Control method, system and equipment of security robot and medium

Abstract

The invention relates to a control method, system and device for a security robot and a medium, and the method solves the problem of perception failure in smog and low-light scenes through the adaptive adjustment of the weight of a multi-modal sensor through dynamic weighting space-time alignment processing. Constructing a cross-modal threat field function based on a lightweight Transform, and realizing quantitative modeling of threats in a complex environment; a federal reinforcement learning framework is utilized to fuse local decision and encrypted global strategy update, and cloud processing delay is avoided while data privacy is guaranteed; a mixed potential field path is generated in combination with the threat field gradient and strategy parameters, and the limitation of traditional path planning on dynamic threat avoidance is broken through; and finally, a multi-robot surrounding track is generated through a distributed cooperation protocol, so that the high-risk target surrounding efficiency and the system robustness are remarkably improved. According to the method, perception-decision-execution full-link optimization is realized, and a high-response, strong-privacy and collaborative intelligent solution is provided for a complex security scene.

Description

Technical Field

[0001] This invention relates to the field of security robot technology, specifically to a control method, system, device, and medium for a security robot. Background Technology

[0002] Security robots, as a core component of modern intelligent security systems, primarily achieve dynamic responses to complex environments through multimodal sensor fusion, real-time decision-making, and collaborative control. These systems are typically deployed in urban security and industrial inspection scenarios, requiring continuous operation under adverse conditions such as smoke, low light, and crowd obstruction. Their core control processes involve environmental perception, threat assessment, path planning, and multi-robot collaboration, aiming to quickly identify potential threats and generate optimal action strategies. Current mainstream methods rely on centralized data processing architectures, constructing environmental models by fusing data from visible light, infrared, and radar sensors, and generating control commands based on preset rules or traditional machine learning algorithms.

[0003] However, existing technologies have significant limitations: at the multimodal perception level, static weight allocation mechanisms struggle to adapt to dynamic environmental changes, leading to a sharp drop in system robustness when visual sensors fail in smoke-filled scenarios; at the decision-making level, centralized cloud processing cannot meet millisecond-level response requirements due to data transmission delays, and the transmission of raw video streams poses a privacy risk; path planning methods lack quantitative modeling of real-time threat fields, causing robots to frequently get stuck in locally dangerous areas; in multi-robot collaboration, fixed communication protocols are susceptible to interference, and the lack of cross-agency security knowledge sharing mechanisms significantly reduces encirclement efficiency. These shortcomings collectively result in systemic bottlenecks for security robots in complex scenarios, including response delays, obstacle avoidance failures, and low collaborative efficiency. Summary of the Invention

[0004] Based on this, the purpose of this invention is to provide a control method, system, device, and medium for a security robot that can dynamically optimize multi-source perception weights, ensure decision-making privacy and security, accurately quantify environmental threats, and achieve efficient collaborative capture.

[0005] The objective of this invention is achieved through the following solution:

[0006] In a first aspect, the present invention provides a control method for a security robot, comprising the following steps:

[0007] S1: Obtain the raw data stream collected by the multimodal sensor array of the security robot, dynamically weight and align the raw data stream in time and space, dynamically adjust the weight of each sensor by environmental factors in the raw data stream and synchronize the timestamps of multi-source heterogeneous data to generate a multimodal perception tensor.

[0008] S2: Perform cross-modal feature fusion processing on the multimodal sensing tensor. Use a lightweight Transformer model to fuse multi-source features in the multimodal sensing tensor and construct a threat field function to generate threat feature vectors and threat field functions.

[0009] S3: Perform federated reinforcement learning decision-making on threat feature vectors, calculate action instructions through local deep double-Q network and update the policy parameters of global policy through security aggregation, bind the updated policy parameters with encryption keys to generate federated communication protocol, and generate federated decision set containing policy parameters and encryption communication protocol.

[0010] S4: Based on the threat field function, perform hybrid potential field path planning on the policy parameters of the federated decision set, generate an executable path by constructing an adaptive repulsive force field and searching for the optimal path in the robot's motion space;

[0011] S5: Generate a collaborative encirclement path using the encrypted communication protocol of the federated decision set. Through distributed optimization calculation of the multi-robot encirclement trajectory of high-risk target locations detected in the executable path, a collaborative encirclement path is generated. The collaborative encirclement path is used to indicate the collaborative encirclement behavior of security robots.

[0012] In one embodiment, S1 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0013] S11: Perform environmental factor analysis on the raw data stream, extract smoke concentration and light intensity parameters, and generate a dynamic environmental factor set;

[0014] S12: Perform weight allocation processing on the dynamic environmental factor set and each sensor, calculate the modal weight coefficients through Sigmoid function mapping, and generate sensor weight vectors;

[0015] S13: Perform spatiotemporal synchronization processing on the sensor weight vector and the original data stream to compensate for the frame rate difference between the camera and the radar and unify the time reference, generating a multimodal perception tensor.

[0016] In one embodiment, S2 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0017] S21: Perform hierarchical feature extraction processing on the multimodal sensing tensor, extract visual features of image data through convolutional neural network, extract spatial features of lidar data through point cloud processing network, and extract temporal features of sonar data through recurrent neural network to generate a feature vector set containing multimodal information.

[0018] S22: Perform cross-modal attention fusion processing on the multi-source feature vector set, calculate the correlation weights between visual, spatial and temporal modalities through a lightweight Transformer, and integrate the complementary feature information of each modality to generate a threat feature vector with a unified representation;

[0019] S23: Perform threat field modeling on the threat feature vectors, construct a spatial threat probability distribution model based on the distribution density and intensity characteristics of the feature vectors, and generate a continuously differentiable threat field function.

[0020] In one embodiment, the threat field function of the control method for a security robot provided by the present invention is expressed as follows:

[0021]

[0022] in, Let be a continuously differentiable two-dimensional threat field function, representing any position in two-dimensional space. The level of threat, The mixing weights for the k-th Gaussian component. Let be the mean vector of a Gaussian distribution, and represent the two-dimensional coordinates of the threat cluster centers. Let covariance matrix be the variance matrix. The attenuation coefficient is... Let L be the gradient L2 norm of the threat feature vector, representing the rate of change of threat intensity.

[0023] In one embodiment, S3 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0024] S31: Perform deep reinforcement learning decision processing on the threat feature vector, calculate the value function of each action through a local deep double-Q network and select the optimal action strategy to generate a local action instruction set;

[0025] S32: Perform federated security aggregation processing on the local action instruction set, use differential privacy technology to protect the policy parameter updates of each node, and achieve global policy fusion through secure multi-party computation to generate updated global policy parameters;

[0026] S33: Perform security protocol encapsulation on the updated global policy parameters, bind the policy parameters with homomorphic encryption keys and add integrity verification information, generate a tamper-proof encrypted communication protocol, and integrate the global policy parameters to generate a federated decision set.

[0027] In one embodiment, S4 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0028] S41: Adaptive repulsive field modeling is performed on the threat field function. The repulsive field function is constructed based on the gradient distribution characteristics of the threat field and the kinematic constraints of the robot to generate an environment adaptive repulsive field model.

[0029] S42: Optimize the potential field parameters of the strategy parameters in the federated decision set, combine the environmental adaptive repulsive field model to analyze the environmental cognitive information contained in the strategy parameters and dynamically adjust the intensity of the potential field to generate a personalized hybrid potential field configuration.

