control logic is highly coupled to specific robot types, making it difficult to adapt to various robot entities. These problems are particularly pronounced in complex, unstructured environments, directly hindering the large-scale deployment and intelligent upgrading of robot management systems in the real world. Summary of the Invention

[0005] The purpose of this invention is to provide a control method for a robot management system based on big data, in order to solve the problems of insufficient multi-dimensional heterogeneous data fusion capability, lag in dynamic disturbance response, lack of global collaborative optimization and weak cross-platform compatibility in existing systems.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: A control method for a robot management system based on big data includes the following specific steps: Step S1: Construct a unified data representation model, collect four types of heterogeneous data: task semantic description, environmental 3D semantic map, robot dynamic parameters and real-time communication status. By embedding an encoder, unstructured text and topological structure are mapped to a 256-dimensional vector space to generate task feature vectors, environmental state tensors, individual capability matrices and link quality sequences with unified formats, which serve as the basic inputs for subsequent joint optimization. Step S2: Establish a multi-machine collaborative decision-making framework. Based on the unified data representation results, deploy a distributed reinforcement learning agent at the central scheduling node. Use a dual-delay deep deterministic policy gradient algorithm for task allocation and resource reservation. The action space is defined as a pair of robot numbers and task identifiers. The reward function is composed of a weighted sum of task completion degree, energy consumption coefficient and time deviation, with weight factors of 0.6, 0.3 and 0.1, respectively. The policy is updated every 10 seconds. Step S3: Implement a dynamic path replanning mechanism. After receiving the initial global path solution, each robot's local controller combines LiDAR and visual sensors to detect changes in obstacle positions in real time. When the actual passage area deviates from the preset trajectory by more than 15 centimeters, an improved dynamic window method is activated for local obstacle avoidance. The velocity sampling resolution is 0.05 m / s, the angular velocity resolution is 0.02 radians / s, and the evaluation function introduces the inverse distance squared term of the pedestrian predicted trajectory to enhance social compliance. Step S4: Construct a time-series state evolution prediction model. Use a long short-term memory network to perform sequence modeling on the historical operation data stream. The input includes the location coordinates, power percentage, load weight and communication delay collected every 200 milliseconds in the past 60 seconds. The output is the state prediction sequence for the next 12 seconds. The prediction step size is 1 second. This is used to identify potential conflict areas or resource competition scenarios in advance and trigger pre-intervention control commands. Step S5: Implement a cross-morphology robot adaptation interface, design a standardized control command conversion layer, receive abstract behavior commands output from higher-level decisions, match execution parameter templates by looking up a table based on the target robot type, and specify the differential drive gain for wheeled robots. 1.2 With a value of 0.8, the legged robot is mapped to a 12-joint servo cycle waveform, the aircraft is parsed into quaternion attitude adjustment commands, and the robotic arm is converted into inverse kinematics solution input, ensuring the consistency and portability of the control logic on different physical platforms.

[0007] In step S1, the embedded encoder uses a hierarchical attention mechanism to process the semantic description of the task. First, it extracts key words through word-granular encoding, and then aggregates the three attributes of task objective, priority and deadline through sentence-level self-attention. Finally, it outputs a fixed-length task feature vector, in which high-priority tasks are assigned a value of 0.9 in the third dimension of the vector, ordinary tasks are assigned a value of 0.3, and low-priority tasks are assigned a value of 0.1.

[0008] In step S1, the environmental 3D semantic map is generated by the fusion of a multi-source SLAM system. It includes a static map layer, a dynamic object layer, and a functional area layer. Each layer is stored in a rasterized format with a resolution of 0.1 meters. The dynamic object layer is refreshed every 500 milliseconds. The functional area layer is labeled with three types of tags: charging area, restricted area, and cooperative area. The tag values ​​are set to 1, 2, and 3, respectively, for the decision-making module to call.

[0009] In step S2, the distributed reinforcement learning agent adopts a parameter-sharing architecture. All robots share a neural network backbone but maintain an independent experience replay buffer. During training, the target network parameters are synchronized using a soft update method. The decay coefficient τ is set to 0.01, the learning rate α is 0.001, and the discount factor γ is 0.99 to ensure the convergence stability of the policy.

