Isolation belt maintenance operation robot system based on multi-AGENT cooperation

Abstract

The invention discloses an isolation belt maintenance operation robot system based on multi-AGENT cooperation, and the system comprises an environment sensing module which is used for generating an environment feature tensor and a global environment map; the state modeling module is used for generating a standardized state embedding vector; the task collaborative allocation module is used for performing task division and collaborative planning through the improved QMIX network; the path planning module is used for generating a path target parameter and a control instruction parameter through an improved TimeSform network; the maintenance operation execution module is used for executing an isolation belt maintenance operation task; the communication interaction module is used for exchanging the standardized state embedding vector and the job progress information of each AGENT through a wireless communication network; and the feedback regulation module is used for collecting energy consumption evaluation indexes of all AGENTs and updating task weights. According to the method, the cooperation efficiency, the energy consumption control capability and the path conflict avoidance capability of the multiple AGENTs in the complex operation area are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent traffic equipment, and in particular to a strip maintenance robot system based on multi-AGENT cooperation. BACKGROUND

[0002] With the continuous expansion of urban transportation infrastructure and the promotion of green and intelligent maintenance concepts, the demand for intelligent maintenance of linear green areas such as road strips is increasing. Existing strip maintenance operations mostly rely on manual driving or single robots to complete tasks such as spraying, pruning, and cleaning, which is low in efficiency and limited in coverage, and lacks an optimization mechanism for collaborative work between multiple robots, which generally has the following problems in actual application:

[0003] Maintenance robots have difficulty in reasonable task allocation and path coordination during multi-task parallel execution, often resulting in overlapping work areas, task conflicts, or blind areas; the path planning and motion control process of each robot often ignores the interaction between them, resulting in path congestion or uneven energy consumption in complex scenarios; traditional path prediction and control methods are mostly based on single AGENT decision models, lacking modeling and feedback mechanisms for group state, resulting in decreased overall work efficiency; existing methods often lack energy self-adaptive optimization mechanisms and cannot dynamically adjust task allocation weights based on historical energy consumption and execution performance, making it difficult to meet the needs of high-frequency, multi-cycle intelligent strip maintenance operations.

[0004] Therefore, how to provide a strip maintenance robot system based on multi-AGENT cooperation is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] One object of the present application is to provide a strip maintenance robot system based on multi-AGENT cooperation. The system uses an improved QMIX network and an improved TimeSformer network to achieve task collaborative allocation, path intelligent planning, and work conflict avoidance between multiple robots, and has closed-loop control capabilities from environment mapping, state perception, path prediction to energy self-adaptive adjustment. By introducing task graph convolution fusion, state coordination matrix, and dynamic task weight mechanism, the present application improves the energy sensitivity and response flexibility of task allocation; by constructing a global interaction attention tensor and combining control instruction parameters, the path prediction accuracy and obstacle avoidance robustness between multiple AGENTS are significantly enhanced. The system has the advantages of high work efficiency, reasonable task division, low energy consumption, and strong cooperation ability.

[0006] According to the strip maintenance robot system based on multi-AGENT cooperation of the present application, the following steps are included: An environment perception module is configured to collect environment data of the isolation belt working scene and generate an environment feature tensor and a global environment map. A state modeling module is configured to collect running state information of each AGENT and generate a standardized state embedding vector. A task coordination and distribution module is configured to input the global environment map and the standardized state embedding vector of each AGENT into an improved QMIX network to perform task division and collaborative planning and generate a task distribution result of each AGENT. A path planning module is configured to input the task distribution result and the environment feature tensor into an improved TimeSformer network to perform spatio-temporal attention modeling and cross-AGENT interaction attention modeling and generate a path target parameter and a control instruction parameter of each AGENT at each time step. A maintenance operation execution module is configured to control a motion trajectory and an operation range of an AGENT body according to the path target parameter and drive an operation execution unit to complete an isolation belt maintenance operation task according to the control instruction parameter. A communication interaction module is configured to exchange the standardized state embedding vector and operation progress information of each AGENT through a wireless communication network during task execution and feed back to the improved QMIX network to update the task distribution result. A feedback adjustment module is configured to collect an energy consumption evaluation index of each AGENT after task completion, calculate a unit path energy consumption prediction value, and update a task weight of each AGENT in a next operation cycle.

[0007] Optionally, the modules are implemented through the following method: Step one: each AGENT collects environment data of the isolation belt working scene, performs preprocessing, generates an environment feature tensor and a global environment map; Step two: running state information of each AGENT is collected and standardized to generate a standardized state embedding vector. Step three: the global environment map and the standardized state embedding vector of each AGENT are input into an improved QMIX network to perform task division and collaborative planning, and a task distribution result of each AGENT is generated. Step four: the task distribution result and the environment feature tensor are input into an improved TimeSformer network to perform spatio-temporal attention modeling and cross-AGENT interaction attention modeling, and a path target parameter and a control instruction parameter are generated. Step five: a maintenance operation execution module is driven according to the control instruction parameter, and each AGENT performs an isolation belt maintenance operation task based on the path target parameter. Step six: During the execution process, each AGENT exchanges standardized state embedding vectors and job progress information through the wireless communication network, and feeds back to the improved QMIX network to update the task allocation result; Step seven: After the isolation strip maintenance task is completed, the energy consumption evaluation index of each AGENT is collected, the unit path energy consumption prediction value is calculated, and the task weight is updated.

