Dynamic collision avoidance path planning method and system for multi-robot task allocation

Abstract

This invention discloses a dynamic collision avoidance path planning method and system for multi-robot task allocation, relating to the field of robotics. The method includes: acquiring multimodal perception data; constructing a sensory control spatiotemporal knowledge graph, using a graph neural network to predict infection risk, and generating a dynamic risk heatmap; based on the heatmap, using multi-agent reinforcement learning to optimize task allocation and outputting the execution sequence of each robot; fusing local sensor and global dynamic obstacle information to construct a spatiotemporal grid map, redefining costs based on risk, and planning paths using incremental search to control robot movement. This addresses the technical problem of simultaneously achieving efficient collaborative task allocation, real-time dynamic collision avoidance, and proactive infection risk avoidance for multiple robots in complex, dynamic hospital environments with strict infection control requirements. It achieves the technical effect of improving global task execution efficiency, reducing human-robot-object collision risks, and enhancing hospital infection control capabilities through the fusion of sensory control spatiotemporal graphs and reinforcement learning.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more specifically to a dynamic collision avoidance path planning method and system for multi-robot task allocation. Background Technology

[0002] With the continuous improvement of the level of intelligence in medical logistics, various service robots are being used more and more widely in modern hospitals, including material delivery robots, disinfection robots, sample transport robots, and guidance service robots. Multi-robot collaborative operations can significantly improve hospital operational efficiency and reduce the burden on medical staff, and have become a key direction in the construction of smart hospitals.

[0003] However, the highly dynamic and complex hospital environment presents significant challenges to multi-robot task allocation and path planning. First, hospital corridors are narrow with high traffic volumes, and dynamic obstacles such as medical staff, patients, beds, and medical carts are densely packed with highly randomized movements. Traditional path planning methods based on static maps struggle to respond to real-time environmental changes, easily leading to frequent robot stops, path deadlocks, and even collisions. Second, hospitals have extremely high infection control requirements, with varying infection control levels in different areas. Furthermore, contamination risks exhibit dynamic propagation characteristics with the movement of personnel and equipment. Existing robot scheduling systems generally only focus on the feasibility of physical space access, lacking the ability to effectively perceive and avoid biosafety risks. Third, multi-robot task allocation often involves dozens of robots and hundreds of task nodes. Traditional rule-based or centralized integer programming methods struggle to achieve globally optimal scheduling when faced with multi-dimensional constraints such as differences in robot battery power, task urgency, regional congestion, and dynamic risk distribution, resulting in low overall execution efficiency. Summary of the Invention

[0004] This application provides a dynamic collision avoidance path planning method and system for multi-robot task allocation, which solves the technical problem that it is difficult for multiple robots to simultaneously achieve efficient collaborative task allocation, real-time dynamic collision avoidance and active avoidance of infection risks in complex, dynamic hospital environments with strict infection control requirements.

[0005] The first aspect of this application provides a dynamic collision avoidance path planning method for multi-robot task allocation, the method comprising: The system acquires multimodal perception data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records. Based on this multimodal perception data, a spatiotemporal knowledge graph of hospital infection control is constructed. A graph neural network is used to infer contamination transmission paths and predict high-risk areas within this knowledge graph, generating a dynamic risk heatmap. A set of collaborative tasks from multiple robots is received. Based on the real-time status data of each robot and the dynamic risk heatmap, a multi-agent reinforcement learning model is used to optimize global task allocation, outputting the task execution sequence for each robot. For each robot's current task, local sensor data from the robot and global dynamic obstacle information from the hospital information system are fused in real-time to construct a spatiotemporal occupancy grid map. Path costs are redefined based on the dynamic risk heatmap, and an incremental search algorithm is used to plan a collision-free path that satisfies collision avoidance and infection risk constraints. The robot is then controlled to travel along this collision-free path.

[0006] A second aspect of this application provides a dynamic collision avoidance path planning system for multi-robot task assignment, the system comprising: The system comprises the following modules: a data acquisition module, a graph reasoning module, and a path planning module. The data acquisition module acquires multimodal perception data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records. The graph reasoning module, based on the multimodal perception data, constructs a spatiotemporal knowledge graph of hospital infection control and uses a graph neural network to reason about contamination transmission paths and predict high-risk areas, generating a dynamic risk heatmap. The task allocation module receives a set of collaborative tasks from multiple robots and, based on the real-time status data of each robot and the dynamic risk heatmap, optimizes global task allocation using a multi-agent reinforcement learning model, outputting the task execution sequence for each robot. The path planning module, for each robot's current task, integrates the robot's local sensor data with the global dynamic obstacle information from the hospital information system in real time to construct a spatiotemporal occupancy grid map. Based on the dynamic risk heatmap, it redefines the path cost and plans a collision-free path that satisfies collision avoidance and infection risk constraints using an incremental search algorithm, controlling the robot to travel along the collision-free path.

[0007] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multi-source heterogeneous data of the hospital environment are collected, including building structure information, real-time personnel location data, medical equipment operating status, and historical infection control records. Then, based on this data, a spatiotemporal knowledge graph of hospital infection control is constructed, modeling the correlation between spatial structure, personnel activities, and infection events. A graph neural network is used to infer and predict contamination transmission paths, generating a dynamically changing risk heat map to characterize the infection risk level of each area. Upon receiving tasks from multiple robots, a multi-agent reinforcement learning algorithm is used to perform global optimization scheduling, combining real-time status information such as the current position, battery level, and task load of each robot with the risk heat map, generating the optimal task execution order for each robot and achieving a balance between overall efficiency and risk control. Then, during the actual task execution, the results of local sensor perception are integrated with global dynamic obstacle information provided by the hospital information system to construct a real-time updated spatiotemporal occupancy grid map. Infection risk is incorporated into the path cost function, and an incremental search algorithm is used for dynamic path planning. Under the premise of satisfying collision avoidance and risk constraints, a safe and efficient collision-free driving path is generated, ensuring that the path avoids high-risk infection areas and has no collision risk. Ultimately, the robot will execute the task according to the planned collision-free path, ensuring the task is carried out smoothly, safely, and efficiently. Attached Figure Description

[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0009] Figure 1 This is a schematic flowchart of a dynamic collision avoidance path planning method for multi-robot task allocation provided in an embodiment of this application.

[0010] Figure 2 This is a schematic diagram of a dynamic collision avoidance path planning system for multi-robot task allocation provided in an embodiment of this application.

[0011] Figure labeling: Data acquisition module 11, graph reasoning module 12, task allocation module 13, path planning module 14. Detailed Implementation

[0012] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0013] Example 1, as Figure 1As shown, this application provides a dynamic collision avoidance path planning method for multi-robot task allocation, wherein the method includes: Acquire multimodal perception data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records.