[0010] In step S2, the task completion rate calculation of the reward function is based on the task progress percentage P and the expected progress. The absolute difference is expressed as: An additional 0.2 reward is added when the task is completed ahead of schedule, and a 0.5 penalty is deducted if the task is interrupted due to robot malfunction, thus maintaining the positive guiding role of the incentive mechanism.

[0011] In step S3, the improved dynamic window method adds communication connectivity constraints to the traditional velocity space constraints. It requires that the endpoint of the candidate trajectory must be located in an area where the signal strength is greater than -85dBm with at least two neighboring robots. Otherwise, the trajectory is forcibly eliminated to ensure that the cluster maintains network topology connectivity during movement.

[0012] In step S3, the social compliance term of the evaluation function is determined by the closest distance d between the predicted path points of surrounding pedestrians in the next 3 seconds and the robot's candidate trajectory, and its contribution value is... The proportionality coefficient k is set to 0.5, which forces the robot to actively maintain a safe distance and move along the natural flow direction in the crowd environment.

[0013] In step S4, the Long Short-Term Memory (LSTM) network contains two hidden layers, each with 128 neurons. The activation function is tanh. During the training phase, the mean squared error loss function is used for backpropagation. The batch size is... Set the parameter to 64, the number of iterations (epochs) to 200, and freeze the model parameters after training for online inference.

[0014] The output of the state prediction sequence in step S4 is used to construct a conflict warning matrix. When the Euclidean distance between the predicted positions of any two robots is less than 0.8 meters within the next 8 seconds, it is determined that there is a risk of collision. The system automatically issues a pre-adjustment command to both robots to decelerate to 70% of their original speed or yaw by 15 degrees to avoid emergency braking causing a chain reaction.

[0015] In step S5, the execution parameter template is stored in a local database in JSON format. Each type of robot configuration file includes a kinematic model identifier and a control cycle. Maximum linear velocity Maximum angular velocity Four core fields, wheeled robot For 50 milliseconds, 1.5 m / s The aircraft has a speed of 1.0 radians per second. For 20 milliseconds, 3.0 m / s The speed is 2.5 radians / second, ensuring that the accuracy and timeliness of instruction parsing match the hardware characteristics.

[0016] The robot management system control method based on big data also includes an abnormal working condition self-recovery mechanism. When the robot fails to respond to the scheduling command or report the fault code three times in a row, the system automatically re-injects its current task into the waiting queue and recalculates the optimal receiver based on the latest environmental state tensor and the available robot capability matrix. The faulty robot enters the diagnostic mode and performs data uploads of battery voltage, motor current and inertial measurement unit until its health status is restored to more than 90% before it can re-register for service.

[0017] The big data-based robot management system control method also includes an energy balance scheduling strategy. The system monitors the remaining battery percentage of all robots in real time. When a robot's battery level drops below 20%, a return-to-home request is triggered. The scheduling module prioritizes planning the shortest path to the nearest available charging station for the robot and suspends the assignment of new tasks. Other robots take over the unfinished work according to the load balancing principle to ensure that the overall battery life of the cluster is maximized.

[0018] The big data-based robot management system control method also includes a federated learning-assisted model update mechanism. The environmental interaction data accumulated by each robot locally is uploaded to the edge server after being de-identified, and participates in the incremental training of the global LSTM prediction model. The gradient is aggregated and updated every 24 hours. The FedAvg algorithm is used to merge local model parameters. The number of communication rounds E is 5, and the client sampling ratio C is 0.3, which improves the model's generalization and adaptability to unknown scenarios.

[0019] The big data-based robot management system control method is applied to a city-level intelligent inspection system, managing no fewer than 500 heterogeneous robots, covering three types of entities: ground inspection vehicles, aerial drones, and underground pipeline robots. The system has an average task response latency of less than 3 seconds, a task completion rate of greater than or equal to 97%, and a maximum daily concurrent task processing capacity of 12,000 items, meeting the stable operation requirements in high-density and complex environments.