[0008] Optionally, the step one specifically comprises: Collecting terrain data and road boundary information through a laser radar sensor, collecting vegetation image data through a camera, collecting distance data of close-range obstacles through an ultrasonic sensor, and collecting attitude change data and acceleration data through an inertial measurement unit to form environment data; Synchronizing the environment data according to a unified time reference; According to the calibration parameters of the sensors in each AGENT, the vegetation image data, the distance data, the attitude change data and the acceleration data are registered to the laser radar coordinate system by using the extrinsic transformation matrix; Performing voxel filtering and statistical outlier rejection on the terrain data and road boundary information, performing median filtering and bilateral filtering on the vegetation image data, performing sliding window smoothing filtering on the distance data, and performing low-pass filtering on the attitude change data and acceleration data; The data in the environment data that exceeds the set threshold range is removed, and the data in the continuous sampling frame that changes by more than a set threshold is differentially detected and removed to generate standardized environment data; Mapping the standardized environment data to the global environment coordinate system constructed by the master AGENT through pose transformation, and discretizing the spatial range of the isolation strip work area according to the set voxel resolution to obtain a plurality of spatial voxel units; Calculate the channel features of each spatial voxel unit, and combine the channel features of all spatial voxel units in spatial order to generate an environment feature tensor; Based on the environment feature tensor, the master AGENT constructs a global environment map through SLAM technology.

[0009] Optionally, the step two specifically comprises: The running state information includes three-dimensional position coordinate information, power information, load information, task progress information and communication state information; Splicing the running state information into a running state vector, and mapping each dimension of the running state vector to the [0, 1] interval using the min-max normalization method to obtain a normalized running state vector; Linearly mapping and dimensionally compressing the normalized running state vector to generate a standardized state embedding vector.

[0010] Optionally, the step three specifically comprises: The improved QMIX network comprises a plurality of local policy networks, a task graph convolution fusion layer, and a hierarchical mixing network, each AGENT corresponds to a local policy network, the hierarchical mixing network comprises an inspection task mixing network, a job task mixing network, and an energy consumption scheduling mixing network, and each of the local policy network, the inspection task mixing network, the job task mixing network, and the energy consumption scheduling mixing network is a two-layer multi-layer perception structure; The main AGENT divides a job area according to a global environment map, constructs a task candidate set, the task candidate set comprises a plurality of task candidates, maps each task candidate to a candidate task action vector through feature splicing and embedding; The standardized state embedding vector of each AGENT is spliced with each candidate task action vector to form a state-action joint vector; The task weight of each AGENT in the current job cycle is introduced, the task weight is dimensionally expanded and multiplied element by element with the state-action joint vector to obtain a weighted state-action joint vector, and the weighted state-action joint vector is input into the local policy network of each AGENT to calculate a task action Q value; In the task graph convolution fusion layer, a task graph is constructed by the main AGENT; The task action Q values of all AGENTS for the same task candidate are spliced to obtain a node feature vector corresponding to a graph node, and all node feature vectors are arranged according to node numbers to form a node feature matrix; A task graph adjacency matrix is constructed, and corresponding elements of the task graph adjacency matrix are a weighted sum of a distance exponential decay term and a task type consistency term; The adjacency matrix and the node feature matrix are subjected to single-layer graph convolution propagation to obtain a graph convolution node feature matrix; A state coordination matrix is constructed, and corresponding elements of the state coordination matrix are weighted inner products between standardized state embedding vectors of AGENTS and are normalized and mapped by a Sigmoid function; The state coordination matrix and the graph convolution node feature matrix are jointly input into the hierarchical mixing network to calculate an inspection task mixing Q value, a job task mixing Q value, and an energy consumption task mixing Q value, respectively, and to generate a global Q value by weighted combination; Under the premise of satisfying that each task candidate is allocated at most one AGENT and each AGENT receives at most one task candidate, the sum of all global Q values is maximized, and a task allocation matrix is generated by using a greedy matching algorithm; According to the task allocation matrix, output the task allocation result corresponding to each AGENT, which includes task identification, spatial coordinates of the target task area, task type code and task priority, and the task priority is set according to the global Q value sorting result of each AGENT.

[0011] Optionally, the step four specifically includes: The main AGENT determines the target task area of each AGENT according to the task allocation result, extracts the matched spatial voxel unit set from the environmental feature tensor, splices the task allocation result of each AGENT with the matched spatial voxel unit set in the channel dimension, and generates the path feature tensor by organizing according to the time step; The path feature tensor of each AGENT is linearly mapped in the channel dimension to generate a path mapping tensor; A time position encoding tensor is introduced through the sine-cosine encoding, and a spatial position encoding tensor is constructed through the relative position bias; The path mapping tensor, the time position encoding tensor and the spatial position encoding tensor are added element by element to obtain the path embedding tensor of each AGENT; In the time step dimension, the spatial position features of the path embedding tensor of each AGENT are executed to generate a spatial attention tensor; In the spatial position dimension, the time sequence features of the path embedding tensor of each AGENT are executed to generate a time attention tensor; The path embedding tensor, the spatial attention tensor and the time attention tensor are executed to generate the path residual embedding tensor of each AGENT through channel residual connection and linear fusion; At each time step, the attention features between the path residual embedding tensors of any two AGENTS are calculated to generate a global interaction attention tensor; Based on the global interaction attention tensor, the global interaction feature vector sequence of each AGENT is extracted; The global interaction feature vector sequence of each AGENT is predicted through a two-layer MLP structure to predict the path target parameters of each AGENT at each time step, and the path target parameters include target path point plane coordinates, speed module length and orientation angle; The global interaction feature vector sequence of each AGENT is predicted through a one-dimensional convolution structure to predict the control instruction parameters of each AGENT at each time step, and the control instruction parameters include execution state identification, steering angle and energy consumption adjustment coefficient.

