Indoor and outdoor logistics robot multi-machine cooperation system

Abstract

The invention relates to the technical field of intelligent robot control, and discloses an indoor and outdoor logistics robot multi-machine cooperation system. The method comprises the following steps: constructing a multi-source sensing data stream, fusing data of a laser radar, a visual sensor and an inertial measurement unit, and generating cooperative sensing data with time-space synchronization; a communication topology network between the robots is established based on the information, and a cooperative relation graph with weights is generated; analyzing the graph by using a space-time graph convolutional network, extracting multi-robot motion mode characteristics, and identifying behavior consistency; a multi-agent reinforcement learning algorithm is adopted, and a multi-machine cooperation strategy adapting to the dynamic environment is generated according to the consistency characteristics; executing the strategy in real time, monitoring the effect, and collecting performance indexes; strategy parameters are updated through an online learning mechanism based on the indexes, and self-adaptive optimization of the system is achieved. The adaptability, stability and efficiency of multi-robot collaborative transportation in indoor and outdoor environments are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent robot control technology, specifically to a multi-robot collaborative system for indoor and outdoor logistics robots. Background Technology

[0002] In hospital logistics scenarios, the precise and efficient delivery of medicines, specimens, and medical supplies is crucial for ensuring the smooth operation of medical processes. Existing technologies typically employ pre-programmed paths or central scheduling systems to enable robots to transport goods within a single indoor area. For example, some solutions rely on laying magnetic strips or QR codes in corridors, with robots traveling along fixed routes and receiving instructions from a central server via a wireless network. For complex tasks requiring movement across floors or connecting inpatient and outpatient departments—indoor-outdoor transition areas—existing technologies often employ segmented scheduling, where robots in different areas work independently, with materials transferred manually or using specific devices at handover points. These existing solutions have significant drawbacks. Robots with pre-programmed paths lack environmental adaptability; when temporary changes occur in the hospital environment, such as temporary construction or dense crowds, the robots cannot autonomously adjust their paths, leading to task interruptions or delays. While central scheduling systems can achieve a certain degree of coordination, their performance heavily relies on a high-quality, uninterrupted communication network. Hospital buildings have complex structures, and indoor and outdoor signals may attenuate or have blind spots. If the central server experiences delays or malfunctions, it can easily cause scheduling chaos or even system paralysis. Segmented scheduling disrupts the continuity of the transportation process. Material transfer not only reduces efficiency but also increases the risk of cross-contamination from human contact. In particular, conventional multi-robot systems struggle to handle the dynamically changing collaborative relationships between robots in hospital settings. During movement, the relative positions and communication links of the robot corps constantly change due to factors such as task allocation and obstacle avoidance; fixed collaborative patterns cannot adapt to this dynamism. Furthermore, the hospital environment requires multiple robots to maintain a high degree of consistency in their movement when collaboratively transporting large or important items to avoid shaking, collisions, or other situations that could damage delicate medical equipment or fragile medications. Existing methods lack the ability to perceive and analyze the fine-grained movement patterns of robot corps, making it difficult to formulate collaborative control strategies adapted to the real-time environment. Summary of the Invention

[0003] The purpose of this invention is to provide a multi-robot collaborative system for indoor and outdoor logistics robots to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides an indoor and outdoor logistics robot multi-robot collaborative system, the system comprising:

[0005] Construct a multi-source perception data stream, collect robot pose information through LiDAR, vision sensors and inertial measurement units, and fuse environmental 3D point cloud data to generate spatiotemporally synchronized collaborative perception data;

[0006] Based on the collaborative perception data, a communication topology network between robots is established, the relative positional relationships of the robots and the data transmission quality are analyzed, and a collaborative relationship graph with connection weights is generated.

[0007] The spatiotemporal graph convolutional network is applied to dynamically analyze the cooperative relationship graph, extract the motion pattern features of multiple robots, and identify the behavioral consistency in the cooperative transportation process.

[0008] A multi-agent reinforcement learning algorithm is used to generate a multi-machine cooperative strategy that adapts to dynamic environments based on the behavioral consistency features.