[0014] In this embodiment, the system first acquires multimodal perception data of the target hospital environment to form the basic dataset required for subsequent modeling and decision-making. Specifically, the system acquires static spatial structure data of the hospital through pre-import or on-site mapping. This static spatial structure data includes at least the hospital floor plan or vector map, functional labels and permission attributes of each area, connectivity of corridors and passages, access control locations, passage widths, passable time periods, and boundary information for clean areas, semi-contaminated areas, and contaminated areas. The system acquires dynamic personnel positioning data through the hospital's existing positioning infrastructure or sensing network. This dynamic personnel positioning data is used to characterize the real-time distribution and movement status of medical staff, patients, and visitors within the hospital, including personnel identity roles, timestamps, location coordinates, velocity vectors, and trajectory sequences. The system collects medical equipment status data through the hospital's IoT platform or equipment management system. This data includes at least the current location, activation status, operating mode, whether the passageway is blocked, and the expected direction and speed of movement for mobile devices such as mobile beds, treatment carts, infusion pump carts, and wheelchairs. For devices that cannot be directly connected to the network, indirect sensing can be achieved by deploying Bluetooth beacons, UWB tags, or visual recognition markers, and the equipment status is recorded in a time-series format. The system also retrieves historical hospital infection control records from the hospital infection management system, disinfection record system, or HIS / EMR related modules. These records include at least the time and location of historical infection events, the type of infection source or risk event, the departments involved and contact chains, the completion status of disinfection operations, records of the activation and deactivation of isolation measures, a list of key risk areas, and historical population density statistics. After the static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records are collected, they will be synchronized in time, aligned in coordinates, filled with missing values, and removed from outliers. They will then be packaged into a multimodal perception dataset according to a unified data interface, providing basic data support for subsequent map construction, risk reasoning, task allocation, and path planning.

[0015] Based on the multimodal sensing data, a spatiotemporal knowledge graph of hospital infection control is constructed by fusion, and a graph neural network is used to infer the pollution transmission path and predict high-risk areas of the hospital infection control spatiotemporal knowledge graph to generate a dynamic risk heat map.

[0016] In one embodiment, after completing the multimodal perception data acquisition, a spatial topology subgraph is first constructed based on floor plans, room function labels, and passageway connections from the static spatial structure data. An entity dynamic distribution subgraph is then constructed based on the real-time location trajectories and roles of medical staff, patients, and visitors from the dynamic personnel positioning data. Finally, an infection transmission time-series subgraph is constructed based on past infection event times, infection sources, and transmission routes from historical hospital infection control records. These subgraphs are then aligned with entities and fused with relationships to construct a hospital infection control spatiotemporal knowledge graph. Subsequently, based on this hospital infection control spatiotemporal knowledge graph, contamination transmission path reasoning and high-risk area prediction are performed using graph convolutional networks and temporal graph attention networks. The predicted contamination probability is then mapped onto the hospital's two-dimensional planar coordinates to establish a continuously distributed dynamic risk heatmap. Each grid cell in this dynamic risk heatmap contains the infection risk value that changes over time, which can serve as the cost input for subsequent task allocation and path planning, enabling proactive avoidance of high-risk areas.

[0017] Furthermore, the integration and construction of a spatiotemporal knowledge graph for hospital infection control includes: Based on the static spatial structure data, floor plans, room function labels, passageway connections, and access control locations of the hospital area are extracted to construct a spatial topology subgraph. Based on the dynamic personnel positioning data, the real-time location trajectories and roles of medical staff, patients, and visitors are extracted, and combined with the medical equipment status data, an entity dynamic distribution subgraph is constructed. Based on the historical hospital infection control records, the time of past infection events, infection sources, transmission routes, and disinfection operation records of each area are extracted to construct an infection transmission timeline subgraph. The spatial topology subgraph, entity dynamic distribution subgraph, and infection transmission timeline subgraph are aligned and fused to generate the hospital infection control spatiotemporal knowledge graph.

[0018] Preferably, when constructing the spatial topology subgraph, the static spatial structure data of the hospital is digitally analyzed and topologically abstracted. For floor plan data derived from BIM models, CAD drawings, or SLAM mapping results, the system first performs primitive analysis, converting geometric elements such as walls, doorways, corridor boundaries, and elevator shafts into computable polygonal regions. Then, based on region segmentation algorithms, such as connected component analysis or manual annotation rules, the continuous space is divided into several minimum functional unit regions, each assigned a unique ID. Attribute fields are established for each region node, including region name, region function type, cleanliness / contamination level, floor number, area, and maximum capacity. Subsequently, adjacency edges between regions are established based on doorway and corridor connections. Edge attributes include travel distance, passage width, whether robot passage is allowed, one-way passage, and access control status. For elevators and stairs, cross-floor connection edges are used for modeling and marked as vertically connected. A complete spatial topology subgraph is constructed using the above nodes and edges and stored in the form of an adjacency matrix or adjacency list, providing a spatial skeleton for subsequent integration.

[0019] When constructing the dynamic distribution subgraph of entities, the system performs time alignment processing on dynamic personnel location data and medical equipment status data. Specifically, all dynamic data adopts a unified timestamp format and is segmented according to fixed time windows. For each time slice, personnel location coordinate data is read, and the coordinates are mapped to the corresponding region nodes through a spatial mapping algorithm, thereby establishing a "entity-location-region" relationship record. For the regional changes of the same entity in consecutive time slices, the system records its trajectory sequence and calculates the velocity vector and dwell time. In addition to basic identity labels, such as doctor, nurse, patient, caregiver, visitor, each entity node also includes a current status field, such as moving, stationary, entering elevator, leaving area, etc. For medical equipment, the same method is used to establish a "device-location-region" relationship, and attributes such as equipment type, operating status, and whether the passage is occupied are added. In the graph structure, entity nodes are connected to region nodes through edges with time labels, forming a dynamic subgraph with a time dimension. To avoid data explosion, a sliding time window mechanism can be used to retain only the dynamic relationships within the most recent preset time range.

[0020] When constructing the infection transmission timeline subgraph, the system performs structured extraction of historical hospital infection control records. Specifically, it parses the infection record text to extract the infection event occurrence time, occurrence area ID, infection type, infection source identifier, involved personnel ID, and transmission route information. For each infection event, an infection event node is created, and a "Infection Event - Occurred in - Area" relationship edge is established. If there is clear information about the infection source and contacts, relationship edges such as "Infection Source - Contact - Personnel" and "Infection Source - Transmission to - Area" are established. Disinfection records are processed similarly, creating disinfection event nodes and establishing a "Disinfection Event - Acted On - Area" relationship, while also adding parameters such as disinfection start time, end time, and effective duration. Then, the infection events are sorted chronologically, and timeline edges are established between time-adjacent infection events, thus forming a time-chain structure.