[0012] Optionally, the step five specifically includes: The maintenance operation execution module includes a pesticide spraying arm, a pruning arm, a cleaning arm and an execution driving unit. At each time step, each AGENT determines the current operation mode according to the execution state identifier, adjusts the moving direction according to the steering angle, and sets the execution energy level according to the energy consumption adjustment coefficient; Each AGENT controls the motion trajectory and speed of the robot body based on the path target parameters, determines the spatial execution area of the current operation, and sets the spraying angle, spraying speed and spraying flow of the spraying arm, the pruning angle and cutting speed of the pruning arm and the sweeping direction and cleaning power of the cleaning arm in combination with the control instruction parameters. The maintenance operation module is driven by the driving unit to complete the isolation belt maintenance operation task.

[0013] Optionally, the step six specifically comprises: At each time step, each AGENT exchanges standardized state embedding vectors and operation progress information through a wireless communication network; The operation progress information includes a current path execution state identifier, a current operation task completion ratio and a current path time step index; The master AGENT receives the standardized state embedding vectors and operation progress information uploaded by all AGENTS, and generates a global state matrix by fusion; The master AGENT inputs the global state matrix into the improved QMIX network to perform network updating and generate an updated task allocation result.

[0014] Optionally, the step seven specifically comprises: The energy consumption evaluation indexes of each AGENT in the current operation cycle are collected, and the energy consumption evaluation indexes include an actual energy consumption value, a remaining power value, a path execution time and an operation path length; Based on the ratio of the actual energy consumption value to the operation path length, a unit path energy consumption prediction value of each AGENT is calculated, and the unit path energy consumption prediction value is compared with a set energy consumption threshold value; if the unit path energy consumption prediction value exceeds the set energy consumption threshold value, the task weight of the current AGENT in the next operation cycle is adjusted; Based on the updated task weight, the task allocation result of each AGENT in the next operation cycle is adjusted.

[0015] The beneficial effects of the present application are: The present application constructs an isolation belt maintenance operation robot system based on multi-AGENT cooperation, which significantly improves the intelligent level and operation efficiency of multi-robot cooperative operation in complex road environment.

[0016] Firstly, the state cooperation modeling between the AGENTS is realized through an improved QMIX network, so that each AGENT can perceive the execution state and job progress of other AGENTS, thereby dynamically adjusting the task allocation strategy; the improved QMIX network introduces a local policy network, a hierarchical hybrid network and a task graph convolution fusion layer, which effectively enhances the cooperation ability of different AGENTS in the path prediction and job execution stage.

[0017] Secondly, the improved TimeSformer network is introduced to uniformly model the standardized state embedding vector sequence and path residual embedding tensor, which significantly enhances the modeling ability of multiple AGENTS for job state changes and path deviation trends at consecutive time steps; the improved TimeSformer network combines a time modeling module and a spatial attention adjustment mechanism, which can identify small deviations and local conflict areas in path residuals with high precision while maintaining long sequence global modeling capability; further, through joint decoding, the path target parameters and control instruction parameters of each AGENT are generated, which improves the spatiotemporal modeling accuracy of path prediction and effectively avoids the collision risk caused by spatial overlap or path intersection between AGENTS, ensuring the safe cooperation and path non-overlapping between multiple AGENTS in high-density job areas.

[0018] In addition, after each job cycle ends, the actual energy consumption value, residual power value, path execution time and job path length of each AGENT are collected as energy consumption evaluation indicators, and the task weight adjustment is performed based on the unit path energy consumption prediction value and the set energy consumption threshold, realizing energy consumption driven task redistribution optimization and path scheduling adjustment, and enhancing the adaptability of the system to resource changes and environmental disturbances.

[0019] In summary, the application realizes structural innovation and performance synergy improvement in key modules such as task allocation optimization, path prediction modeling and job energy regulation, and has good popularization value and engineering practicability. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings are included to provide a further understanding of the application, and constitute a part of the specification, together with the embodiments of the application, to explain the application, and do not constitute a limitation on the application. In the drawings:

[0021] Fig. 1 is a schematic diagram of a multi-AGENT cooperative isolation belt maintenance robot system proposed by the application; Fig. 2 is a task allocation flowchart of an improved QMIX network in a multi-AGENT cooperative isolation belt maintenance robot system proposed by the application; Fig. 3Is an improved TimeSformer network path prediction flowchart of a barrier maintenance operation robot system based on multi-AGENT cooperation proposed by the application. DETAILED DESCRIPTION

[0022] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams which show only the basic structure of the application in a schematic manner, and thus only show the components relevant to the application.