[0030] S43: Perform optimal path search processing on personalized hybrid potential field configuration, and use the fast exploration random tree algorithm to explore the robot's feasible motion trajectory under the constraints of the hybrid potential field, and generate a safe and executable optimal path.

[0031] In one embodiment, S5 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0032] S51: Perform secure parsing on the encrypted communication protocols in the federated decision set, decrypt the protocol content using the bound homomorphic encryption key and verify its integrity, and generate a usable set of cooperative control parameters;

[0033] S52: Perform high-risk target trajectory extrapolation on executable paths, analyze the historical movement patterns of targets in the path and infer their future position sequences, and generate target trajectory predictions.

[0034] S53: Perform distributed cooperative encirclement planning processing on the target motion trajectory prediction, combine the cooperative control parameter set to calculate the motion trajectory of multiple robots forming an encirclement formation, and generate a spatiotemporally synchronized cooperative encirclement path.

[0035] Secondly, the present invention provides a control system for a security robot, which is configured with the following modules:

[0036] The multimodal spatiotemporal alignment module is used to acquire the raw data stream collected by the multimodal sensor array of the security robot, dynamically weight and align the raw data stream in a spatiotemporal manner, dynamically adjust the weight of each sensor by environmental factors in the raw data stream and synchronize the timestamps of multi-source heterogeneous data to generate a multimodal perception tensor.

[0037] The cross-modal feature fusion module is used to perform cross-modal feature fusion processing on the multimodal perception tensor. It fuses multi-source features in the multimodal perception tensor through a lightweight Transformer model and constructs a threat field function to generate threat feature vectors and threat field functions.

[0038] The federated reinforcement learning module is used to make federated reinforcement learning decisions on threat feature vectors. It calculates action instructions through a local deep double-Q network and updates the policy parameters of the global policy through security aggregation. The updated policy parameters are bound with the encryption key to generate a federated communication protocol, and a federated decision set containing policy parameters and encryption communication protocol is generated.

[0039] The hybrid potential field path planning module is used to perform hybrid potential field path planning based on the policy parameters of the federated decision set according to the threat field function. It generates an executable path by constructing an adaptive repulsive force field and searching for the optimal path in the robot's motion space.

[0040] The collaborative encirclement path generation module is used to generate collaborative encirclement paths based on the encrypted communication protocol of the federated decision set. It generates collaborative encirclement paths by calculating the multi-robot encirclement trajectory of high-risk target locations detected in the executable path through distributed optimization. The collaborative encirclement path is used to instruct the collaborative encirclement behavior of security robots.

[0041] Thirdly, this application provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement any of the above-mentioned control methods for security robots.

[0042] Fourthly, this application provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements any of the above-described control methods for a security robot.

[0043] In summary, this application provides a security robot that achieves intelligent collaborative security in complex scenarios through a closed-loop end-to-end technology. At the perception level, the dynamic weighted spatiotemporal alignment mechanism adaptively adjusts sensor weights based on environmental factors such as smoke concentration and light intensity, addressing the perception failure problem of traditional static fusion under adverse conditions and significantly improving the spatiotemporal consistency of multi-source heterogeneous data. The threat field function constructed by fusing features from multiple sources such as visible light, infrared, and radar using a cross-modal Transformer model enables accurate quantitative modeling of dynamic threats, effectively overcoming the risk of missed detection of concealed targets by traditional methods. At the decision-making level, the federated reinforcement learning framework generates action instructions through a local deep double-Q network and updates the global policy using differential privacy and secure multi-party computation. This avoids the risk of original data leakage while solving the latency bottleneck of cloud-based decision-making, ensuring millisecond-level real-time response capabilities. In the path planning stage, a hybrid potential field model dynamically couples the threat field gradient with a repulsive field driven by policy parameters, enabling the robot to autonomously adjust obstacle avoidance intensity based on the environmental threat situation. This overcomes the conservative limitations of traditional artificial potential field methods in dynamic obstacle avoidance, generating an optimal path that balances safety and efficiency. Ultimately, the distributed collaborative optimization mechanism based on homomorphic encryption protocol can realize the spatiotemporal synchronous encirclement and capture of high-risk targets by multiple robots. By collaboratively generating the encirclement trajectory, the interception success rate of moving targets and the overall robustness of the system are greatly improved.

[0044] To better understand and implement this invention, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0045] Figure 1 A flowchart illustrating a control method for a security robot provided in an embodiment of this application;

[0046] Figure 2 A schematic diagram of the process for generating a continuously differentiable threat field function provided in an embodiment of this application;

[0047] Figure 3 This is a schematic diagram of the structure of a control device for a security robot, provided in another embodiment of this application. Detailed Implementation

[0048] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0049] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0050] In one embodiment, such as Figure 1 As shown, a control method for a security robot is provided. This embodiment illustrates the method applied to a terminal. It is understood that this method can also be applied to a server, and further to a system including both a terminal and a server, and is implemented through interaction between the terminal and the server. In this embodiment, the method includes the following steps:

[0051] S1: Obtain the raw data stream collected by the multimodal sensor array of the security robot, dynamically weight and align the raw data stream in time and space, dynamically adjust the weight of each sensor by environmental factors in the raw data stream and synchronize the timestamps of multi-source heterogeneous data to generate a multimodal perception tensor.

[0052] Specifically, the system is equipped with a multimodal sensor array. Various sensors are deployed at different locations on the robot according to a pre-defined layout and connect to the robot's main control unit via a designated bus to achieve high-speed transmission of raw data streams. The raw data streams collected by the sensor array cover various types, including image data, point cloud data, distance data, and time-series data. The system performs preprocessing operations on the collected raw data streams. For image data, the system can use filtering algorithms to remove noise and histogram equalization technology to enhance image details. For point cloud data, the system can use statistical filtering methods to remove outliers and downsampling technology to reduce the data volume. For time-series data, the system uses pre-defined criteria to identify abnormal data points and uses interpolation to supplement missing values.

[0053] Furthermore, the system defines and quantitatively models multiple core environmental factors, including factors related to light intensity, smoke concentration, occlusion rate, and target movement speed. Normalized coefficients for each factor are obtained by mapping or calculating parameters collected by corresponding sensors. The system constructs a sensor weight adjustment model. The weight of each sensor is obtained by multiplying its inherent weight by the environmental influence function and then dividing by the sum of all corresponding product functions. The environmental influence function is designed for different sensor characteristics and calculates the impact of various environmental factors. The system employs a spatiotemporal alignment mechanism to process multi-source heterogeneous data. Timestamp synchronization is based on the high-precision clock of the main control unit, and interpolation is used to calibrate the time of sensor data at different sampling rates, ensuring consistency of data across time dimensions. Spatial alignment uses a hand-eye calibration method to establish a transformation matrix from the local coordinate system of each sensor to the robot's body coordinate system, uniformly mapping the coordinate data collected by different sensors to the body coordinate system.