[0020] The big data-based robot management system control method supports hot-swappable device access. After a new robot is powered on, it broadcasts its own model and capability tags. The central node automatically loads the corresponding control template and safety constraint rules by matching the pre-stored device library. System integration can be completed without manual programming intervention. The entire device registration process takes no more than 8 seconds, significantly improving operation and maintenance efficiency.

[0021] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention achieves deep integration of four-dimensional information—task, environment, individual, and communication—by constructing a unified data representation model, eliminating the semantic gap between high-level decision-making and low-level execution, and enabling the robot management system to understand complex unstructured scenarios.

[0022] This invention employs a distributed reinforcement learning agent to achieve dynamic task allocation and resource reservation within a multi-machine collaborative decision-making framework. Compared to static allocation strategies, this improves the task completion rate by at least 18 percentage points and significantly enhances the resilience of the robot management system in responding to sudden task insertions and robot failures.

[0023] This invention introduces a local replanning mechanism based on an improved dynamic window method, embedding communication connectivity and social behavior constraints in the velocity space, enabling robots to maintain swarm collaboration and interpersonal interaction safety even in dynamic obstacle environments.

[0024] This invention establishes a state evolution prediction model based on long short-term memory networks, which can accurately predict the system state within the next 12 seconds, identify potential conflicts in advance and issue pre-intervention instructions, and reduce the frequency of emergency braking events to less than 23% of the original level.

[0025] This invention designs a standardized control command conversion layer and an extensible execution parameter template library, supporting plug-and-play access for various robot forms such as wheeled, legged, flying, and robotic arms. It completely breaks the dependence of control logic on specific hardware platforms, forming a truly universal robot management paradigm.

[0026] This invention integrates three enhancement mechanisms: self-recovery from abnormal operating conditions, balanced energy scheduling, and federated learning-assisted updates. These mechanisms further improve the robustness, endurance, and model evolution speed of the robot management system, fully meeting the long-term stable operation requirements of demanding application scenarios such as industrial automation, urban inspection, and disaster relief. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the overall technical solution architecture of the robot management system control method based on big data proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the multi-machine collaborative decision-making framework in this invention; Figure 3 This is a schematic diagram illustrating the specific steps of the control method for a robot management system based on big data proposed in this invention. Detailed Implementation

[0028] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0029] Example 1 Currently, with the deep integration of artificial intelligence and big data technologies, big data-based robot management systems are demonstrating significant application value in diverse scenarios such as industrial automation, intelligent services, and special operations. These systems integrate historical robot trajectories, real-time environmental perception data, task semantic descriptions, and resource status information to construct a data-driven decision-making mechanism. This enables intelligent scheduling and closed-loop control of robot behavior, thereby improving overall operational efficiency, system robustness, and task completion quality. Against this macro-technological backdrop, robot management has evolved from the independent control of single devices to a complex systems engineering approach oriented towards multi-machine collaboration, multi-task concurrency, and dynamic environmental adaptation. Its core lies in establishing a unified data fusion framework and a globally optimized control strategy. To address the aforementioned technical issues, this invention proposes a unified data representation model that deeply integrates four-dimensional information—task, environment, individual, and communication—eliminating the semantic gap between high-level decision-making and low-level execution, enabling the system to understand complex unstructured scenarios. A distributed reinforcement learning agent is employed in a multi-machine collaborative decision-making framework to achieve dynamic task allocation and resource reservation, improving task completion rate by at least 18 percentage points compared to static allocation strategies and significantly enhancing the system's resilience in the face of unexpected task insertions and robot failures. A local replanning mechanism based on an improved dynamic window method is introduced, embedding communication connectivity and social behavior constraints in the velocity space, enabling robots to maintain cluster collaboration and interpersonal interaction safety even in dynamic obstacle environments. Finally, a state evolution prediction model based on a long short-term memory network is established. The system can accurately predict the system state within the next 12 seconds, identify potential conflicts in advance, and issue pre-intervention commands, reducing the frequency of emergency braking events to less than 23% of the original level. It designs a standardized control command conversion layer and an extensible execution parameter template library, supporting plug-and-play access for various robot forms such as wheeled, legged, flying, and robotic arms, completely breaking the dependence of control logic on specific hardware platforms and forming a truly universal robot management paradigm. It integrates three enhancement mechanisms: abnormal working condition self-recovery, energy balance scheduling, and federated learning-assisted updates, further improving the system's robustness, endurance, and model evolution speed. It fully meets the long-term stable operation requirements of high-demand application scenarios such as industrial automation, urban inspection, and disaster relief, and is applied to a big data-based robot management system control method.