[0023] Reference Figs. 1-3 A barrier maintenance operation robot system based on multi-AGENT cooperation, comprising the following steps: An environment perception module for collecting environment data of the barrier operation scene and generating an environment feature tensor and a global environment map; A state modeling module for collecting the running state information of each AGENT and generating a standardized state embedding vector; A task coordination and distribution module for inputting the global environment map and the standardized state embedding vector of each AGENT to an improved QMIX network for task division and collaborative planning, and generating a task allocation result for each AGENT; A path planning module for inputting the task allocation result and the environment feature tensor to an improved TimeSformer network to perform spatiotemporal attention modeling and cross-AGENT interaction attention modeling, and generate path target parameters and control instruction parameters for each AGENT at each time step; A maintenance operation execution module for controlling the AGENT body motion trajectory and operation range according to the path target parameters, and driving the operation execution unit to complete the barrier maintenance operation task according to the control instruction parameters; A communication interaction module for exchanging the standardized state embedding vector and the operation progress information of each AGENT through a wireless communication network during task execution, and feeding back to the improved QMIX network to update the task allocation result; A feedback adjustment module for collecting the energy consumption evaluation index of each AGENT after the task is completed, calculating the unit path energy consumption prediction value and updating the task weight of each AGENT in the next operation cycle.

[0024] In this embodiment, the modules are realized through the following methods: Step 1: Each AGENT collects environment data of the barrier operation scene and pre-processes it to generate an environment feature tensor and a global environment map; Step 2: Collect the running state information of each AGENT and perform standardized processing to generate a standardized state embedding vector; Step three: the global environment map and the standardized state embedding vector of each AGENT are input into the improved QMIX network to perform task division and collaborative planning, and the task allocation result of each AGENT is generated; Step four: the task allocation result and the environment feature tensor are input into the improved TimeSformer network to perform spatio-temporal attention modeling and cross-AGENT interaction attention modeling, and the path target parameter and the control instruction parameter are generated; Step five: the maintenance operation execution module is driven according to the control instruction parameter, and each AGENT performs the median strip maintenance task based on the path target parameter; Step six: in the execution process, each AGENT exchanges the standardized state embedding vector and the operation progress information through the wireless communication network, and feeds back to the improved QMIX network to update the task allocation result; Step seven: after the completion of the median strip maintenance task, the energy consumption evaluation index of each AGENT is collected, the unit path energy consumption prediction value is calculated, and the task weight is updated.

[0025] In the embodiment, the step one specifically comprises: The terrain data and the road boundary information are collected by the laser radar sensor, the vegetation image data are collected by the camera, the distance data of the near-range obstacle are collected by the ultrasonic sensor, and the attitude change data and the acceleration data are collected by the inertial measurement unit, to constitute the environment data; The environment data are synchronized according to the unified time reference; According to the calibration parameters of the sensors in each AGENT, the vegetation image data, the distance data, the attitude change data and the acceleration data are registered to the laser radar coordinate system by using the external parameter transformation matrix; The terrain data and the road boundary information are subjected to voxel filtering and statistical outlier rejection, the vegetation image data are subjected to median filtering and bilateral filtering, the distance data are subjected to sliding window smoothing filtering, and the attitude change data and the acceleration data are subjected to low-pass filtering processing; The data in the environment data that exceeds the set threshold range is rejected, and the data in the continuous sampling frame that changes by more than the set threshold is subjected to difference detection and rejection operation, to generate the standardized environment data; The standardized environment data are mapped to the global environment coordinate system constructed by the master AGENT through pose transformation, and the spatial range of the median strip operation area is discretized according to the set voxel resolution, to obtain a plurality of spatial voxel units; The channel features of each spatial voxel unit are calculated, the channel features include the occupancy state, the obstacle distance, the color channel, the image texture, the normal vector and the attitude estimation information, and all the channel features of the spatial voxel units are combined in spatial order to generate the environment feature tensor; Based on the environmental feature tensor, the main AGENT constructs a global environment map through SLAM technology.

[0026] In this embodiment, the step two specifically includes: The running state information includes three-dimensional position coordinate information, power information, load information, task progress information, and communication state information. The running state information is spliced into a running state vector, and the minimum-maximum normalization method is used to map each dimension of the running state vector to the interval [0, 1] to obtain a normalized running state vector. The normalized running state vector is linearly mapped and dimensionally compressed to generate a standardized state embedding vector.

[0027] In this embodiment, the step three specifically includes: The improved QMIX network includes a plurality of local policy networks, a task graph convolution fusion layer, and a hierarchical mixing network, each AGENT corresponding to a local policy network, the hierarchical mixing network including an inspection task mixing network, a job task mixing network, and an energy consumption scheduling mixing network, the local policy network, the inspection task mixing network, the job task mixing network, and the energy consumption scheduling mixing network all being two-layer multilayer perceptron structures. The main AGENT divides the job area according to the global environment map and constructs a task candidate set, the task candidate set including a plurality of task candidates, each task candidate consisting of a task identifier, a task area spatial coordinate, and a task type code; each task candidate is mapped to a candidate task action vector through feature splicing and embedding; The standardized state embedding vector of each AGENT is spliced with each candidate task action vector to form a state-action joint vector. The task weight of each AGENT in the current job cycle is introduced, the task weight is dimensionally expanded and multiplied element by element with the state-action joint vector to obtain a weighted state-action joint vector, which is input into the local policy network of each AGENT to calculate the task action Q value. In the task graph convolution fusion layer, the main AGENT constructs a task graph, specifically: each task candidate in the task candidate set is taken as a graph node, and if the Euclidean distance between two task candidates in spatial position is less than a set threshold, the task types are the same or there is an execution order relationship, then a connection edge is established between the graph nodes. The task action Q values of all AGENTS for the same task candidate are spliced to obtain a node feature vector corresponding to the graph node, and all node feature vectors are arranged according to node number to form a node feature matrix. constructing a task graph adjacency matrix, a corresponding element of the task graph adjacency matrix being a weighted sum of a distance exponential decay term and a task type consistency term, the distance exponential decay term being a Euclidean distance between task candidates calculated by an exponential decay function, the task type consistency term being equal to 1 if task type encodings of two task candidates are the same, otherwise the task type consistency term being equal to 0; performing single-layer graph convolution propagation on the adjacency matrix and the node feature matrix to obtain a graph convolution node feature matrix; In the application, the main AGENT constructs a task graph based on task candidates, and constructs a task graph adjacency matrix in combination with distance exponential decay and task type consistency, so as to ensure that the task graph structure takes into account both physical layout and semantic constraints. Through graph convolution propagation operation, the Q value information of different AGENTS on the same task candidate is effectively fused, the modeling capability of the global collaborative relationship among multiple AGENTS is improved, and the distribution discrimination and robustness of the system in the task intensive area are enhanced.