[0021] Then, based on a unified system of region IDs and entity IDs, the system uses region nodes in the spatial topology subgraph as the basic master nodes, mapping region references in the entity dynamic distribution subgraph and the infection transmission timeline subgraph to the same node object. For personnel nodes, if the same personnel ID appears in the infection record and the location data, they are merged into a single node, integrating all their attributes and relationships. Duplicate or semantically consistent relationship types are normalized and encoded. Next, a graph database is used for storage, writing nodes, edges, and their attributes into the database in a structured form and creating indexes to support efficient queries. Time attributes are represented using timestamp fields or independent time nodes, enabling the graph to possess both spatial structure and temporal evolution capabilities. Through these steps, a multi-type, multi-relationship, time-attributed spatiotemporal knowledge graph for hospital infection control, containing spatial structure, personnel dynamics, equipment status, and infection transmission information, is finally generated, providing a complete and computable data foundation for subsequent node embedding computation and risk reasoning in graph neural networks.

[0022] Furthermore, generating a dynamic risk heatmap includes: The node features and edge features in the hospital infection control spatiotemporal knowledge graph are input into a pre-trained graph convolutional network. The graph convolutional network learns the infection risk representation of nodes through multi-layer neighbor aggregation. A temporal graph attention network is used to capture the evolution law of infection risk in the time dimension and outputs the pollution probability of each spatial region in the preset future time domain. The pollution probability is mapped to the two-dimensional plane coordinates of the hospital and a continuously distributed dynamic risk heat map is generated using the kernel density estimation method.

[0023] Optionally, after constructing the hospital infection control spatiotemporal knowledge graph, the system converts it into structured input data that can be processed by a graph neural network. Specifically, it encodes spatial region nodes, personnel nodes, equipment nodes, and infection event nodes according to node type, and constructs a fixed-dimensional feature vector for each node. These node feature vectors include at least the region's historical infection frequency, population density statistics, recent disinfection time interval, current number of personnel, region functional category code, cleanliness / contamination level code, and time / location code. For personnel or equipment nodes, they also include numerical features such as role or equipment category, activity frequency, and movement speed statistics. Edge features include attributes such as spatial connectivity distance, contact frequency, propagation relationship weight, and time interval, and are uniformly normalized. Subsequently, the graph structure is represented as an adjacency matrix, node feature matrix, and edge feature tensor, and input into a pre-trained graph convolutional network. In the graph convolutional network, a multi-layer neighbor aggregation mechanism is used to embed and update nodes. Specifically, in each layer of graph convolution, for a given region node, its hidden representation is obtained by weighted aggregation of the features of its neighboring nodes. By stacking multiple layers of graph convolutional networks, node representations can incorporate multiple neighborhood information, enabling regional nodes to reflect not only their own historical infection data but also the influence of surrounding areas, related personnel, and infection events. The final output is a risk embedding vector for each spatial region node, serving as an infection risk representation. Subsequently, to model the evolution of risk over time, a temporal graph attention network is introduced to process the graph embedding results of consecutive time slices. Specifically, the knowledge graph is used to generate a sequence of graph snapshots according to fixed time windows, and the regional infection risk representations obtained from each time slice through the graph convolutional network are used as temporal inputs. The temporal graph attention network employs a self-attention mechanism to model the dependencies between different time slices. For a region's representation at time t, its future risk prediction value is obtained by weighted summation of historical time slice representations. Through this mechanism, the network can capture the lag effect, periodic changes, and sudden growth trends of infection risk. The final network output is the contamination probability value for each spatial region node within a preset future time range, which is normalized to the 0-1 range using a sigmoid function. Then, the predicted contamination probabilities of each region are mapped to the hospital's two-dimensional plane coordinate system. That is, based on the geometric boundary corresponding to each region in the spatial topology sub-map, the region's contamination probability is assigned to the grid cells covered by that region. To obtain a continuous and smooth risk distribution field, a kernel density estimation method is used to spatially smooth the discrete risk values. In this process, for any grid point, the spatial distance between that grid point and the centroids of all regions is calculated. Then, based on the Euclidean distance between the grid point and the region centroids, a distance decay weight is calculated. This distance decay weight is determined by a preset Gaussian kernel function; that is, the closer the distance, the larger the weight; the farther the distance, the smaller the weight, and the weight decays exponentially.Next, the pollution probability of each region is multiplied by the corresponding distance attenuation weight to obtain the risk contribution value of that region to the current grid point. Then, the contribution values ​​of all regions to the grid point are summed to obtain the final risk value of the grid point. By traversing and calculating across all map grids, a continuously distributed risk field matrix is ​​generated. Finally, this matrix is ​​output in the form of a heatmap, where each grid contains a real-time updated risk cost value, which is updated continuously over a time window, forming a dynamic risk heatmap. This dynamic risk heatmap serves as the risk input for subsequent task allocation and path planning, enabling proactive avoidance and dynamic adjustment of high-risk areas.

[0024] Furthermore, the pre-trained graph convolutional network and temporal graph attention network include: Collect multimodal perception data of historical hospital infection control event sequences and corresponding time periods, construct a sample knowledge graph set, and use the actual infection location and time as supervision labels; construct a composite loss function, which includes node classification loss and link prediction loss, and train it together with gradient descent until the network converges; deploy the trained graph convolutional network and temporal graph attention network to an online inference environment for real-time dynamic risk heat map generation.