[0054] The system organizes the aligned multi-source data into a multimodal perception tensor with a four-dimensional structure of time-space-modality-feature. The time dimension is determined based on a sliding time window, the spatial dimension corresponds to the three-dimensional coordinate axes of the vehicle coordinate system, the modal dimension corresponds to various types of sensors, and the feature dimension is obtained by dimensionality reduction of the basic features extracted from each type of sensor. The tensor is stored using a specified data type and storage efficiency is optimized through a tensor compression algorithm.

[0055] S2: Perform cross-modal feature fusion processing on the multimodal sensing tensor. Use a lightweight Transformer model to fuse multi-source features in the multimodal sensing tensor and construct a threat field function to generate threat feature vectors and threat field functions.

[0056] Specifically, the system employs a lightweight Transformer model to achieve cross-modal feature fusion. Dedicated encoders are designed for different modalities of data. Image data extracts feature maps through a pre-defined network structure; point cloud data extracts features by dividing it into units and performing convolutional operations; and temporal data extracts temporal features through a combination of one-dimensional convolution and a recurrent neural network. The system maps the features extracted from each modality to a unified dimension through a modal projection layer, generating a modal feature matrix. A local window attention mechanism reduces computational load, implementing attention calculations through query, key, and value matrix operations. A multi-head attention mechanism further enhances the feature fusion effect. The fusion output layer processes features through a multi-layer Transformer encoder, with each encoder layer containing a multi-head attention module and a feedforward network, outputting a fused feature map.

[0057] Furthermore, the system constructs a threat field function, which employs a potential energy superposition model and is defined as the threat potential energy at any point in the robot's motion space, obtained by superimposing an attractive potential field and a repulsive potential field. The attractive potential field is generated by high-risk targets and calculated based on the target's three-dimensional coordinates, threat level, and potential energy attenuation coefficient. The target's three-dimensional coordinates are identified through a target detection algorithm using fused features, and the threat level is quantified based on the target's movement speed, carried item characteristics, and behavioral patterns. The repulsive potential field is jointly generated by obstacles and high-risk areas. When the distance between a spatial point and the center of an obstacle or high-risk area is within the influence radius, the repulsive potential field value is calculated using an adaptive repulsion coefficient related to distance. When the distance exceeds the influence radius, the repulsive potential field value is zero. Obstacles are detected by designated sensors, and high-risk areas are filtered using attractive potential field values. The influence radius is dynamically adjusted based on the obstacle's volume, and the adaptive repulsion coefficient is calculated by combining the target threat level and the robot's current speed. The system samples within the robot's perception range at preset step lengths, extracting the potential energy value of each sampling point. These potential energy values ​​are then concatenated with the target's core features, including position, speed, and threat level, to generate a threat feature vector.

[0058] S3: Perform federated reinforcement learning decision-making on threat feature vectors, calculate action instructions through a local deep double-Q network and update the policy parameters of the global policy through security aggregation, bind the updated policy parameters with the encryption key to generate a federated communication protocol, and generate a federated decision set containing policy parameters and encryption communication protocol.

[0059] Specifically, each security robot in the system acts as a federated learning client, deploying a local deep double-Q network. The input layer of this network receives threat feature vectors, the hidden layers consist of multiple fully connected layers, and the features are transformed using activation functions. The output layer discretizes the action space into multiple actions and outputs the Q-value of each action. The reinforcement learning reward function designed by the system integrates threat response, obstacle avoidance performance, and collaborative requirements. Threat response rewards are calculated based on changes in the distance between the robot and the target; a decrease in distance corresponds to a positive reward. Obstacle avoidance rewards are determined based on whether a collision occurs; a positive reward is given when no collision occurs, and a negative reward is given when a collision occurs. Collaborative rewards are calculated based on the deviation between the actual distance between the robot and its collaborating partner and the optimal collaborative distance.

[0060] Preferably, the system can adopt a federated learning server-side and client-side architecture to implement global policy updates. The client trains a deep double-Q network based on local data to obtain local model parameters. The client encrypts the local model parameters using a homomorphic encryption algorithm, adds noise, and uploads it to the server through a secure channel, with the noise intensity meeting differential privacy requirements. The server receives the encrypted parameters from all clients, performs global parameter aggregation calculation, and decrypts them to obtain global policy parameters. The server encrypts the global policy parameters and sends them to each client, and the clients update their local model parameters.

[0061] The system binds global policy parameters with encryption keys to construct a federated communication protocol. The protocol frame structure comprises several components: the frame header includes the protocol version number, data type identifier, and reserved fields; the key hash is the hash value of the encryption key, used for client key verification; the parameter length field indicates the length of the encrypted parameter, which is the encrypted and encoded global policy parameter; and the checksum field is calculated using a specified algorithm to ensure data integrity. Finally, the system generates a federated decision set, which contains the encrypted global policy parameters and the federated communication protocol specification document.

[0062] S4: Based on the threat field function, perform hybrid potential field path planning on the policy parameters of the federated decision set. Generate an executable path by constructing an adaptive repulsive force field and searching for the optimal path in the robot's motion space.

[0063] Specifically, the system employs an improved A algorithm to achieve hybrid potential field path planning based on the threat field function and global policy parameters. The path starting point is obtained through GPS and IMU fusion positioning, and the target point is decoded from the global policy parameters, selecting the target location with the highest threat level from the threat feature vector. The system rasterizes the robot's movement range to construct a raster map. The potential field value of each raster is calculated by superimposing an attractive potential field and a repulsive potential field. The attractive potential field represents the gravitational potential of the target point on the raster, while the repulsive potential field represents the repulsive potential within the raster. The system uses the improved A algorithm for path search, and the designed heuristic function balances path length and safety, combining the distance from the raster to the target point with the raster potential field value in the calculation.

[0064] After obtaining the path, the system prunes the path, removing redundant inflection points and retaining key turning points. Then, it uses curve fitting to smooth the path, adhering to preset constraints to ensure smooth robot movement. The generated executable path is output as a sequence of path points, each containing three-dimensional coordinates and a timestamp. The timestamp is calculated based on the initial time, the path point interval, and the preset speed of the corresponding road segment. The preset speed is determined by global strategy parameters.

[0065] S5: Generate a collaborative encirclement path using the encrypted communication protocol of the federated decision set. Through distributed optimization calculation of the multi-robot encirclement trajectory of high-risk target locations detected in the executable path, a collaborative encirclement path is generated. The collaborative encirclement path is used to indicate the collaborative encirclement behavior of security robots.

[0066] Specifically, the system aggregates target detection results from all robots through a security aggregation protocol, selecting targets whose threat levels meet preset criteria as capture targets. Each robot converts the target coordinates in its local coordinate system to the global GPS coordinate system. This coordinate conversion is based on the robot's GPS positioning information and vehicle coordinate system calibration parameters, ensuring accurate position synchronization. The system employs a distributed optimization algorithm to optimize the allocation of capture positions among multiple robots. The constructed objective function integrates the robot's movement distance from its current position to the capture position with the capture error. The capture error is the deviation between the center of the capture circle and the target position. The objective function solution process follows the constraints of the minimum safe distance between robots and the capture radius, which is related to the target's movement speed.