[0030] refer to Figure 1The overall technical architecture of the big data-based robot management system control method proposed in this invention comprises five main parts: a central scheduling node, a heterogeneous robot cluster, a multi-source perception network, an edge computing unit, and a remote monitoring terminal. The central scheduling node, serving as the decision-making hub of the entire system, is equipped with four core functional modules: a unified data representation model, a multi-machine collaborative decision-making framework, a temporal state evolution prediction model, and a cross-morphological adaptation interface. The heterogeneous robot cluster consists of no fewer than 500 ground inspection vehicles, aerial drones, and underground pipeline robots. Each robot is equipped with a local controller, LiDAR, vision sensor, inertial measurement unit, and wireless communication module. The multi-source perception network is responsible for collecting 3D semantic maps of the environment and dynamic obstacle information. The edge computing unit is used to perform federated learning to assist in model updates and local data anonymization processing. The remote monitoring terminal provides a human-machine interface and visualization of the system's health status. All components are interconnected through high-speed, low-latency communication links, forming a closed-loop feedback, dynamically evolving intelligent control system.

[0031] In a big data-based robot management system control method, step S1 involves constructing a unified data representation model, collecting four types of heterogeneous data: task semantic description, environmental 3D semantic map, robot dynamic parameters, and real-time communication status. By embedding an encoder, unstructured text and topological structure are mapped to a 256-dimensional vector space, generating task feature vectors, environmental state tensors, individual capability matrices, and link quality sequences with a unified format, which serve as the basic inputs for subsequent joint optimization.

[0032] Specifically, in step S1, the embedded encoder uses a hierarchical attention mechanism to process the task semantic description. First, key words are extracted through word-granular encoding. Then, the task objective, priority, and deadline are aggregated through sentence-level self-attention. Finally, a fixed-length task feature vector is output, where high-priority tasks are assigned a value of 0.9 in the third dimension of the vector, ordinary tasks are assigned a value of 0.3, and low-priority tasks are assigned a value of 0.1. The task semantic description is input in natural language form, such as "complete the inspection of power facilities in area A within 10 minutes". After word segmentation and word embedding, an initial word vector sequence is formed. Then, the inter-word dependencies are calculated through a multi-head self-attention mechanism to extract key entities such as "power facilities", "area A", and "10 minutes". Finally, the representation is aggregated into a sentence-level representation through a gated recurrent unit and mapped to a fixed position in a 256-dimensional vector space. The environmental 3D semantic map is generated by a multi-source SLAM system, comprising a static map layer, a dynamic object layer, and a functional area layer. Each layer is stored in a rasterized format with a resolution of 0.1 meters. The dynamic object layer is refreshed every 500 milliseconds. The functional area layer is labeled with three categories: charging area, restricted area, and cooperative area, with label values ​​set to 1, 2, and 3, respectively, for use by the decision-making module. The static map layer constructs the indoor building structure using laser SLAM. The dynamic object layer integrates visual object detection and millimeter-wave radar point cloud tracking results. The functional area layer is pre-labeled by maintenance personnel and imported through a geographic information system. Robot dynamic parameters, including maximum linear velocity, maximum angular velocity, acceleration limits, mass distribution, and joint degrees of freedom, are reported to the central node in structured JSON format. Real-time communication status is generated by each robot periodically broadcasting signal strength, channel occupancy, and packet loss rate, forming a link quality sequence with a sampling frequency of 10 Hz. All four types of data are processed by an embedded encoder, outputting a tensor structure of the same dimension, facilitating subsequent joint feature extraction and optimization calculations by the neural network.

[0033] In a big data-based robot management system control method, step S2 involves establishing a multi-machine collaborative decision-making framework. Based on unified data representation results, a distributed reinforcement learning agent is deployed at the central scheduling node. A dual-delay deep deterministic policy gradient algorithm is used for task allocation and resource reservation. The action space is defined as a pairing set of robot numbers and task identifiers. The reward function consists of a weighted sum of task completion degree, energy consumption coefficient, and time deviation, with weight factors of 0.6, 0.3, and 0.1, respectively. A policy update is performed every 10 seconds.