[0028] constructing a state collaboration matrix, a corresponding element of the state collaboration matrix being a weighted inner product between normalized state embedding vectors of AGENTS and being normalized and mapped by a Sigmoid function: ; wherein, denotes a state collaboration degree between the i th AGENT and the j th AGENT, denotes a normalized state embedding vector of the i th AGENT, denotes a normalized state embedding vector of the j th AGENT, denotes a state collaboration weighted matrix, denotes a Sigmoid function; inputting the state collaboration matrix and the graph convolution node feature matrix into a hierarchical hybrid network, respectively calculating a hybrid Q value of a patrol task, a hybrid Q value of a work task and a hybrid Q value of an energy consumption task, and performing weighted combination to generate a global Q value; In the application, by constructing a state collaboration matrix and inputting the state collaboration matrix and the graph convolution node feature matrix into a hierarchical hybrid network, explicit modeling of state relationships among multiple AGENTS and collaborative evaluation of global task values are realized. The state collaboration matrix describes state correlations among AGENTS through weighted inner product and Sigmoid normalization operation, and can dynamically reflect the cooperation degree of different AGENTS in terms of position, power, load and communication state. The hierarchical hybrid network generates hybrid Q values in three dimensions of patrol, work and energy consumption respectively and performs weighted fusion, which not only ensures the multi-objective balance of task execution, but also enhances the adaptability and decision accuracy of global task distribution, thereby improving the collaborative optimization capability of the system in a complex dynamic environment.

[0029] Maximizing the sum of all global Q values under the premise of meeting the maximum allocation of one AGENT for each task candidate and the maximum reception of one task candidate for each AGENT, and using a greedy matching algorithm to generate a task allocation matrix, if the task candidate j is allocated to the i-th AGENT, the corresponding element of the task allocation matrix is equal to 1, otherwise the corresponding element of the task allocation matrix is equal to 0; According to the task allocation matrix, output the task allocation result corresponding to each AGENT, the task allocation result includes task identification, spatial coordinates of target task area, task type code and task priority, the task priority is set according to the global Q value sorting result of each AGENT, the larger the value is, the higher the task priority is, and the task priority is an integer value label.

[0030] In this embodiment, the step four specifically includes: The main AGENT determines the target task area of each AGENT according to the task allocation result, extracts the matched spatial voxel unit set from the environmental feature tensor, splices the task allocation result of each AGENT with the matched spatial voxel unit set in the channel dimension, and generates the path feature tensor by organizing in the time step; Linearly mapping the path feature tensor of each AGENT in the channel dimension to generate a path mapping tensor; Introducing a time position encoding tensor through a sine-cosine encoding, and constructing a spatial position encoding tensor through a relative position bias; Element-wise adding the path mapping tensor, the time position encoding tensor and the spatial position encoding tensor to obtain the path embedding tensor of each AGENT; In the time step dimension, performing attention operation on the spatial position features of the path embedding tensor of each AGENT to generate a spatial attention tensor; In the spatial position dimension, performing attention operation on the time sequence features of the path embedding tensor of each AGENT to generate a time attention tensor; Performing channel residual connection and linear fusion on the path embedding tensor, the spatial attention tensor and the time attention tensor to generate the path residual embedding tensor of each AGENT; At each time step, calculate the attention features between the path residual embedding tensors of any two AGENTS to generate a global interaction attention tensor, which is used to represent the path interaction relationship and conflict association between all AGENTS at all time steps; In the present application, in order to depict the interaction relationship and possible conflicts between the paths of all AGENTS in the execution process, attention feature calculation needs to be performed on the path residual embedding tensors of any two AGENTS at each time step. For any two AGENTS, the path residual embedding tensors respectively represent the task area space features and target path embedding features of the AGENTS at each time step. First, linear mapping is performed on the two path residual embedding tensors to generate corresponding query tensors, key tensors and value tensors. At the current time step, the query tensor of one AGENT is matched with the key tensor of another AGENT for similarity, and the attention scores are obtained through dot product operation, and all attention scores are normalized. The normalized attention scores are multiplied by the value tensor of the AGENT being focused on to aggregate the spatial path features and complete the interaction information extraction. Each pair of AGENTS performs once at each time step, thereby constructing a global interaction attention tensor covering all AGENTS and all time steps.

[0031] Based on the global interaction attention tensor, a global interaction feature vector sequence of each AGENT is extracted; The global interaction feature vector sequence of each AGENT is predicted through a two-layer MLP structure to predict the path target parameters of each AGENT at each time step, including target path point plane coordinates, speed module length and orientation angle. The global interaction feature vector sequence of each AGENT is predicted through a one-dimensional convolution structure to predict the control instruction parameters of each AGENT at each time step, including execution state identifier, steering angle and energy consumption adjustment coefficient.