[0067] Each robot generates a local trajectory based on its assigned encirclement position using a model predictive control algorithm. The prediction and control time domains are set according to preset parameters, with the optimization objectives being minimum trajectory tracking error and smooth velocity. Robots share trajectory information in real time via a federated communication protocol, adjusting their local trajectories using a distributed gradient descent method and iteratively optimizing position parameters using a consensus algorithm. During iteration, a step size parameter is introduced, and adjustments are made with reference to the position information of the communication neighbor set. The system generates a cooperative encirclement path for each robot, containing commands such as trajectory point sequences, movement speeds, and turning angles, ensuring all robots arrive at the encirclement position synchronously to form a closed encirclement. The cooperative encirclement path is output to the robot motion control unit in a standardized command format. The control unit tracks the path using a control algorithm to achieve precise cooperative encirclement.

[0068] In summary, this application provides a security robot that achieves intelligent collaborative security in complex scenarios through a closed-loop end-to-end technology. At the perception level, the dynamic weighted spatiotemporal alignment mechanism adaptively adjusts sensor weights based on environmental factors such as smoke concentration and light intensity, addressing the perception failure problem of traditional static fusion under adverse conditions and significantly improving the spatiotemporal consistency of multi-source heterogeneous data. The threat field function constructed by fusing features from multiple sources such as visible light, infrared, and radar using a cross-modal Transformer model enables accurate quantitative modeling of dynamic threats, effectively overcoming the risk of missed detection of concealed targets by traditional methods. At the decision-making level, the federated reinforcement learning framework generates action instructions through a local deep double-Q network and updates the global policy using differential privacy and secure multi-party computation. This avoids the risk of original data leakage while solving the latency bottleneck of cloud-based decision-making, ensuring millisecond-level real-time response capabilities. In the path planning stage, a hybrid potential field model dynamically couples the threat field gradient with a repulsive field driven by policy parameters, enabling the robot to autonomously adjust obstacle avoidance intensity based on the environmental threat situation. This overcomes the conservative limitations of traditional artificial potential field methods in dynamic obstacle avoidance, generating an optimal path that balances safety and efficiency. Ultimately, the distributed collaborative optimization mechanism based on homomorphic encryption protocol can realize the spatiotemporal synchronous encirclement and capture of high-risk targets by multiple robots. By collaboratively generating the encirclement trajectory, the interception success rate of moving targets and the overall robustness of the system are greatly improved.

[0069] In one embodiment, S1 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0070] S11: Perform environmental factor analysis on the raw data stream, extract smoke concentration and light intensity parameters, and generate a dynamic environmental factor set.

[0071] Specifically, the system receives raw data streams collected by a multimodal sensor array, which contain raw signals output by various sensors. The system performs environmental factor analysis on the raw data stream, following a pre-defined data processing flow. The system extracts signal components related to smoke concentration from the raw data stream using signal separation technology. These signal components are generated by gas sensors. The system amplifies and filters the extracted signal components to remove external electromagnetic interference and inherent sensor noise. The smoke concentration parameter is then obtained by converting the signal amplitude into environmental parameters.

[0072] Furthermore, the system extracts light intensity-related information from image data acquired by a visible light camera. By analyzing the grayscale distribution characteristics of image pixels, it calculates the overall grayscale mean and variance of the image. Combined with the exposure parameters at the time of image acquisition, it obtains the light intensity parameters through a preset transformation model. The system organizes the extracted smoke concentration and light intensity parameters according to a time series. Each time point corresponds to a complete set of smoke concentration and light intensity parameters, forming a dynamic environmental factor set. This dynamic environmental factor set is updated in real time as it changes, ensuring that it reflects the actual state of the current environment.

[0073] S12: Perform weight allocation processing on the dynamic environmental factor set and each sensor, calculate the modal weight coefficients through Sigmoid function mapping, and generate sensor weight vectors.

[0074] Specifically, the system acquires a dynamic set of environmental factors and the inherent attribute information of each sensor. This inherent attribute information includes the sensor's detection range, response characteristics, and environmental adaptability. Based on smoke concentration and light intensity parameters from the dynamic environmental factor set, combined with the inherent attribute information of each sensor, the system determines the effective output capability of each sensor in the current environment. The system uses the effective output capability parameters of each sensor as input, substitutes them into the Sigmoid function for mapping calculation, and the input variable of the Sigmoid function is the normalized result of the effective output capability parameters. Through function calculation, the system obtains the weight coefficients corresponding to each sensor. The value range of the weight coefficients is within a preset interval to ensure that the weight proportion of each sensor is reasonable. The system arranges the weight coefficients corresponding to all sensors sequentially according to a preset sensor sorting rule, forming a sensor weight vector. Each element in the sensor weight vector corresponds one-to-one with the corresponding sensor, intuitively reflecting the data contribution weight of each sensor in the current environment.

[0075] S13: Perform spatiotemporal synchronization processing on the sensor weight vector and the original data stream to compensate for the frame rate difference between the camera and the radar and unify the time reference, generating a multimodal perception tensor.

[0076] Specifically, the system invokes the sensor weight vector and the preprocessed raw data stream to initiate the spatiotemporal synchronization processing flow. To compensate for the frame rate difference between the camera and radar, the system employs time interpolation technology. Based on the difference in their sampling frequencies, supplementary data calculated from adjacent data points is inserted into the sensor data sequence with the lower frame rate, ensuring that the data output frequencies of the two types of sensors remain consistent. The system selects the high-precision clock of the robot's main control unit as a unified time reference, reads the raw timestamps of each sensor's data, and converts all sensor data timestamps into time identifiers under this unified time reference, achieving synchronization alignment in the time dimension.

[0077] In the spatial dimension, the system transforms the spatial coordinate data collected by different sensors into a unified vehicle coordinate system based on preset sensor calibration parameters, ensuring the consistency of the data from each sensor in spatial location. The system fuses the spatiotemporally synchronized raw data stream with the sensor weight vectors, organizing the data according to a four-dimensional structure of time-space-modality-feature. The time dimension corresponds to the synchronized time series, the spatial dimension corresponds to the three-dimensional coordinate axes of the vehicle coordinate system, the modality dimension corresponds to each sensor type, and the feature dimension is the result of feature extraction from each sensor data. The elements in the weight vector are used as weight coefficients for the corresponding modal data and are incorporated into the feature dimension calculation, ultimately generating a multimodal perception tensor. This tensor fully preserves the spatiotemporal information and weight proportions of each sensor data, providing basic data support for subsequent feature fusion.

[0078] In one embodiment, such as Figure 2 As shown, step S2 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0079] S21: Perform hierarchical feature extraction processing on the multimodal sensing tensor, extract visual features of image data through convolutional neural network, extract spatial features of lidar data through point cloud processing network, and extract temporal features of sonar data through recurrent neural network to generate a feature vector set containing multimodal information.

[0080] Specifically, the system receives a multimodal sensing tensor and initiates a hierarchical feature extraction process. For the image data within the tensor, the system performs feature extraction using a convolutional neural network, and the output of the convolutional layer is calculated using the following formula:

[0081]

[0082] in, For the extracted visual features, For convolution kernel parameters, For input image data, This represents the convolution operation. For bias terms, The activation function is used. For LiDAR data, the system processes it using a point cloud processing network, extracting spatial features using the following formula:

[0083]

[0084] in, For spatial features, This is the network weight matrix. For point cloud data matrix, Here, `maxpool` represents the maximum pooling operation, acting as a bias parameter. For sonar data, the system extracts temporal features using a recurrent neural network and calculates the following formula:

[0085]

[0086] in, For the time series characteristics at the current moment, This is the hidden state from the previous moment. and These are the weights for the hidden layer and the input layer, respectively. The sonar data is at time t. For the circular layer bias, The activation function is used. The system organizes visual features, spatial features, and temporal features according to the time dimension and modality dimension, ensuring the dimensional consistency and temporal synchronization of each feature, forming a feature vector set containing multimodal information. This vector set completely retains the core feature information of various sensor data.