[0034] Specifically, in step S2, the distributed reinforcement learning agent adopts a parameter-sharing architecture. All robots share a common neural network backbone but maintain independent experience replay buffers. During training, soft updates are used to synchronize the target network parameters. The decay coefficient τ is set to 0.01, the learning rate α is 0.001, and the discount factor γ is 0.99 to ensure policy convergence stability. The central scheduling node maintains a global state observer, aggregating the state tensors of all robots and the queue of tasks to be assigned every 10 seconds to form an environmental snapshot at the current decision moment. This snapshot serves as the input to the reinforcement learning agent, is compressed by the encoder, and fed into the Actor network, outputting continuous action vectors, which are then discretized and mapped to specific robot-task pairings. The Critic network evaluates the long-term reward of this pairing to guide the policy optimization direction. The task completion degree of the reward function is calculated based on the task progress percentage P and the expected progress. The absolute difference is expressed as: An additional 0.2 reward is added when the task is completed ahead of schedule; a 0.5 penalty is deducted if the task is interrupted due to robot malfunction, maintaining the positive guiding effect of the incentive mechanism. The energy consumption coefficient is determined by the ratio of the robot's current power consumption rate to the historical average, and the time deviation is the absolute difference between the actual start time and the planned start time divided by the task window length. The weighted sum of these three factors constitutes the total reward signal, driving the strategy towards high efficiency, energy saving, and timeliness. The experience replay buffer capacity is set to 10,000 trajectories, and a priority experience replay mechanism is used to accelerate the learning of key samples. The strategy update process is executed in parallel on a GPU-accelerated server to ensure real-time requirements within a 10-second cycle.

[0035] In a big data-based robot management system control method, step S3 involves implementing a dynamic path replanning mechanism. After receiving the initial global path solution, each robot's local controller combines LiDAR and visual sensors to detect changes in obstacle positions in real time. When the actual passage area deviates from the preset trajectory by more than 15 centimeters, an improved dynamic window method is activated for local obstacle avoidance. The velocity sampling resolution is 0.05 m / s, and the angular velocity resolution is 0.02 radians / s. The evaluation function incorporates the squared inverse distance term of the pedestrian predicted trajectory to enhance social compliance.

[0036] Specifically, in step S3, the improved dynamic window method adds communication connectivity constraints to the traditional velocity space constraints. It requires that the endpoint of a candidate trajectory must be located within an area where the signal strength of at least two neighboring robots is greater than -85dBm; otherwise, the trajectory is forcibly eliminated, ensuring that the cluster maintains network topology connectivity during movement. The local controller performs obstacle detection every 100 milliseconds. The LiDAR point cloud is segmented into individual obstacles using Euclidean clustering, and the vision sensor identifies pedestrians using the YOLOv5 model and predicts their trajectory for the next 3 seconds. Local replanning is triggered when the robot's current position deviates laterally from the global path by more than 15 centimeters or when a new obstacle appears ahead. The velocity sampling space is in... and Within the range, the system discretizes the trajectory with step sizes of 0.05 m / s and 0.02 radians / s to generate candidate velocity pairs. Each candidate velocity pair is simulated forward for 2 seconds to generate a sequence of trajectory points. The evaluation function consists of four terms: target orientation, velocity smoothing, obstacle distance, and social compliance. The social compliance term is determined by the closest distance *d* between the predicted path points of surrounding pedestrians in the next 3 seconds and the robot's candidate trajectory, and its contribution value is... A proportionality coefficient k is set to 0.5, forcing the robot to actively maintain a safe distance and move along the natural flow direction in crowded environments. After comprehensive scoring of all candidate trajectories, the highest scorer is selected as the execution command for the next control cycle. This mechanism ensures obstacle avoidance safety while taking into account swarm communication connectivity and human social habits, making it suitable for high-density human-robot coexistence scenarios such as shopping malls and train stations.