[0025] Optionally, the system extracts infection event sequences within a certain time range from the hospital's historical database and simultaneously retrieves the corresponding multimodal sensing data for that time period, including spatial structure information, personnel dynamic distribution, equipment operating status, and disinfection records. Using fixed time windows as units, the continuous time period is divided into several time slices. For each time slice, a corresponding snapshot of the hospital's infection control spatiotemporal knowledge graph is constructed. Multiple consecutive time slices constitute a sample sequence. For each sample sequence, the actual location and time of the infection event are recorded, and the spatial area where infection occurred is marked as a high-risk category, while the area where no infection occurred is marked as a low-risk category, thus forming node-level supervisory labels. Simultaneously, based on the actual transmission chain records, pairs of areas or personnel-area pairs with actual transmission relationships are marked for link prediction supervision. Finally, a sample knowledge graph set is formed, consisting of "graph structure data, time series, node labels, and transmission relationship labels." Subsequently, a spatial-dimensional graph convolutional network and a temporal-dimensional temporal graph attention network are constructed. The graph convolutional network adopts a multi-layer stacked structure, including an input layer, two to three layers of graph convolutional hidden layers, and an output embedding layer. The input layer receives node feature vectors and edge attribute encodings. Each graph convolutional hidden layer contains a neighbor aggregation unit, a feature linear transformation unit, and a non-linear activation unit. The neighbor aggregation unit integrates the feature information of neighboring nodes, the linear transformation unit maps the feature dimension, and the activation unit enhances expressive power. The output dimension of the hidden layer is compressed or kept consistent layer by layer, ultimately obtaining the risk embedding representation of each spatial region node. Next, the region embedding vectors obtained from consecutive time slices are input into the temporal graph attention network, which consists of a temporal encoding layer, a multi-head attention layer, and a temporal fusion layer. The temporal encoding layer adds time position identifiers to different time slices; the multi-head attention layer calculates the importance of each historical time slice to the current time slice, modeling different temporal dependency patterns in parallel through multiple attention heads; the temporal fusion layer integrates the attention weighting results and outputs the risk prediction value of each region within the future time window.

[0026] Before training, a composite loss function is constructed to achieve multi-objective joint optimization. This composite loss function includes node classification loss and link prediction loss. The node classification loss constrains the difference between the model's output of the region risk prediction result and the actual infection label. The error is calculated by comparing the consistency between the predicted category and the actual high-risk region label, and can be quantified based on the cross-entropy loss function. The link prediction loss constrains the model's ability to predict potential propagation relationships. The error is calculated by judging whether the propagation connection predicted by the model is consistent with the actual propagation record, and can be quantified based on the contrastive loss function. The system weights and sums the two losses according to preset weights to form the composite loss function. During training, a mini-batch iterative approach is used. The sample knowledge graph sequence is input into the network, the error between the prediction result and the actual label is calculated, and then the error signal is passed layer by layer to each network parameter through the backpropagation algorithm. Gradient descent optimization algorithms, such as adaptive learning rate optimization methods, are used to update the weight parameters. During training, the trend of loss change on the validation set is monitored. When the node classification accuracy and link prediction accuracy steadily improve and the overall loss tends to stabilize within several consecutive iterations, the network is considered to have converged and training is stopped. Finally, the parameters of the trained graph convolutional network and temporal graph attention network are solidified and exported as an inference model file, which is then deployed to the online inference server or edge computing node of the hospital robot. During online operation, the system receives the latest multimodal perception data in real time, constructs a snapshot of the knowledge graph at the current moment, and inputs it into the graph convolutional network to obtain regional risk embeddings. These embeddings are then combined with recent time-slice sequences input into the temporal graph attention network to output the contamination probability of each spatial region within a preset future time range. The prediction results are spatially mapped to generate a dynamic risk heatmap, which is updated at a fixed refresh rate. This provides real-time risk input for task allocation and path planning, improving the robot's safety and proactive control capabilities in complex hospital environments.

[0027] Receive a set of collaborative tasks for multiple robots, optimize global task allocation using a multi-agent reinforcement learning model based on the real-time status data of each robot and the dynamic risk heatmap, and output the task execution sequence of each robot.

[0028] In one embodiment, upon receiving a set of multi-robot collaborative tasks from an upper-layer business system, such as a delivery system, disinfection system, or inspection system, the set of tasks is structured. Each task is abstracted into a task node, and task attribute fields are generated. These task attributes include task type, starting and target point area numbers, priority, latest completion deadline, estimated service duration, whether it needs to pass through a designated area, and required load capacity or interface capabilities. Then, real-time status data of each robot is collected and combined with the task attributes as input for allocation decisions. The multi-robot task allocation process is defined as a constrained multi-agent decision problem, aiming to minimize the overall task completion time and total risk exposure. Framework learning and task allocation are performed using Markov decision theory and a multi-agent deep deterministic policy gradient algorithm, generating a final ordered task list for each robot (i.e., a sequentially arranged task list). Rolling reallocation is triggered when new tasks are added, robot malfunctions occur, congestion suddenly increases, or the risk heatmap changes significantly, updating the robot queues. This achieves a globally optimal or near-optimal multi-robot collaborative task allocation effect oriented towards risk constraints and efficiency goals.

[0029] Furthermore, the task execution sequence of each robot is output, including: Real-time status data of each robot is acquired, including current position coordinates, current battery level, length of the assigned task queue, and estimated task completion value. Based on the dynamic risk heatmap, the infection risk cost of each task target area and its path is extracted. The multi-robot task allocation problem is modeled as a constrained Markov decision process with the goal of minimizing the overall task completion time and total risk exposure. A multi-agent deep deterministic policy gradient algorithm is used to learn the task allocation strategy through centralized training and a distributed execution framework, outputting an ordered task list for each robot.

[0030] Preferably, the system first sends status acquisition commands to each robot scheduling interface at a fixed decision cycle to obtain real-time operational status data for each robot. This status data includes the robot's current global coordinates, current battery percentage and estimated remaining range, current assigned task queue length, estimated completion time of each task, and estimated remaining service time. This data is then normalized and mapped to form an individual robot state vector. Subsequently, all robot states are concatenated to form a global state representation, which, combined with the task attributes of currently unassigned tasks, constitutes the environmental state input for the reinforcement learning model. Next, a dynamic risk heatmap is invoked to obtain the risk value corresponding to the target area for each candidate task. Based on the current robot position and the task target position, a simplified cost map is used to obtain candidate paths from the robot's current position to the task target area, along with the cumulative risk heatmap value on the grid cells covered by these paths, thereby calculating the infection risk cost of the path. For robots with existing task queues, the complete execution path after task sequential insertion is also simulated. Finally, corresponding infection risk cost and time cost features are generated for each "robot-task" combination as part of the decision input. Subsequently, the multi-robot task allocation problem is formalized as a constrained Markov decision process. The state space is defined as the state vectors of all robots, the features of the unfinished task set, and the current risk heatmap summary features. The action space is defined as a set of actions for each robot to select a task to be executed or to adjust the task order. The state transition mechanism is defined as the changes in robot state, battery level, task queue, and environmental time progression after executing the selected action. To reflect the optimization objective, a comprehensive reward function is constructed, which considers both the total task completion time and the cumulative risk exposure value, and incorporates constraints such as energy consumption constraints, task timeout penalties, and illegal allocation penalties. Through the design of the reward function, the strategy tends to shorten the overall completion time and reduce risk exposure in the long run. Then, a multi-agent deep deterministic policy gradient algorithm is used for training. Specifically, an actor network is constructed for each robot to output continuous or discretized task selection decisions based on the current local observation state. At the same time, a centralized critic network is constructed to receive the state and action information of all robots during the training phase and evaluate the global value of the joint actions. During training, the system repeatedly generates task flows and risk-changing scenarios in a simulation environment. The agent outputs actions based on the current policy, and the environment updates its state and calculates rewards based on these actions. All state, action, reward, and next state data are stored in an experience replay buffer. Next, small batches of data are randomly sampled from the buffer for network updates. The critic network adjusts its parameters by minimizing the value assessment error, while the actor network updates its parameters using a policy gradient method to ensure its output actions improve global value.