[0087] S22: Perform cross-modal attention fusion processing on the multi-source feature vector set, calculate the correlation weights between visual, spatial and temporal modalities through a lightweight Transformer, and integrate the complementary feature information of each modality to generate a threat feature vector with a unified representation.

[0088] Specifically, the system acquires a multi-source feature vector set and initiates cross-modal attention fusion processing. The system calculates the correlation weights between each modality using a lightweight Transformer and employs a scaled dot product attention mechanism. The weight calculation formula is as follows:

[0089]

[0090] in, The query matrix is ​​obtained by visual feature transformation. The key matrix is ​​obtained by spatial feature transformation. The value matrix is ​​obtained by transforming time series features. To query the dimensions of the matrix and the key matrix, This represents matrix transpose multiplication. is the scaling factor, and softmax is the normalization function. The system uses this formula to obtain the correlation weights between each modality, with the weight values ​​reflecting the complementarity of different modal features. The system multiplies the correlation weights by the feature vectors of each modality, sums them element-wise, and integrates them. The integration process follows the feature dimension alignment principle to ensure that the features of each modality are fused in a unified dimensional space. Through the above processing, the system complementaryly integrates the detailed information of visual features, the positional information of spatial features, and the dynamic information of temporal features, eliminating information redundancy between modalities and generating a threat feature vector with a unified representation. This vector contains complete threat-related information after multimodal fusion, providing a foundation for subsequent threat field modeling.

[0091] S23: Perform threat field modeling on the threat feature vectors, construct a spatial threat probability distribution model based on the distribution density and intensity characteristics of the feature vectors, and generate a continuously differentiable threat field function.

[0092] Specifically, the system receives threat feature vectors and performs threat field modeling. First, the system analyzes the distribution density and intensity characteristics of the threat feature vectors. The distribution density is determined by the clustering of the feature vectors in high-dimensional space, and the intensity is represented by the magnitude of the feature vector amplitude. Based on the analysis results, the system constructs a spatial threat probability distribution model, employing a Gaussian mixture model combined with a threat intensity attenuation mechanism to generate a continuously differentiable threat field function. The expression for the threat field function is:

[0093]

[0094] in, Let be a continuously differentiable two-dimensional threat field function, representing any position in two-dimensional space. The level of threat, The mixing weights for the k-th Gaussian component. Let be the mean vector of a Gaussian distribution, and represent the two-dimensional coordinates of the threat cluster centers. Let covariance matrix be the variance matrix. The attenuation coefficient is... Let L2 be the gradient norm of the threat feature vector, representing the rate of change of threat intensity. The system uses this function to map the high-dimensional information of the threat feature vector to a two-dimensional space, forming a continuously distributed threat field. The magnitude of the threat field directly corresponds to the threat level at a spatial location.

[0095] In one embodiment, S3 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0096] S31: Perform deep reinforcement learning decision processing on the threat feature vector, calculate the value function of each action through a local deep double-Q network and select the optimal action strategy to generate a local action instruction set.

[0097] Specifically, the system receives a unified threat feature vector and invokes a locally deployed deep double-Q network. This network includes an evaluation network and a target network, both with identical structures and using the threat feature vector as input state s. The system calculates the value function corresponding to each action 'a' through the evaluation network. The value function formula is:

[0098]

[0099] in, For the value of the action, To evaluate network parameters, and These are the weight matrices from the input layer to the hidden layer and from the hidden layer to the output layer, respectively. and For the corresponding layer bias, Let be the activation function, and s be the state represented by the threat feature vector. These are actions in a discrete action space. The system generates a target value function through a target network. , The target network parameters are periodically updated from the evaluation network. The system employs a greedy strategy to select the optimal action, that is, to select the action that maximizes the action's value. The motion space encompasses core control dimensions such as robot speed and turning angle. The system organizes the optimal motion strategy into a structured set of local motion instructions based on time sequence and decision-making scenario. This set includes corresponding motion instructions for different states, ensuring the consistency of instruction timing and the feasibility of execution.

[0100] S32: Performs federated security aggregation processing on the local action instruction set, uses differential privacy technology to protect the policy parameter updates of each node, and achieves global policy fusion through secure multi-party computation to generate updated global policy parameters.

[0101] Specifically, the system collects the set of local action commands uploaded by each distributed node and extracts the corresponding local policy parameters from the set. The federated security aggregation process is initiated. The system employs differential privacy technology to protect the privacy of policy parameters at each node. Privacy protection is achieved by adding noise to the local policy parameters; the noise addition formula is as follows:

[0102]

[0103] in, The parameters for the privacy protection strategy after adding noise. These are the original local policy parameters. For parameters related to privacy budget, For noise variables that conform to a preset distribution, the system achieves global policy fusion through secure multi-party computation technology. Each node only uploads encrypted parameter information and does not disclose the original data. The global policy fusion formula is:

[0104]

[0105] in, For the updated global policy parameters, The number of nodes participating in federated learning. The weight coefficients for each node are determined by the node's data quality and participation, ensuring that the sum of all node weight coefficients equals 1. Through this fusion process, the system integrates the local decision-making experience of each node, eliminates the influence of single-node data bias, and generates updated global policy parameters with global optimality.

[0106] S33: Perform security protocol encapsulation on the updated global policy parameters, bind the policy parameters with homomorphic encryption keys and add integrity verification information, generate a tamper-proof encrypted communication protocol, and integrate the global policy parameters to generate a federated decision set.

[0107] Specifically, the system obtains the updated global policy parameters, generates an encryption key K using a homomorphic encryption algorithm, and binds the global policy parameters to the encryption key using the following binding formula:

[0108]

[0109] in, For encrypted global policy parameters, It is a homomorphic encryption algorithm. For the generated homomorphic encryption key, this binding process ensures that only the node holding the corresponding decryption key can parse the policy parameters. The system performs integrity verification on the bound encryption parameters and key, calculating the verification information using a hash algorithm. The verification information formula is:

[0110]

[0111] in, For hash value, For the preset hash algorithm, This represents a string concatenation operation, and the verification information is used by the receiving end to check whether the data has been tampered with. The system encapsulates the encrypted global policy parameters, homomorphic encryption keys, and integrity verification information according to a preset protocol format to generate a tamper-proof encrypted communication protocol. The system integrates the updated global policy parameters and the encrypted communication protocol to form a federated decision set. This set contains both the globally optimal decision-making basis and ensures the security and integrity of the transmission and storage process through encryption and verification mechanisms.

[0112] In one embodiment, S4 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0113] S41: Perform adaptive repulsive field modeling on the threat field function. Construct the repulsive field function based on the gradient distribution characteristics of the threat field and the kinematic constraints of the robot to generate an environment-adaptive repulsive field model.