[0037] In a big data-based robot management system control method, step S4 involves constructing a temporal state evolution prediction model. This model utilizes a long short-term memory network to perform sequence modeling on historical operational data streams. The input includes location coordinates, battery percentage, load weight, and communication delay collected every 200 milliseconds over the past 60 seconds. The output is a state prediction sequence for the next 12 seconds, with a prediction step size of 1 second. This model is used to identify potential conflict areas or resource competition scenarios in advance and trigger pre-intervention control commands.

[0038] Specifically, in step S4, the Long Short-Term Memory (LSTM) network contains two hidden layers, each with 128 neurons. The activation function is tanh. During the training phase, the mean squared error loss function is used for backpropagation, and the batch size is... The system is configured with a parameter set to 64 and an iteration count of 200 epochs. After training, the model parameters are frozen for online inference. The data acquisition module obtains state data from each robot at a frequency of 5 Hz, forming a time series. Each sequence contains 300 time steps (60 seconds × 5 Hz), and each time step includes 4-dimensional features: x-coordinate, y-coordinate, z-coordinate (if applicable), battery percentage, load weight (kg), and communication latency (milliseconds). After normalization, the sequence is input into an LSTM network, which outputs predictions for the next 12 time steps (12 seconds). The output of the state prediction sequence is used to construct a conflict warning matrix. If the predicted Euclidean distance between any two robots is less than 0.8 meters within the next 8 seconds, a collision risk is identified, and the system automatically issues a pre-emptive order to both robots to decelerate to 70% of their original speed or yaw by 15 degrees to avoid a chain reaction caused by emergency braking. The warning matrix is ​​updated every second and maintained by a dedicated conflict detection thread. Furthermore, if the predicted battery depletion time is less than the remaining task time, an energy balancing scheduling strategy is triggered in advance. The predictive model is deployed on an edge server, with inference latency controlled within 50 milliseconds to ensure the timeliness of pre-intervention instructions.

[0039] In a big data-based robot management system control method, step S5 involves implementing a cross-morphological robot adaptation interface, designing a standardized control command conversion layer, receiving abstract behavioral commands output from higher-level decisions, and matching execution parameter templates based on the target robot type, with the wheeled robot corresponding to the differential drive gain. 1.2 With a value of 0.8, the legged robot is mapped to a 12-joint servo cycle waveform, the aircraft is parsed into quaternion attitude adjustment commands, and the robotic arm is converted into inverse kinematics solution input, ensuring the consistency and portability of the control logic on different physical platforms.

[0040] Specifically, in step S5, the execution parameter template is stored in a local database in JSON format, and each type of robot configuration file contains a kinematic model identifier and a control cycle. Maximum linear velocity Maximum angular velocity Four core fields, wheeled robot For 50 milliseconds, 1.5 m / s The aircraft has a speed of 1.0 radians per second. For 20 milliseconds, 3.0 m / s The speed is 2.5 radians / second to ensure that the instruction parsing accuracy and timeliness match the hardware characteristics. Abstract behavior commands are generated by the high-level decision-making module and take the following form: The control command translation layer first parses the target robot ID, queries its type tag, and then loads the corresponding parameter template. For wheeled robots, the target position is converted into left and right wheel speed commands, and the gain is calculated using the differential drive formula. Used for linear velocity amplification Used for angular velocity suppression to prevent over-steering. For legged robots, a central pattern generator is invoked to generate phase coordination waveforms for 12 joints, achieving stable gait. The aircraft receives the quaternion target attitude and adjusts the motor speed via a PID controller. The robotic arm uses the Jacobian matrix to solve the inverse kinematics and outputs the joint angle sequence. All conversion processes are completed on the local controller, and the central node does not need to know the underlying hardware details, thus achieving decoupling between the control logic and the physical platform.

[0041] Building upon the aforementioned methods, the big data-based robot management system control method further integrates an abnormal operating condition self-recovery mechanism. When a robot fails to respond to scheduling commands or report fault codes three consecutive times, the system automatically re-injects its current task into the assignment queue and recalculates the optimal assignee based on the latest environmental state tensor and the available robot capability matrix. The faulty robot enters diagnostic mode, performing battery voltage, motor current, and inertial measurement unit (IMU) data uploads until its health status recovers to over 90% before it can re-register for service. The health status is calculated using a weighted average of multi-dimensional indicators, including battery internal resistance, motor temperature rise, and IMU zero-bias stability; thresholds are determined through statistical analysis of historical fault data.