[0032] In the present embodiment, the step five specifically comprises: The maintenance operation execution module comprises a pesticide spraying arm, a pruning arm, a cleaning arm and an execution driving unit. At each time step, each AGENT determines the current operation mode according to the execution state identifier, adjusts the moving direction according to the steering angle, and sets the execution energy level according to the energy consumption adjustment coefficient. Each AGENT controls the body motion trajectory and speed based on the path target parameters, determines the spatial execution area of the current operation, and sets the spraying angle, spraying speed and spraying flow of the pesticide spraying arm, the pruning angle and cutting speed of the pruning arm, and the sweeping direction and cleaning power of the cleaning arm in combination with the control instruction parameters. The maintenance operation execution module is driven by the execution driving unit to complete the isolation belt maintenance operation task.

[0033] In the present embodiment, the step six specifically comprises: At each time step, each AGENT exchanges standardized state embedding vectors and operation progress information through a wireless communication network. The job progress information includes the current path execution status identifier, the current job task completion percentage, and the current path time step index; The main agent receives standardized state embedding vectors and job progress information uploaded by all agents and fuses them to generate a global state matrix; The main agent inputs the global state matrix into the improved QMIX network to perform network updates and generate updated task assignment results.

[0034] In this embodiment, step seven specifically includes: Collect energy consumption assessment indicators for each agent in the current work cycle. The energy consumption assessment indicators include actual energy consumption value, remaining power value, path execution time and work path length. Based on the ratio of actual energy consumption to job path length, the predicted energy consumption per unit path for each agent is calculated. This predicted energy consumption per unit path is then compared to a set energy consumption threshold. If the predicted energy consumption per unit path exceeds the set threshold, the task weight of the current agent in the next job cycle is adjusted. ; in, Index representing the job cycle, Indicates the first In the first work cycle The task weight of each agent Indicates the first In the first work cycle The task weight of each agent Represents an exponential function. This represents the energy consumption weighting adjustment coefficient. Indicates the first Predicted energy consumption per path for each agent This indicates the setting of an energy consumption threshold; Based on the updated task weights, adjust the task allocation results for each agent in the next job cycle.

[0035] Example 1 To verify the feasibility of this invention in practice, it was applied to a scenario of refined maintenance of urban greenbelts by a municipal department in a city. On-site deployment and simulation evaluation of multi-Agent robot collaborative spraying and pruning tasks were conducted. This urban greenbelt is densely distributed between main roads and expressways, with a narrow work area, numerous obstacles, and complex tasks, encompassing pest and disease control, pruning, irrigation, and other operations. Traditional manual or single-robot inspections suffer from path planning conflicts, inefficient collaboration, and unreasonable energy allocation, especially in high-temperature or nighttime operation scenarios where efficiency is extremely low, and the mixing of humans and robots poses significant safety hazards.

[0036] In the implementation process, 6 robots are deployed to cover about 2400 meters of main road sections, the robots are equipped with a visual perception system, an odometer and an electric quantity monitoring module, and have the functions of pesticide spraying, pruning and watering; a standardized state embedding vector is constructed through a state modeling module; a task allocation matrix is generated according to the work density, path residual error and current energy consumption weight, and the work content is issued to each AGENT. Each AGENT completes the pesticide spraying and pruning work according to the instructions, uploads the current work frame number, path time step index, remaining proportion of electric quantity and current position coordinates to the main AGENT every 30 seconds during the work, the main AGENT summarizes the state and updates the task weight to optimize the task division in the next work cycle. Feedback evaluation is carried out through path execution time, electric quantity drop amplitude and unit path energy consumption, energy consumption adaptive scheduling is realized, energy distribution is more balanced, and overload of high-frequency work AGENT is avoided.

[0037] To verify the overall performance of the system, the system of the application is compared with the traditional static centralized scheduling scheme, the QMIX scheduling scheme without energy consumption feedback and the path scheduling method based on genetic algorithm. Six robots are selected to perform maintenance tasks for 12 hours in three typical isolated belt scenes, the task content is consistent, and the evaluation indexes include: average path overlap rate, unit path energy consumption, task completion time, system overall electric quantity fluctuation rate, path conflict times and scheduling convergence rounds, and the experimental results are shown in Table 1.

[0038] Table 1 Comparison table of comprehensive performance of the application and the comparison schemes

[0039] As can be seen from the data in Table 1, the application is better than the comparison methods in multiple performance indicators. In terms of path overlap rate, the application is based on improved TimeSformer for path prediction modeling and combined with AGENT interaction attention, which effectively avoids path intersection and redundant task execution, so that the average path overlap rate is reduced to 4.2%; in terms of unit path energy consumption, the application realizes more balanced energy consumption allocation and path selection through dynamic adjustment of task weight and energy consumption prediction feedback, so that the unit path energy consumption is reduced to 2.31 Wh / m, which is obviously better than the comparison schemes; at the same time, the task completion time is shortened by about 20% compared with the comparison schemes, and the electric quantity fluctuation rate is reduced to 8.1%, which fully illustrates the practicability and robustness of the energy consumption scheduling strategy. The application can complete multi-AGENT strategy scheduling convergence within 12 rounds, which is much lower than 30 rounds of the path scheduling scheme based on genetic algorithm, and significantly improves the strategy stability and scheduling efficiency. The path conflict times are zero, which verifies the guarantee ability of the improved TimeSformer network to the safety of multi-AGENT cooperation.