[0114] Specifically, the system receives a continuously differentiable threat field function. By solving for the gradient of the threat field function, it obtains the gradient distribution characteristics of the threat field. The gradient distribution characteristics intuitively reflect the changing trend of the threat level at various locations in space; a larger gradient magnitude indicates a more drastic change in threat. Simultaneously, the system obtains the robot's kinematic constraints. These constraints cover the limitations on the robot's core motion parameters such as speed, turning angle, and acceleration, ensuring that the constructed repulsive field model matches the robot's actual motion capabilities. Preferably, the system constructs a repulsive field function based on the threat field gradient distribution characteristics and kinematic constraints, with the function expression as follows:

[0115]

[0116] in, The strength of the repulsive field. This is the repulsion adjustment coefficient. For the threat field in position gradient vector, Let be the magnitude of the gradient vector. As the kinematic constraint factor, For the robot's movement speed, For the robot's turning angle, the kinematic constraint factor transforms motion parameter restrictions into adjustments to the repulsive force intensity through a mapping relationship. The system uses this function to combine threat field information with the robot's motion capabilities, generating an environment-adaptive repulsive force field model. This model can dynamically adjust the repulsive force effect based on the spatial threat distribution and the robot's motion state.

[0117] S42: Optimize the potential field parameters of the strategy parameters in the federated decision set, combine the environmental cognitive information contained in the strategy parameters with the environmental adaptive repulsive field model, and dynamically adjust the strength of the potential field to generate a personalized hybrid potential field configuration.

[0118] Specifically, the system acquires global policy parameters from the federated decision set, initiates a potential field parameter optimization process based on the environmental adaptive repulsive field model, analyzes the global policy parameters, and extracts the environmental cognitive information contained within them. This environmental cognitive information includes the priority of responses to different types of threats, the distribution characteristics of environmental obstacles, and adaptation rules. Based on the analyzed environmental cognitive information, the system dynamically adjusts the intensity of the potential field and constructs a potential field intensity adjustment formula:

[0119]

[0120] in, The potential field intensity coefficient, For parameter mapping functions, For the updated global policy parameters, this formula transforms the environmental awareness information in the global policy parameters into potential field strength adjustment commands through a mapping relationship. The system adjusts the weight ratio of the repulsive and gravitational fields through intensity coefficients. The gravitational field is generated from the target point in path planning and is used to guide the robot to move in the target direction. The system integrates the adjusted repulsive and gravitational fields according to preset fusion rules to generate a personalized hybrid potential field configuration. This configuration can adapt to the current environmental characteristics and global decision-making needs, providing accurate potential field constraints for path search.

[0121] S43: Perform optimal path search processing on personalized hybrid potential field configuration, and use the fast exploration random tree algorithm to explore the robot's feasible motion trajectory under the constraints of the hybrid potential field, and generate a safe and executable optimal path.

[0122] Specifically, the system acquires a personalized hybrid potential field configuration and initiates the optimal path search process using a fast-exploration random tree algorithm. The system initializes a fast-exploration random tree structure, with the robot's current position as the root node and the target position specified by the global policy parameters as the path endpoint. During algorithm iteration, the system randomly samples candidate nodes within the robot's motion space and performs feasibility assessments on these candidate nodes using the personalized hybrid potential field configuration. This involves determining whether the repulsive field strength of the candidate node is within a preset safe range, while simultaneously ensuring that the motion state corresponding to the node conforms to the robot's kinematic constraints. The system constructs a path evaluation function:

[0123]

[0124] in, For path cost, As the starting point of the path, The endpoint of the path. For the gravitational field strength, The system uses a path element as an example. It iteratively expands a random tree, connecting candidate nodes that meet the constraints to form multiple feasible motion trajectories. Based on the path evaluation function, the cost of each trajectory is calculated, and the trajectory with the minimum cost is selected as the optimal path. The system then smooths the optimal path, eliminating redundant inflection points to ensure the path meets the robot's motion stability requirements, ultimately generating a safe and executable optimal path.

[0125] In one embodiment, S5 of the control method for a security robot provided by the present invention specifically includes the following steps:

[0126] S51: Perform secure parsing of the encrypted communication protocols in the federated decision set, decrypt the protocol content using the bound homomorphic encryption key and verify its integrity, and generate a usable set of cooperative control parameters.

[0127] Specifically, the system receives encrypted communication protocols from the federated decision set, calls the homomorphic encryption key bound to the protocol, performs a decryption operation to obtain the plaintext content of the protocol, and the decryption process follows a preset homomorphic decryption algorithm. The decryption formula is as follows:

[0128]

[0129] in, This is the plaintext of the decrypted protocol. For homomorphic decryption operators, For encrypted data in encrypted communication protocols, This is the bound homomorphic encryption key. After decryption, the system initiates an integrity verification process, comparing the calculated result of the decrypted plaintext with the integrity verification information carried in the protocol. The verification formula is:

[0130]

[0131] in, To verify the results, For comparison functions, To decrypt the plaintext, the hash value is calculated using a preset hash algorithm. This is the integrity verification hash value built into the encrypted communication protocol. Once the verification result confirms that the data has not been tampered with, the system extracts the relevant parameters for collaborative control from the decrypted plaintext. These parameters cover core aspects such as multi-robot communication timing rules, encirclement formation constraints, and target tracking accuracy requirements. The system categorizes and organizes these parameters according to their functions and usage scenarios, forming a well-structured and directly callable set of collaborative control parameters. This ensures that the parameters can effectively play a constraining and guiding role in subsequent collaborative encirclement planning.

[0132] S52: Perform high-risk target trajectory extrapolation processing on executable paths, analyze the historical movement patterns of targets in the path and infer their future position sequences, and generate target trajectory predictions.

[0133] Specifically, the system acquires the safe and executable optimal path, extracts historical motion data of high-risk targets within the path's coverage area, preprocesses the historical motion data, filters out valid data points and sorts them by time series, extracts key features such as position coordinates, motion speed, and motion direction, and analyzes the historical motion patterns presented by these features, including uniform motion characteristics, variable speed motion patterns, turning frequency, and angle distribution. Based on the historical motion patterns, the system constructs a trajectory prediction model and uses a time-series prediction method to infer the future position sequence of the target. The prediction formula is:

[0134]

[0135] in, Let m be the predicted location of the target at the m-th time point in the future. For trajectory prediction function, Given the historical position coordinates of the target over the past n+1 consecutive time points, The time interval between adjacent time nodes. For the current time point, To predict the number of time points, the system uses this formula, combined with the inherent patterns in historical movement models, to correct prediction biases and ensure that the predicted position of each future time point conforms to the target's movement characteristics. The system combines the predicted positions of all future time points in chronological order to generate a target movement trajectory prediction containing position coordinates and corresponding time identifiers, providing a dynamic basis for collaborative encirclement planning.

[0136] S53: Perform distributed cooperative encirclement planning processing on the target motion trajectory prediction, combine the cooperative control parameter set to calculate the motion trajectory of multiple robots forming an encirclement formation, and generate a spatiotemporally synchronized cooperative encirclement path.