[0031] After the model training converges, the actor networks corresponding to each robot are deployed to an online scheduling system, which employs a centralized training and distributed execution framework. During online execution, each robot independently outputs task selection decisions based solely on its own state and shared risk and task information. The system then aggregates these decisions to generate an ordered task list for each robot, clearly defining the order of task execution and achieving a balanced, efficient, and safe allocation of tasks across multiple robots. This addresses the problems of uneven allocation, low efficiency, and neglect of infection control risks inherent in traditional methods.

[0032] Furthermore, the task allocation strategy is learned through centralized training and a distributed execution framework, including: For each robot, an actor network and a critic network are constructed. The actor network outputs a task selection probability distribution based on local observation states, while the critic network receives state-action pairs from all robots during training to calculate the global action value. An experience replay buffer is designed to store the states, actions, rewards, and next states of all robots at each time step. The reward function integrates task completion efficiency, energy saving, and sensory control risk avoidance. Small batches of samples are randomly sampled from the experience replay buffer to update the critic network parameters, and the actor network parameters are updated through policy gradients. This process is iterated until the reward curve converges.

[0033] Optionally, an actor network and a critic network are first established for each robot. The actor network is used to output task selection decisions based on the robot's own local observation state. This local observation state includes at least the current region number or coordinates, power level, current task queue length, estimated remaining working time, executable task list features, and local risk summary information from the risk heatmap. The actor network preferably adopts a multi-layer fully connected neural network structure, including an input layer, two to three hidden layers, and an output layer. The hidden layers use non-linear activation functions to enhance expressive power. The output layer generates a probability distribution vector of the corresponding dimension based on the number of tasks, representing the probability of selecting each candidate task. The critic network adopts a centralized structure during the training phase. Its input is the state-action combination vector of all robots, that is, concatenating the current states of all robots and adding the joint action encoding of the corresponding time step. The critic network also adopts a multi-layer fully connected structure, and its output is a scalar value used to represent the value evaluation result of the joint action in the current global state. The critic network is only used during the training phase and does not participate in decision-making during the online execution phase. Subsequently, in the experience data organization phase, an experience replay buffer is designed to store interaction data. Each time the system enters a decision time step, all robots update their robot status, task completion status, battery level changes, and risk exposure based on the current actor network output actions and action execution results. An immediate reward is calculated using a reward function, which consists of multiple indicators, including task completion efficiency, energy saving, infection control risk avoidance, and necessary constraint penalties. The task completion efficiency indicator is the number of tasks completed per unit time; the energy saving indicator is the difference between battery consumption and the optimal energy consumption benchmark; the infection control risk avoidance indicator is a negative weighted average of the cumulative path risk exposure; and the constraint penalties are a weighted average of task timeout duration, battery shortage value, or number of conflicts. The system packages the global state of the current time step, the actions of all robots, the calculated global reward, and the global state of the next time step into a single empirical data point and stores it in an experience replay buffer. The buffer uses a first-in, first-out (FIFO) mechanism and has a maximum capacity. When the capacity reaches its limit, the oldest data is overwritten to ensure the temporal diversity and stability of the sample distribution.

[0034] Subsequently, the system randomly samples a small batch of data from the experience replay buffer at a preset frequency, inputs the states and actions in the samples into the critic network, calculates the current estimated value, and then constructs a value error by combining the immediate reward and the target value of the next state. Next, this error is minimized using the backpropagation algorithm to update the critic network parameters. Then, the critic network parameters are fixed, and the parameters of each actor network are updated using the policy gradient method, so that the actions output by the actor network can obtain higher value scores under the critic network's evaluation. To improve training stability, a target network mechanism can be adopted, that is, a delayed-update target network is set for both actors and critics, and the parameters are gradually synchronized through soft updates. Finally, key indicators such as the global cumulative reward curve, average task completion time, and average risk exposure value are continuously monitored during training. When the cumulative reward value tends to stabilize over several consecutive training cycles, and the performance indicators in the verification scenario no longer significantly improve, the policy network is considered to have converged. After convergence, the parameters of each actor network are fixed and deployed. In the actual operating environment, each robot independently calls its corresponding actor network to select tasks, thereby achieving the optimized effect of multi-robot collaborative task allocation through centralized training and distributed execution.

[0035] For each robot's current task, the local sensor data of the robot and the global dynamic obstacle information of the hospital information system are integrated in real time to construct a spatiotemporal occupancy grid map. Based on the dynamic risk heat map, the path cost is redefined, and a collision-free path that meets the collision avoidance constraint and infection risk constraint is planned through an incremental search algorithm. The robot is then controlled to travel along the collision-free path.

[0036] In one embodiment, once a robot determines its current task, it first continuously acquires point cloud data of the surrounding environment using its onboard LiDAR, depth camera, or millimeter-wave radar. The raw point cloud data is then filtered, denoised, and segmented to identify static obstacles (such as walls and fixed equipment) and dynamic obstacles (such as people and mobile carts). Subsequently, the data collected by these local sensors is uniformly converted to a global coordinate system along with the global obstacle data from the hospital information system. This data is then fused based on the confidence level of the data source to establish a spatiotemporal occupancy grid map. This grid map is divided into two-dimensional grid cells of fixed resolution, each containing spatial occupancy probability and temporal prediction information. Next, a dynamic risk heatmap is overlaid onto the spatiotemporal occupancy grid map to redefine path costs. An incremental search algorithm is then used to search for the minimum cost path, starting from the robot's current position and ending at the task target point. This yields a collision-free trajectory consisting of continuous grid cells that satisfies collision avoidance and infection risk constraints. This trajectory drives the robot to move smoothly along the path, ensuring efficient and safe operation in complex hospital environments.