[0040] The isolation belt maintenance operation robot system based on multi-AGENT cooperation has significant intelligent scheduling advantages, energy consumption optimization capabilities and path conflict inhibition effects in a real operation scene, and can effectively improve the intelligent level, energy saving level and system stability of urban maintenance operation.

[0041] The above merely describes a preferred embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can make equivalent replacements or changes to the technical solution and the inventive concept of the present application within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A robot system for maintaining isolation zones based on multi-agent collaboration, characterized in that, include: The environmental perception module is used to collect environmental data of the isolation zone operation scenario and generate environmental feature tensors and a global environmental map; The state modeling module is used to collect the running state information of each agent and generate standardized state embedding vectors. The task collaborative allocation module is used to input the global environment map and the standardized state embedding vectors of each agent into the improved QMIX network for task partitioning and collaborative planning, and generate the task allocation results for each agent. The path planning module is used to input the task allocation results and environmental feature tensors into the improved TimeSformer network, perform spatiotemporal attention modeling and cross-AGENT interaction attention modeling, and generate the path target parameters and control command parameters for each AGENT at each time step. The maintenance operation execution module is used to control the movement trajectory and operation range of the agent body according to the path target parameters, and drive the operation execution unit to complete the isolation zone maintenance operation task according to the control command parameters; The communication and interaction module is used to exchange standardized status embedding vectors and job progress information of each agent through a wireless communication network during task execution, and to feed back the task allocation results to the improved QMIX network. The feedback adjustment module is used to collect energy consumption assessment indicators of each agent after the task is completed, calculate the predicted value of energy consumption per unit path, and update the task weight of each agent in the next work cycle.

2. The multi-agent collaborative isolation zone maintenance robot system according to claim 1, characterized in that, The modules are connected in the following way: Step 1: Each agent collects environmental data of the isolation zone operation scenario, performs preprocessing, and generates environmental feature tensors and a global environmental map; Step 2: Collect the running status information of each agent, perform standardization processing, and generate standardized state embedding vectors; Step 3: The global environment map and the standardized state embedding vectors of each agent are used to perform task partitioning and collaborative planning through an improved QMIX network to generate task allocation results for each agent; Step 4: Input the task allocation results and environmental feature tensors into the improved TimeSformer network to perform spatiotemporal attention modeling and cross-AGENT interaction attention modeling, and generate path target parameters and control command parameters; Step 5: Drive the maintenance operation execution module according to the control command parameters, and each agent executes the isolation strip maintenance operation task based on the path target parameters; Step Six: During execution, each agent exchanges standardized state embedding vectors and job progress information through the wireless communication network and feeds them back to the improved QMIX network to update the task allocation results; Step 7: After the maintenance work on the isolation zone is completed, collect the energy consumption assessment indicators of each agent, calculate the predicted energy consumption per unit path, and update the task weights.

3. The isolation zone maintenance robot system based on multi-agent collaboration according to claim 2, characterized in that, Step one specifically includes: Environmental data is constructed by collecting terrain data and road boundary information through lidar sensors, vegetation image data through cameras, distance data of nearby obstacles through ultrasonic sensors, and attitude change data and acceleration data through inertial measurement units. Synchronize environmental data according to a unified time base; Based on the calibration parameters of the sensors in each agent, the vegetation image data, distance data, attitude change data and acceleration data are registered to the lidar coordinate system using the extrinsic parameter transformation matrix; Voxel filtering and statistical outlier removal are performed on terrain data and road boundary information; median filtering and bilateral filtering are performed on vegetation image data; sliding window smoothing filtering is performed on distance data; and low-pass filtering is performed on attitude change data and acceleration data. Data exceeding a set threshold range in the environmental data is removed, and differential detection and removal operations are performed on data whose variation exceeds a set threshold in consecutive sampling frames to generate standardized environmental data. The standardized environmental data is mapped to the global environmental coordinate system constructed by the main agent through pose transformation, and the spatial range of the isolation zone operation area is discretized according to the set voxel resolution to obtain several spatial voxel units. Calculate the channel features of each spatial voxel unit, and combine the channel features of all spatial voxel units in spatial order to generate the environment feature tensor. Based on the environmental feature tensor, the main agent constructs a global environmental map using SLAM technology.

4. The multi-agent collaborative isolation zone maintenance robot system according to claim 2, characterized in that, Step two specifically includes: The operational status information includes three-dimensional position coordinate information, power information, load information, task progress information, and communication status information; The running status information is concatenated into a running status vector. The minimum-maximum normalization method is used to map each dimension of the running status vector to the interval [0,1] to obtain a normalized running status vector. The normalized running state vector is linearly mapped and its dimensions are compressed to generate a standardized state embedding vector.