[0137] Specifically, the system acquires the target motion trajectory prediction and the set of available cooperative control parameters, analyzes the cooperative control parameter set, and extracts key constraints such as encirclement formation parameters, inter-robot safety distance constraints, and spatiotemporal synchronization accuracy requirements as the basis for encirclement trajectory planning. The system constructs a trajectory generation model for multi-robot cooperative encirclement, and calculates the motion trajectory of each participating robot based on the target motion trajectory prediction. The trajectory generation formula is:

[0138]

[0139] in, For the first A robot in time The trajectory of movement, For distributed trajectory planning functions, Let be the initial position of the i-th robot. The predicted position of the target at time t. These are the constraint parameters in the collaborative control parameter set. The system coordinates the trajectory generation process of each robot through a distributed optimization algorithm, ensuring that the motion trajectories of all robots meet the encirclement formation requirements, maintain a safe distance between robots, and synchronize their movement rhythm with the target trajectory. The system performs spatiotemporal synchronization calibration on the motion trajectories of each robot, enabling all robots to reach the preset encirclement position at the same time point, forming a closed encirclement posture. The system integrates the synchronized trajectories of all participating robots to generate a spatiotemporally synchronized collaborative encirclement path. The path clearly defines the position, speed, and turning commands of each robot at different time points, directly instructing the collaborative encirclement behavior of the security robots.

[0140] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0141] Based on the same inventive concept, this application also provides a control device for a security robot to implement the control method for the security robot described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the control device for the security robot provided below can be found in the limitations of the control method for the security robot described above, and will not be repeated here.

[0142] Preferably, such as Figure 3 As shown, the present invention provides a control system 600 for a security robot, which is configured with the following modules:

[0143] The multimodal spatiotemporal alignment module 610 is used to acquire the raw data stream collected by the multimodal sensor array of the security robot, dynamically weight and align the raw data stream in a spatiotemporal manner, dynamically adjust the weight of each sensor through environmental factors in the raw data stream and synchronize the timestamps of multi-source heterogeneous data to generate a multimodal perception tensor.

[0144] The cross-modal feature fusion module 620 is used to perform cross-modal feature fusion processing on the multimodal perception tensor. It fuses multi-source features in the multimodal perception tensor through a lightweight Transformer model and constructs a threat field function to generate threat feature vectors and threat field functions.

[0145] The federated reinforcement learning module 630 is used to perform federated reinforcement learning decisions on threat feature vectors. It calculates action instructions through a local deep double-Q network and updates the policy parameters of the global policy through security aggregation. The updated policy parameters are bound with the encryption key to generate a federated communication protocol, and a federated decision set containing policy parameters and encryption communication protocol is generated.

[0146] The hybrid potential field path planning module 640 is used to perform hybrid potential field path planning based on the policy parameters of the federated decision set according to the threat field function. It generates an executable path by constructing an adaptive repulsive force field and searching for the optimal path in the robot's motion space.

[0147] The collaborative encirclement path generation module 650 is used to generate collaborative encirclement paths based on the encrypted communication protocol of the federated decision set. It generates collaborative encirclement paths by calculating the multi-robot encirclement trajectory of high-risk target locations detected in the executable path through distributed optimization. The collaborative encirclement paths are used to instruct the collaborative encirclement behavior of security robots.

[0148] Preferably, the multimodal spatiotemporal alignment module 610 provided in this application is configured with the following units:

[0149] The environmental factor analysis unit is used to perform environmental factor analysis on the raw data stream, extract smoke concentration and light intensity parameters, and generate a dynamic set of environmental factors.

[0150] The sensor weight allocation unit is used to perform weight allocation processing on the dynamic environmental factor set and each sensor. It calculates the weight coefficient of each mode through the Sigmoid function mapping and generates the sensor weight vector.

[0151] The spatiotemporal synchronization processing unit is used to perform spatiotemporal synchronization processing on the sensor weight vector and the raw data stream, compensate for the frame rate difference between the camera and the radar, unify the time reference, and generate a multimodal sensing tensor.

[0152] Preferably, the cross-modal feature fusion module 620 provided in this application is configured with the following units:

[0153] The hierarchical feature extraction unit is used to perform hierarchical feature extraction processing on the multimodal sensing tensor. It extracts visual features from image data through convolutional neural networks, spatial features from lidar data through point cloud processing networks, and temporal features from sonar data through recurrent neural networks, generating a feature vector set containing multimodal information.

[0154] The cross-modal attention fusion unit is used to perform cross-modal attention fusion processing on multi-source feature vector sets. It calculates the correlation weights between visual, spatial and temporal modalities through a lightweight Transformer, and integrates the complementary feature information of each modality to generate a threat feature vector with a unified representation.

[0155] The threat field modeling unit is used to perform threat field modeling on threat feature vectors. Based on the distribution density and intensity characteristics of the feature vectors, it constructs a spatial threat probability distribution model and generates a continuously differentiable threat field function.

[0156] Preferably, the federated reinforcement learning module 630 provided in this application is configured with the following units:

[0157] The local reinforcement decision unit is used to perform deep reinforcement learning decision processing on the threat feature vector. It calculates the value function of each action through a local deep double-Q network and selects the optimal action strategy to generate a set of local action instructions.

[0158] The Federation Security Aggregation Unit is used to perform federated security aggregation processing on the local action instruction set. It uses differential privacy technology to protect the policy parameter updates of each node and achieves global policy fusion through secure multi-party computation to generate updated global policy parameters.

[0159] The security protocol encapsulation unit is used to encapsulate the updated global policy parameters using a security protocol, bind the policy parameters to homomorphic encryption keys and add integrity verification information, generate a tamper-proof encrypted communication protocol, and integrate the global policy parameters to generate a federated decision set.

[0160] Preferably, the hybrid potential field path planning module 640 provided in this application is configured with the following units:

[0161] The adaptive repulsive field modeling unit is used to perform adaptive repulsive field modeling on the threat field function. It constructs the repulsive field function based on the gradient distribution characteristics of the threat field and the kinematic constraints of the robot, and generates an adaptive repulsive field model of the environment.

[0162] The potential field parameter optimization unit is used to optimize the potential field parameters of the strategy parameters in the federated decision set. It combines the environmental adaptive repulsive field model to analyze the environmental cognitive information contained in the strategy parameters and dynamically adjusts the intensity of the potential field to generate a personalized hybrid potential field configuration.

[0163] The optimal path search unit is used to perform optimal path search processing on personalized hybrid potential field configurations. It uses a fast exploration random tree algorithm to explore feasible motion trajectories of the robot under the constraints of the hybrid potential field and generate a safe and executable optimal path.

[0164] Preferably, the collaborative encirclement path generation module 650 provided in this application is configured with the following units:

[0165] The protocol security parsing unit is used to perform security parsing on the encrypted communication protocols in the federated decision set, decrypt the protocol content using the bound homomorphic encryption key and verify its integrity, and generate a set of usable cooperative control parameters.

[0166] The target trajectory extrapolation unit is used to extrapolate the motion trajectory of high-risk targets on executable paths, analyze the historical motion patterns of targets in the path and infer their future position sequences, and generate target motion trajectory predictions.

[0167] The collaborative encirclement planning unit is used to perform distributed collaborative encirclement planning processing on the predicted target motion trajectory. It combines the collaborative control parameter set to calculate the motion trajectory of multiple robots forming an encirclement formation and generates a spatiotemporally synchronized collaborative encirclement path.

[0168] In one embodiment, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described control method for the security robot.