[0042] The big data-driven robot management system simultaneously implements an energy-balanced scheduling strategy, monitoring the remaining battery percentage of all robots in real time. When a robot's battery drops below 20%, a return-to-home request is triggered. The scheduling module prioritizes planning the shortest path to the nearest available charging station for that robot, while suspending new task assignments. Other robots take over the unfinished work according to load balancing principles, ensuring maximum overall cluster endurance. The charging station status is reported in real time by IoT sensors, and path planning uses A / B algorithm. The algorithm combines a dynamic traffic flow model to avoid highly congested areas.

[0043] Furthermore, a big data-based robot management system control method is implemented, deploying a federated learning-assisted model update mechanism. Environmental interaction data accumulated locally by each robot is anonymized and then uploaded to an edge server to participate in the incremental training of the global LSTM prediction model. Gradient updates are aggregated every 24 hours, and the Fed Avg algorithm is used to merge local model parameters. The communication rounds E is 5, and the client sampling ratio C is 0.3, improving the model's generalization adaptability to unknown scenarios. Anonymization processes include removing precise geographic locations, obfuscating timestamps, and adding differential privacy noise to ensure user data security.

[0044] In a specific application example of a city-level intelligent inspection system, this invention manages no fewer than 500 heterogeneous robots, encompassing three types of entities: ground inspection vehicles, aerial drones, and underground pipeline robots. Ground inspection vehicles are responsible for inspecting roads and building facades, equipped with 360-degree LiDAR and high-definition cameras; aerial drones perform high-altitude power line and roof inspections, equipped with thermal imagers; and underground pipeline robots enter narrow pipe networks, using ultrasonic and gas sensors. The system's average task response latency is less than 3 seconds, the task completion rate is greater than or equal to 97%, and the maximum daily concurrent task processing capacity reaches 12,000 items. The new robots support hot-swappable device access; upon power-up, they broadcast their model and capability tags. The central node automatically loads the corresponding control templates and safety constraint rules by matching the pre-stored device library, completing system integration without manual programming intervention. The entire device registration process takes no more than 8 seconds.

[0045] refer to Figure 2 The core principles of the multi-machine collaborative decision-making framework include a state encoder, an Actor-Critic network, an experience replay pool, and a policy update engine. The state encoder receives the output of a unified data representation model and generates a 1024-dimensional state vector. The Actor network is a three-layer fully connected network that outputs the action probability distribution. The Critic network evaluates the value of actions. The experience replay pool stores data in partitions by robot ID, and the policy update engine triggers a training loop every 10 seconds. This framework effectively solves the problem of dynamic task allocation in large-scale heterogeneous clusters.

[0046] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A control method for a robot management system based on big data, characterized in that, The specific steps include the following: Step S1: Construct a unified data representation model, collect four types of heterogeneous data: task semantic description, environmental 3D semantic map, robot dynamic parameters and real-time communication status, and map unstructured text and topology to a 256-dimensional vector space by embedding an encoder to generate task feature vectors, environmental state tensors, individual capability matrices and link quality sequences with unified formats. Step S2: Establish a multi-machine collaborative decision-making framework, deploy distributed reinforcement learning agents at the central scheduling node, and use a dual-delay deep deterministic policy gradient algorithm for task allocation and resource reservation. The action space is defined as a set of paired robot numbers and task identifiers, and the reward function is composed of a weighted sum of task completion degree, energy consumption coefficient and time deviation. Step S3: Implement a dynamic path replanning mechanism. After receiving the initial global path solution, each robot's local controller combines LiDAR and visual sensors to detect changes in obstacle positions in real time. When the actual passage area deviates from the preset trajectory, the improved dynamic window method is activated for local obstacle avoidance. The evaluation function introduces the inverse distance squared term of the pedestrian predicted trajectory to enhance social compliance. Step S4: Construct a time-series state evolution prediction model. Use a long short-term memory network to perform sequence modeling on the historical operation data stream. Input location coordinates, power percentage, load weight and communication delay, output state prediction sequence, identify potential conflict areas or resource competition scenarios in advance, and trigger pre-intervention control commands. Step S5: Implement cross-morphology robot adaptation interface, design standardized control command conversion layer, receive abstract behavior commands output by high-level decision-making, and match execution parameter templates by looking up a table according to the target robot type.