[0037] Furthermore, constructing a spatiotemporal occupancy raster map includes: The robot acquires point cloud data of the surrounding environment using its LiDAR and depth camera, and identifies static obstacles and dynamic moving entities through filtering and clustering. It obtains the real-time location and movement speed of medical staff ID cards, medical cart Bluetooth beacons, and hospital bed mobile devices through the hospital information system interface, and converts them into occupancy probabilities in the global coordinate system. The local perceived obstacles and global dynamic obstacles are aligned in coordinate and fused with confidence to generate a local spatiotemporal occupancy raster map centered on the robot.

[0038] Preferably, the robot first acquires 2D or 3D point cloud data of the surrounding environment through periodic scanning with LiDAR, and simultaneously acquires depth maps and image data through a depth camera. After scanning, the raw point cloud is preprocessed, including outlier removal, noise filtering, ground segmentation, and voxel downsampling, to improve data quality and processing efficiency. Subsequently, a clustering method based on Euclidean distance is used to group the remaining point cloud, dividing spatially continuous points into several obstacle candidates. For detected obstacles, their velocity and direction of motion are estimated through continuous multi-frame data matching and target tracking algorithms. If an obstacle's position remains basically unchanged within a continuous time window, it is determined to be a static obstacle; if its position changes significantly over time, it is determined to be a dynamically moving entity, and its current position, velocity vector, and predicted trajectory are recorded. Subsequently, the system obtains real-time location and speed information of medical staff ID cards, medical cart Bluetooth beacon data, and movable hospital bed equipment through the hospital information system interface. These are typically provided in the form of global coordinates or region numbers. If the data is a region number, the system converts it into coordinate values ​​through region center point mapping; if it is local coordinates, it maps them to a unified hospital global coordinate system through a preset coordinate transformation matrix. For each dynamic entity, the system generates its current position coordinates, movement direction, velocity magnitude, and data update timestamp, and assigns an initial confidence value based on the reliability of the data source. Then, obstacle data detected by local sensors and dynamic entity data provided by the hospital information system are uniformly mapped to the same global coordinate system. During this process, the system determines whether the two data sources correspond to the same physical entity based on a spatial distance threshold. If the distance is less than the preset threshold, they are considered the same target, and their position and velocity are merged and updated using a confidence-weighted fusion method; if only a single source of data exists, that data is retained and weighted according to its source confidence. This fusion process forms a unified list of dynamic obstacles, each obstacle containing its position, velocity, predicted movement direction, and overall confidence. Next, a fixed-range local two-dimensional grid map is established centered on the robot's current position, and the grid resolution is set. For each grid cell, the system not only records whether it is currently occupied but also predicts the probability of it being occupied within a preset time window. Specifically, for grid cells containing static obstacles, their occupancy probability is set to a high value or close to 1 throughout the entire time window; for dynamically moving entities, their trajectory is extrapolated based on their current speed and direction to calculate their possible locations at future time steps, and the occupancy probability of the corresponding grid cells is attenuated according to their distance from the current time; for areas without obstacles, their occupancy probability remains low. If a grid cell is simultaneously affected by predictions from multiple entities, the overall occupancy probability is calculated by summing the probabilities or taking the maximum value.Finally, all the local spatiotemporal occupancy grid maps constructed centered on the robot are stitched together to form a spatiotemporal occupancy grid map. This grid map contains two-dimensional spatial coordinates and time window dimension information, with the value of each grid cell representing the probability that the location is occupied within the set time window. The system updates the grid map at a fixed refresh rate and uses it as input for path planning. When the occupancy probability exceeds a preset threshold, the grid is considered an impassable or high-cost area during the planning phase. Through the above process, a local spatiotemporal occupancy grid map integrating local perception and global dynamic information is constructed, providing accurate and predictable environmental occupancy information for subsequent incremental path planning.

[0039] Furthermore, an incremental search algorithm is used to plan collision-free paths that satisfy both collision avoidance and infection risk constraints, including: The cost of impassable grid cells in the spatiotemporal occupancy grid is set to infinity, and the infection risk cost is superimposed on the passable grid cells according to the dynamic risk heatmap; an improved method is adopted. The algorithm starts from the robot's current position and ends at the task objective point, incrementally searching for the minimum cost path in a dynamically changing cost map. During the search process, affected path nodes are locally repaired based on the real-time updates of the spatiotemporal occupancy grid map, and a collision-free trajectory is output.

[0040] Preferably, a comprehensive cost map is first generated based on a local spatiotemporal occupancy grid map. During this process, the path cost of grid cells deemed impassable is set to a maximum value, allowing the planning algorithm to automatically avoid these areas during the search. These impassable grid cells are static obstacle areas with an occupancy probability exceeding a preset threshold or areas with a high probability of being occupied by dynamic entities within a future time window. For traversable grids, an infection risk cost is superimposed on the basic distance cost. This infection risk cost is mapped from the risk value of the corresponding grid in the dynamic risk heatmap and weighted using preset weights, increasing the path cost for areas with higher risks. The comprehensive cost map thus simultaneously reflects spatial distance cost, dynamic collision avoidance constraints, and infection risk constraints. Subsequently, starting from the grid cell where the robot is currently located and ending at the grid cell corresponding to the task target point, a minimum cost path search is performed on the comprehensive cost map. Specifically, the improved... The algorithm maintains a priority queue and cost consistency conditions to generate a path with the minimum total cost from the starting point to the ending point during the initial planning phase, and saves the cost value and parent node information of each path node. When the environmental cost remains unchanged, the robot follows this path. When the cost map is locally updated, instead of performing a full map search from scratch, the robot updates the cost value of the affected grid cells and adds the relevant nodes to the update queue, thus performing a local repair-style re-search to improve real-time performance. During the search and execution process, when the spatiotemporal occupancy grid map is updated due to newly detected obstacles or changes in the predicted motion of dynamic entities, the system triggers a local replanning mechanism. In this process, the affected grid regions and their adjacent path nodes are identified, their costs are updated, and the affected path segments are recalculated without changing the unaffected path portions. If an impassable area appears within a certain distance ahead on the current path, local repair is immediately performed and a new path segment is generated, ensuring that the robot always travels along a feasible path. Then, curve fitting is performed on the obtained discrete grid path. For example, spline curve interpolation or corner reduction methods are used to eliminate sharp corners and form a continuous curved path. This is then combined with the robot's own kinematic model, such as a differential drive model or an omnidirectional wheel model, to constrain the path curvature and ensure that the turning radius meets the robot's minimum turning radius constraint. Simultaneously, velocity and acceleration are constrained and planned to avoid sudden acceleration or sharp turns. The resulting trajectory is a smooth, continuous, collision-free trajectory that satisfies the robot's kinematic constraints and does not spatially conflict with dynamic or static obstacles. Finally, speed and steering control commands are generated based on this trajectory to drive the robot safely along the planned path, enabling the robot to perform tasks smoothly and safely in the complex and dynamic hospital environment.