5. A robot system for maintaining isolation zones based on multi-agent collaboration according to claim 2, characterized in that, Step three specifically includes: The improved QMIX network includes several local policy networks, task graph convolutional fusion layers, and hierarchical hybrid networks. Each agent corresponds to a local policy network. The hierarchical hybrid network includes an inspection task hybrid network, a job task hybrid network, and an energy consumption scheduling hybrid network. The local policy network, inspection task hybrid network, job task hybrid network, and energy consumption scheduling hybrid network are all two-layer multilayer perceptron structures. The main agent divides the operation area according to the global environment map and constructs a task candidate set. The task candidate set includes several task candidates. Each task candidate is generated into a candidate task action vector through feature concatenation and embedding mapping. The normalized state embedding vector of each AGENT is concatenated with the action vector of each candidate task to form a joint state-action vector; Introduce the task weight of each agent in the current job cycle, expand the dimension of the task weight and multiply it element-wise with the state-action joint vector to obtain the weighted state-action joint vector, and input it into the local policy network of each agent to calculate the task action Q value. In the task graph convolutional fusion layer, the task graph is constructed through the main agent; The task action Q values ​​of all agents for the same task candidate are concatenated to obtain the node feature vector of the corresponding graph node, and all node feature vectors are arranged by node number to form a node feature matrix. Construct a task graph adjacency matrix, wherein the corresponding elements of the task graph adjacency matrix are the weighted sum of the distance exponential decay term and the task type consistency term; The adjacency matrix and the node feature matrix are propagated by a single-layer graph convolution to obtain the graph convolution node feature matrix; Construct a state coordination matrix, the corresponding elements of which are: the weighted inner product between the normalized state embedding vectors of the AGENT and normalized by the Sigmoid function; The state collaboration matrix and the graph convolution node feature matrix are jointly input into the hierarchical hybrid network to calculate the hybrid Q-values ​​of inspection tasks, operation tasks, and energy consumption tasks, respectively, and then perform weighted combination to generate the global Q-value. Under the premise that each task candidate can be assigned to at most one agent and each agent can receive at most one task candidate, the sum of all global Q values ​​is maximized, and a greedy matching algorithm is used to generate the task assignment matrix. Based on the task allocation matrix, the task allocation result corresponding to each AGENT is output. The task allocation result includes the task identifier, the spatial coordinates of the target task area, the task type code, and the task priority. The task priority is set according to the global Q value sorting result of each AGENT.

6. The multi-agent collaborative isolation zone maintenance robot system according to claim 2, characterized in that, Step four specifically includes: The main agent determines the target task region for each agent based on the task allocation results, and extracts a set of spatial voxel units that match the target task region from the environmental feature tensor. The task allocation results of each agent and the set of matching spatial voxel units are concatenated in the channel dimension and the path feature tensor is generated by organizing it according to the time step. The path feature tensor of each agent is linearly mapped along the channel dimension to generate a path mapping tensor. A temporal position coding tensor is introduced by sine and cosine coding, and a spatial position coding tensor is constructed by relative position offset. The path mapping tensor is added element-wise to the temporal location encoding tensor and the spatial location encoding tensor to obtain the path embedding tensor for each agent. In the time step dimension, attention operations are performed on the spatial location features of the path embedding tensor of each AGENT to generate a spatial attention tensor; In the spatial location dimension, attention operations are performed on the temporal series features of the path embedding tensor for each AGENT to generate a temporal attention tensor; Perform channel residual join and linear fusion with the path embedding tensor, spatial attention tensor, and temporal attention tensor to generate the path residual embedding tensor for each agent. At each time step, the attention features between the path residual embedding tensors of any two AGENTs are calculated to generate a global interaction attention tensor; Based on the global interaction attention tensor, extract the global interaction feature vector sequence for each agent; The global interaction feature vector sequence of each agent is used to predict the path target parameters of each agent at each time step through a two-layer MLP structure. The path target parameters include the planar coordinates of the target path point, the velocity magnitude, and the orientation angle. The global interaction feature vector sequence of each agent is used to predict the control command parameters of each agent at each time step through a one-dimensional convolutional structure. The control command parameters include execution status identifier, steering angle and energy consumption adjustment coefficient.

7. The isolation zone maintenance robot system based on multi-agent collaboration according to claim 2, characterized in that, Step five specifically includes: The maintenance operation execution module includes a spraying arm, a pruning arm, a cleaning arm, and an execution drive unit; At each time step, each agent determines its current operating mode based on the execution status flag, adjusts its movement direction based on the steering angle, and sets its execution level based on the energy consumption adjustment coefficient. Each agent controls the machine's movement trajectory and speed based on path target parameters, determines the current work space execution area, and sets the spraying angle, spraying speed and spraying flow rate of the spraying arm, the trimming angle and cutting speed of the trimming arm, and the sweeping direction and cleaning power of the cleaning arm in combination with control command parameters. The maintenance operation module is driven by the drive unit to complete the maintenance operation task of the isolation zone.

8. A robot system for maintaining isolation zones based on multi-agent collaboration according to claim 2, characterized in that, Step six specifically includes: At each time step, each agent exchanges standardized state embedding vectors and job progress information with each other via a wireless communication network; The job progress information includes the current path execution status identifier, the current job task completion percentage, and the current path time step index; The main agent receives standardized state embedding vectors and job progress information uploaded by all agents and fuses them to generate a global state matrix; The main agent inputs the global state matrix into the improved QMIX network to perform network updates and generate updated task assignment results.

9. A robot system for maintaining isolation zones based on multi-agent collaboration according to claim 2, characterized in that, Step seven specifically includes: Collect energy consumption assessment indicators for each agent in the current work cycle. The energy consumption assessment indicators include actual energy consumption value, remaining power value, path execution time and work path length. Based on the ratio of actual energy consumption to work path length, calculate the unit path energy consumption prediction value for each agent, compare the unit path energy consumption prediction value with the set energy consumption threshold, and if the unit path energy consumption prediction value exceeds the set energy consumption threshold, adjust the task weight of the current agent in the next work cycle. Based on the updated task weights, adjust the task allocation results for each agent in the next job cycle.

Images