2. The robot management system control method based on big data according to claim 1, characterized in that: In step S1, the embedded encoder uses a hierarchical attention mechanism to process the semantic description of the task. First, it extracts key words through word-granular encoding, and then aggregates the three attributes of task objective, priority and deadline through sentence-level self-attention. Finally, it outputs a fixed-length task feature vector, in which high-priority tasks are assigned a value of 0.9 in the third dimension of the vector, ordinary tasks are assigned a value of 0.3, and low-priority tasks are assigned a value of 0.

1.

3. The robot management system control method based on big data according to claim 1, characterized in that: In step S1, the environmental 3D semantic map is generated by the fusion of a multi-source SLAM system. It includes a static map layer, a dynamic object layer, and a functional area layer. Each layer is stored in a rasterized format with a resolution of 0.1 meters. The dynamic object layer is refreshed every 500 milliseconds. The functional area layer is labeled with three types of tags: charging area, restricted area, and cooperative area. The tag values ​​are set to 1, 2, and 3, respectively.

4. The robot management system control method based on big data according to claim 1, characterized in that: In step S2, the distributed reinforcement learning agent adopts a parameter-sharing architecture. All robots share a neural network backbone but maintain an independent experience replay buffer. During training, the target network parameters are synchronized using a soft update method. The decay coefficient τ is set to 0.01, the learning rate α is 0.001, and the discount factor γ is 0.

99.

5. The robot management system control method based on big data according to claim 1, characterized in that: In step S2, the task completion rate calculation of the reward function is based on the absolute difference between the task progress percentage P and the expected progress Pe. When the task is completed ahead of schedule, an additional reward increment of 0.2 is added. If the interruption is caused by robot failure, a penalty of 0.5 is deducted.

6. The robot management system control method based on big data according to claim 1, characterized in that: In step S3, the improved dynamic window method adds a communication connectivity constraint to the traditional velocity space constraint. It requires that the endpoint of the candidate trajectory must be located in an area where the signal strength is greater than -85dBm with at least two neighboring robots; otherwise, the trajectory is forcibly eliminated.

7. The robot management system control method based on big data according to claim 1, characterized in that: In step S3, the social compliance term of the evaluation function is determined by the closest distance between the predicted path points of surrounding pedestrians in the next 3 seconds and the robot's candidate trajectory.

8. The robot management system control method based on big data according to claim 1, characterized in that: In step S4, the Long Short-Term Memory network contains two hidden layers, each with 128 neurons. The activation function is tanh. During the training phase, the mean squared error loss function is used for backpropagation. The batch size is set to 64, and the number of iterations is 200.

9. The robot management system control method based on big data according to claim 1, characterized in that: The output of the state prediction sequence in step S4 is used to construct a conflict warning matrix. When the Euclidean distance between the predicted positions of any two robots is less than 0.8 meters within the next 8 seconds, it is determined that there is a risk of collision. The system automatically issues a pre-adjustment order to both robots to decelerate to 70% of their original speed or yaw by 15 degrees.

10. The robot management system control method based on big data according to claim 1, characterized in that: In step S5, the execution parameter template is matched by looking up a table according to the target robot type, and the differential drive gain is corresponding to the wheeled robot. 1.2 The value is 0.

8. The legged robot is mapped to a 12-joint servo cycle waveform, the aircraft is analyzed as a quaternion attitude adjustment command, and the robotic arm is converted into an inverse kinematics solution input. In step S5, the execution parameter template is stored in a local database in JSON format. Each type of robot configuration file includes a kinematic model identifier, control period Tc, and maximum linear velocity. Maximum angular velocity Four core fields, wheeled robot For 50 milliseconds, 1.5 m / s The aircraft has a speed of 1.0 radians per second. For 20 milliseconds, 3.0 m / s It is 2.5 radians per second.

Images