[0041] In summary, the embodiments of this application have at least the following technical effects: First, multimodal perception data of the target hospital environment is acquired, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records. Then, based on this multimodal perception data, a spatiotemporal knowledge graph of hospital infection control is constructed. A graph neural network is then used to infer contamination transmission paths and predict high-risk areas within this knowledge graph, generating a dynamic risk heatmap. Next, a set of multi-robot collaborative tasks is received. Based on the real-time status data of each robot and the dynamic risk heatmap, a multi-agent reinforcement learning model is used to optimize global task allocation, outputting the task execution sequence for each robot. Finally, for each robot's current task, local sensor data from the robot and global dynamic obstacle information from the hospital information system are fused in real-time to construct a spatiotemporal occupancy grid. Path costs are redefined based on the dynamic risk heatmap, and an incremental search algorithm is used to plan a collision-free path that satisfies collision avoidance and infection risk constraints. The robot is then controlled to travel along this collision-free path. It solves the technical problem that multiple robots in complex, dynamic hospital environments with strict infection control requirements cannot simultaneously achieve efficient collaborative task allocation, real-time dynamic collision avoidance, and proactive avoidance of infection risks. It achieves the technical effect of improving the overall task execution efficiency, reducing the risk of human-machine-object collisions, and enhancing hospital infection control capabilities by integrating infection control spatiotemporal mapping and reinforcement learning.

[0042] Example 2, based on the same inventive concept as the dynamic collision avoidance path planning method for multi-robot task allocation in the foregoing examples, such as... Figure 2 As shown, this application provides a dynamic collision avoidance path planning system for multi-robot task allocation, wherein the system includes: Data acquisition module 11: Acquires multimodal perception data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records; Graph reasoning module 12: Based on the multimodal perception data, it fuses and constructs a hospital infection control spatiotemporal knowledge graph, and uses a graph neural network to reason about the pollution transmission path and predict high-risk areas in the hospital infection control spatiotemporal knowledge graph, generating a dynamic risk heat map; Task allocation module 13: Receives a set of multi-robot collaborative tasks, and optimizes global task allocation using a multi-agent reinforcement learning model based on the real-time status data of each robot and the dynamic risk heat map, outputting the task execution sequence of each robot; Path planning module 14: For the current task of each robot, it fuses the robot's local sensor data and the global dynamic obstacle information of the hospital information system in real time to construct a spatiotemporal occupancy grid map, redefines the path cost based on the dynamic risk heat map, and plans a collision-free path that satisfies collision avoidance constraints and infection risk constraints through an incremental search algorithm, controlling the robot to travel along the collision-free path.

[0043] Furthermore, the graph inference module 12 is used to perform the following methods: Based on the static spatial structure data, floor plans, room function labels, passageway connections, and access control locations of the hospital area are extracted to construct a spatial topology subgraph. Based on the dynamic personnel positioning data, the real-time location trajectories and roles of medical staff, patients, and visitors are extracted, and combined with the medical equipment status data, an entity dynamic distribution subgraph is constructed. Based on the historical hospital infection control records, the time of past infection events, infection sources, transmission routes, and disinfection operation records of each area are extracted to construct an infection transmission timeline subgraph. The spatial topology subgraph, entity dynamic distribution subgraph, and infection transmission timeline subgraph are aligned and fused to generate the hospital infection control spatiotemporal knowledge graph.

[0044] Furthermore, the graph inference module 12 is used to perform the following methods: The node features and edge features in the hospital infection control spatiotemporal knowledge graph are input into a pre-trained graph convolutional network. The graph convolutional network learns the infection risk representation of nodes through multi-layer neighbor aggregation. A temporal graph attention network is used to capture the evolution law of infection risk in the time dimension and outputs the pollution probability of each spatial region in the preset future time domain. The pollution probability is mapped to the two-dimensional plane coordinates of the hospital and a continuously distributed dynamic risk heat map is generated using the kernel density estimation method.

[0045] Furthermore, the graph inference module 12 is used to perform the following methods: Collect multimodal perception data of historical hospital infection control event sequences and corresponding time periods, construct a sample knowledge graph set, and use the actual infection location and time as supervision labels; construct a composite loss function, which includes node classification loss and link prediction loss, and train it together with gradient descent until the network converges; deploy the trained graph convolutional network and temporal graph attention network to an online inference environment for real-time dynamic risk heat map generation.

[0046] Furthermore, the task allocation module 13 is used to perform the following method: Real-time status data of each robot is acquired, including current position coordinates, current battery level, length of the assigned task queue, and estimated task completion value. Based on the dynamic risk heatmap, the infection risk cost of each task target area and its path is extracted. The multi-robot task allocation problem is modeled as a constrained Markov decision process with the goal of minimizing the overall task completion time and total risk exposure. A multi-agent deep deterministic policy gradient algorithm is used to learn the task allocation strategy through centralized training and a distributed execution framework, outputting an ordered task list for each robot.

[0047] Furthermore, the task allocation module 13 is used to perform the following method: For each robot, an actor network and a critic network are constructed. The actor network outputs a task selection probability distribution based on local observation states, while the critic network receives state-action pairs from all robots during training to calculate the global action value. An experience replay buffer is designed to store the states, actions, rewards, and next states of all robots at each time step. The reward function integrates task completion efficiency, energy saving, and sensory control risk avoidance. Small batches of samples are randomly sampled from the experience replay buffer to update the critic network parameters, and the actor network parameters are updated through policy gradients. This process is iterated until the reward curve converges.

[0048] Furthermore, the path planning module 14 is used to perform the following methods: The robot acquires point cloud data of the surrounding environment using its LiDAR and depth camera, and identifies static obstacles and dynamic moving entities through filtering and clustering. It obtains the real-time location and movement speed of medical staff ID cards, medical cart Bluetooth beacons, and hospital bed mobile devices through the hospital information system interface, and converts them into occupancy probabilities in the global coordinate system. The local perceived obstacles and global dynamic obstacles are aligned in coordinate and fused with confidence to generate a local spatiotemporal occupancy raster map centered on the robot.

[0049] Furthermore, the path planning module 14 is used to perform the following methods: The cost of impassable grid cells in the spatiotemporal occupancy grid is set to infinity, and the infection risk cost is superimposed on the passable grid cells according to the dynamic risk heatmap; an improved method is adopted. The algorithm starts from the robot's current position and ends at the task objective point, incrementally searching for the minimum cost path in a dynamically changing cost map. During the search process, affected path nodes are locally repaired based on the real-time updates of the spatiotemporal occupancy grid map, and a collision-free trajectory is output.