[0169] In one embodiment, this application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the above-described control method for a security robot.

[0170] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of those different embodiments or examples.

[0171] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The components described as separate parts may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0172] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any person skilled in the art can easily conceive of various variations or substitutions within the technical scope disclosed in this application, and these should all be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A security robot control method, characterized by, The method comprises the following steps: S1: acquiring a raw data stream collected by a multi-modal sensor array of a security robot, performing dynamic weighted space-time alignment on the raw data stream, dynamically adjusting the weights of each sensor and synchronizing the timestamps of multi-source heterogeneous data through environmental factors in the raw data stream, and generating a multi-modal perception tensor; S2: performing cross-modal feature fusion processing on the multi-modal perception tensor, fusing multi-source features in the multi-modal perception tensor through a lightweight Transformer model, and constructing a threat field function to generate a threat feature vector and a threat field function; S3: performing federated reinforcement learning decision-making on the threat feature vector, calculating action instructions through a local deep double Q network, updating the policy parameters of the global policy through safe aggregation, binding the updated policy parameters with an encryption key to generate a federated communication protocol, and generating a federated decision set containing policy parameters and an encrypted communication protocol; S4: performing hybrid potential field path planning on the policy parameters of the federated decision set based on the threat field function, constructing an adaptive repulsive field, and searching for an optimal path in the robot motion space to generate an executable path; S5: performing cooperative hunting path generation on the encrypted communication protocol of the federated decision set, calculating a multi-robot surrounding trajectory of a high-risk target position detected in the executable path through distributed optimization, and generating a cooperative hunting path, which is used to indicate the cooperative hunting behavior of the security robot.

2. The method of claim 1, wherein, The S1 comprises: S11: performing environmental factor analysis processing on the raw data stream, extracting smoke concentration and light intensity parameters, and generating a dynamic environmental factor set; S12: performing weight distribution processing on the dynamic environmental factor set and each sensor, calculating each modal weight coefficient through a Sigmoid function mapping, and generating a sensor weight vector; S13: performing space-time synchronization processing on the sensor weight vector and the raw data stream, compensating for the frame rate difference between the camera and the radar, and unifying the time reference to generate a multi-modal perception tensor.

3. The method of claim 1, wherein, The S2 comprises: S21: performing hierarchical feature extraction processing on the multi-modal perception tensor, extracting visual features of image data through a convolutional neural network, extracting spatial features of laser radar data through a point cloud processing network, and extracting time sequence features of sonar data through a recurrent neural network to generate a feature vector set containing multi-modal information; S22: performing cross-modal attention fusion processing on the multi-source feature vector set, calculating the correlation degree weight between visual, spatial and time sequence modes through a lightweight Transformer, and integrating the complementary feature information of each mode to generate a threat feature vector with unified representation; S23: performing threat field modeling processing on the threat feature vector, constructing a spatial threat probability distribution model based on the distribution density and intensity features of the feature vector to generate a continuous and differentiable threat field function.

4. The method of claim 3, wherein, The expression of the threat field function is: wherein, is a continuous differentiable two-dimensional threat field function, representing the threat level at any position in a two-dimensional space, is a continuous differentiable two-dimensional threat field function, representing the threat level at any position in a two-dimensional space, is a mixing weight of the kth Gaussian component, is a mean vector of Gaussian distribution, representing the two-dimensional coordinates of the threat cluster center, is a covariance matrix, is a decay coefficient, is a gradient two-norm of the threat feature vector, representing the threat intensity change rate.

5. The method of claim 1, wherein, The S3 comprises: S31: performing deep reinforcement learning decision processing on the threat feature vector, calculating the value function of each action through a local deep double Q network and selecting the optimal action policy to generate a local action instruction set; S32: Perform federated security aggregation processing on the local action instruction set, protect the policy parameter update of each node using differential privacy technology, and realize global policy fusion through secure multi-party computation to generate updated global policy parameters; S33: Perform security protocol encapsulation processing on the updated global policy parameters, bind the policy parameters with homomorphic encryption keys and add integrity verification information to generate tamper-proof encrypted communication protocols, and integrate the global policy parameters to generate a federated decision set.

6. The method of claim 1, wherein, The S4 includes: S41: Perform adaptive repulsive force field modeling processing on the threat field function, construct a repulsive force field function according to the gradient distribution characteristics of the threat field and the kinematic constraints of the robot, and generate an environment-adaptive repulsive force field model; S42: Perform potential field parameter optimization on the policy parameters in the federated decision set, analyze the environmental awareness information contained in the policy parameters in combination with the environment-adaptive repulsive force field model, and dynamically adjust the action strength of the potential field to generate a personalized hybrid potential field configuration; S43: Perform optimal path search processing on the personalized hybrid potential field configuration, use the rapid exploration random tree algorithm to explore the feasible motion trajectory of the robot under the constraint of the hybrid potential field, and generate a safe and executable optimal path.

7. The method according to any one of claims 1 to 6, characterized in that, The S5 includes: S51: Perform security analysis processing on the encrypted communication protocol in the federated decision set, use the bound homomorphic encryption key to decrypt the protocol content and verify its integrity, and generate a set of available cooperative control parameters; S52: Perform high-risk target motion trajectory deduction processing on the executable path, analyze the historical motion pattern of the target in the path and infer its future position sequence to generate a target motion trajectory prediction; S53: Perform distributed cooperative hunting planning processing on the target motion trajectory prediction, calculate the motion trajectory of the multi-robot formation of the surrounding formation in combination with the cooperative control parameter set, and generate a time and space synchronized cooperative hunting path. 8.A control system of a security robot, characterized by, The system includes: A multi-modal spatio-temporal alignment module for obtaining raw data streams collected by a multi-modal sensor array of a security robot, performing dynamic weighted spatio-temporal alignment on the raw data streams, dynamically adjusting the weights of each sensor through environmental factors in the raw data streams and synchronizing the timestamps of multi-source heterogeneous data, and generating a multi-modal perception tensor; A cross-modal feature fusion module for performing cross-modal feature fusion processing on the multi-modal perception tensor, fusing multi-source features in the multi-modal perception tensor through a lightweight Transformer model and constructing a threat field function, and generating a threat feature vector and a threat field function; A federated reinforcement learning module for federated reinforcement learning decision-making on the threat feature vector, calculating action instructions through a local deep double Q network and updating policy parameters of a global policy through secure aggregation, binding the updated policy parameters with encryption keys to generate a federated communication protocol, and generating a federated decision set containing policy parameters and encrypted communication protocols; The mixed potential field path planning module is configured to perform mixed potential field path planning on the policy parameters of the federated decision set based on the threat field function, generate an executable path by constructing an adaptive repulsive field and searching for an optimal path in a robot motion space, and output the executable path. The cooperative hunting path generation module is configured to perform cooperative hunting path generation on the encrypted communication protocol of the federated decision set, generate a cooperative hunting path by performing distributed optimization on a multi-robot surrounding trajectory of a high-risk target position detected in the executable path, and output the cooperative hunting path, where the cooperative hunting path is used to indicate a cooperative hunting behavior of the security robots. 9.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-8 when the computer program is executed by the processor. The computer program is executed by the processor to implement the method in any one of claims 1 to 7.

10. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the method in any one of claims 1 to 7.

Images