[0050] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic collision avoidance path planning method for multi-robot task allocation, characterized in that, The method includes: Acquire multimodal perception data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records; Based on the multimodal sensing data, a spatiotemporal knowledge graph of hospital infection control is constructed by fusion, and a graph neural network is used to infer the pollution transmission path and predict high-risk areas of the hospital infection control spatiotemporal knowledge graph to generate a dynamic risk heat map. Receive a set of collaborative tasks for multiple robots, optimize global task allocation using a multi-agent reinforcement learning model based on the real-time status data of each robot and the dynamic risk heatmap, and output the task execution sequence of each robot. For each robot's current task, the local sensor data of the robot and the global dynamic obstacle information of the hospital information system are integrated in real time to construct a spatiotemporal occupancy grid map. Based on the dynamic risk heat map, the path cost is redefined, and a collision-free path that meets the collision avoidance constraint and infection risk constraint is planned through an incremental search algorithm. The robot is then controlled to travel along the collision-free path.

2. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 1, characterized in that, The hospital infection control spatiotemporal knowledge graph is constructed by integration, including: Based on the static spatial structure data, the floor plan, room function labels, passageway connections, and access control locations of the hospital area are extracted to construct a spatial topology sub-graph. Based on the dynamic personnel positioning data, the real-time location trajectories and roles of medical staff, patients and visitors are extracted, and combined with the medical equipment status data, a dynamic distribution sub-graph of entities is constructed. Based on the historical hospital infection control records, the time of past infection events, sources of infection, transmission routes and disinfection operation records of each area were extracted to construct an infection transmission timeline sub-graph. The spatial topology subgraph, the entity dynamic distribution subgraph, and the infection transmission time sequence subgraph are aligned and fused to generate the hospital infection control spatiotemporal knowledge graph.

3. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 2, characterized in that, Generate a dynamic risk heatmap, including: The node features and edge features in the hospital infection control spatiotemporal knowledge graph are input into a pre-trained graph convolutional network, which learns the infection risk representation of nodes through multi-layer neighbor aggregation. The temporal graph attention network is used to capture the evolution of infection risk over time and output the probability of contamination in each spatial region within a preset future time domain. The pollution probability is mapped to the two-dimensional plane coordinates of the hospital, and a continuously distributed dynamic risk heat map is generated using the kernel density estimation method.

4. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 3, characterized in that, The pre-trained graph convolutional network and temporal graph attention network include: Collect multimodal perception data of historical hospital infection control event sequences and corresponding time periods, construct a sample knowledge graph set, and use the actual location and time of infection as supervision labels; A composite loss function is constructed, which includes node classification loss and link prediction loss, and the network is jointly trained using gradient descent until convergence. The trained graph convolutional network and temporal graph attention network are deployed to an online inference environment for real-time dynamic risk heatmap generation.

5. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 1, characterized in that, Output the task execution sequence for each robot, including: Acquire real-time status data for each robot, including current position coordinates, current battery level, length of the assigned task queue, and estimated task completion value; Based on the dynamic risk heat map, the infection risk value of each task target area and the path along the route is extracted; The multi-robot task allocation problem is modeled as a constrained Markov decision process with the goal of minimizing the overall task completion time and total risk exposure. A multi-agent deep deterministic policy gradient algorithm is adopted to learn the task allocation strategy through centralized training and distributed execution framework, and output an ordered task list for each robot.

6. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 5, characterized in that, Task allocation strategies are learned through a centralized training and distributed execution framework, including: For each robot, an actor network and a critic network are constructed. The actor network outputs a task selection probability distribution based on the local observation state, and the critic network receives state-action pairs from all robots during training to calculate the global action value. The design includes an experience playback buffer to store the state, actions, rewards, and next state of all robots at each time step. The reward function integrates task completion efficiency, energy saving, and infection control risk avoidance. Randomly sample small batches of samples from the experience replay buffer, update the critic network parameters, and update the actor network parameters through the policy gradient. Iterate until the reward curve converges.

7. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 1, characterized in that, Constructing a spatiotemporal occupancy raster map includes: The robot acquires point cloud data of the surrounding environment using its LiDAR and depth camera, and identifies static obstacles and dynamic moving entities through filtering and clustering. The real-time location and movement speed of medical staff ID cards, medical cart Bluetooth beacons, and hospital bed mobile devices are obtained through the hospital information system interface and converted into occupancy probability in the global coordinate system. By aligning the coordinates of locally perceived obstacles with global dynamic obstacles and fusing their confidence scores, a local spatiotemporal occupancy grid map centered on the robot is generated.

8. The dynamic collision avoidance path planning method for multi-robot task allocation as described in claim 1, characterized in that, Collision-free paths that satisfy collision avoidance and infection risk constraints are planned using an incremental search algorithm, including: The cost of impassable grids in the spatiotemporal occupancy grid is set to infinity, and the infection risk cost is superimposed on the passable grids according to the dynamic risk heat map. Adopting improved The algorithm starts from the robot's current position and ends at the task objective point, incrementally searching for the minimum cost path in a dynamically changing cost map. During the search process, affected path nodes are locally repaired based on the real-time updates of the spatiotemporal occupancy grid, and collision-free trajectories are output.

9. A dynamic collision avoidance path planning method for multi-robot task allocation, characterized in that, The method for implementing the dynamic collision avoidance path planning method for multi-robot task allocation according to any one of claims 1-8, the method comprising: Data acquisition module: Acquires multimodal sensing data of the target hospital environment, including static spatial structure data, dynamic personnel positioning data, medical equipment status data, and historical hospital infection control records; Graph reasoning module: Based on the multimodal perception data, a spatiotemporal knowledge graph of hospital infection control is constructed by fusion, and a graph neural network is used to reason about the pollution transmission path and predict high-risk areas of the hospital infection control spatiotemporal knowledge graph to generate a dynamic risk heat map; Task allocation module: Receives a set of collaborative tasks for multiple robots, optimizes global task allocation based on the real-time status data of each robot and the dynamic risk heatmap, and outputs the task execution sequence of each robot. Path planning module: For each robot's current task, the module integrates the robot's local sensor data with the global dynamic obstacle information of the hospital information system in real time to construct a spatiotemporal occupancy grid map. Based on the dynamic risk heat map, the module redefines the path cost and plans a collision-free path that meets the collision avoidance constraints and infection risk constraints through an incremental search algorithm. The module then controls the robot to travel along the collision-free path.

Images