[0009] The multi-machine collaboration strategy is executed in real time, and the execution effect is monitored through a distributed sensor network to collect collaboration performance indicators.

[0010] Based on the aforementioned collaborative performance indicators, an online learning mechanism is used to update the collaborative strategy parameters, thereby achieving adaptive optimization of the multi-machine collaborative system.

[0011] Preferably, the construction of the multi-source sensing data stream includes:

[0012] The robot's centimeter-level position coordinates are obtained by deploying ultra-wideband positioning base stations in indoor and outdoor environments, while the robot's speed is measured by using an odometer on board.

[0013] A stereo vision camera is used to capture depth information of surrounding obstacles, and an infrared sensor is used to detect ambient lighting conditions.

[0014] Timestamps are applied to data from various sensors, and Kalman filtering is used to fuse the data and eliminate measurement errors.

[0015] Establish a unified coordinate system to transform multi-source data into the same reference frame and generate spatiotemporally synchronized collaborative sensing data.

[0016] Preferably, the data fusion using the Kalman filter algorithm includes:

[0017] Establish the robot's motion state equation, with position, velocity, and acceleration as state variables;

[0018] Design observation equations and use ultra-wideband positioning data and visual odometry readings as observation inputs;

[0019] The state estimate is updated through a predictive-corrective iterative cycle, and the covariance matrix is ​​calculated to reflect the estimation uncertainty.

[0020] When abnormal sensor data is detected, the observation noise parameters are adaptively adjusted to improve fusion robustness.

[0021] Preferably, establishing the inter-robot communication topology network includes:

[0022] Establish an initial communication connection based on the distance threshold between robots, and calculate the signal transmission delay and packet loss rate;

[0023] Analyze network connectivity to identify key relay nodes and vulnerable links in the network;

[0024] Real-time monitoring of communication quality changes and dynamic adjustment of network topology;

[0025] Assign quality weights to each communication link to generate a weighted collaboration graph.

[0026] Preferably, the analysis of network connectivity includes:

[0027] Identifying critical network connections using the minimum cut set algorithm from graph theory;

[0028] Calculate the node betweenness centrality index to assess the importance of the robot in the communication network;

[0029] When a risk of network segmentation is detected, the robot's position is proactively adjusted to maintain network connectivity.

[0030] Establish redundant network paths to improve communication reliability.

[0031] Preferably, the application of the spatiotemporal graph convolutional network for dynamic analysis includes:

[0032] By treating robots as graph nodes and communication links as edges, a spatiotemporal graph data structure is constructed.

[0033] We designed multi-layer graph convolution operations to extract spatial features and combined them with gated recurrent units to capture temporal evolution patterns.

[0034] Neighbor node information is weighted and aggregated using an attention mechanism to highlight important collaborative relationships;

[0035] Output a behavior consistency score for each robot and identify abnormal collaboration patterns.

[0036] Preferably, the step of using a multi-agent reinforcement learning algorithm to generate a multi-machine cooperative strategy adapted to a dynamic environment based on the behavioral consistency features includes:

[0037] Define an intelligent agent for each robot, whose state space is the set of behavioral consistency characteristics and its own sensor data, and whose action space is speed, steering angle and communication switch commands;

[0038] Design a collaborative reward function that simultaneously considers overall task progress, inter-robot collision avoidance distance, and energy consumption efficiency.

[0039] By using a framework of centralized training and decentralized execution, each agent learns a policy network that maximizes collaborative rewards.

[0040] Preferably, the multi-machine cooperation strategy is specified as motion coordination instructions between robots, including:

[0041] Generate speed commands to maintain formation for a swarm of robots that are jointly transporting the same object;

[0042] Generate dynamic avoidance commands based on priority rules for a group of robots moving towards each other in a channel;

[0043] And generate patrol instructions for robots performing area patrol tasks that cover the maximum area and avoid repeating paths.

[0044] Preferably, the monitoring strategy execution effect, and the collection of collaborative performance indicators, include:

[0045] Real-time calculation of task completion rate, average robot travel distance per unit time, average distance deviation between robots, and communication load rate;

[0046] When the task completion rate is consistently below the threshold or the average distance deviation exceeds the safe range, it is determined to be an abnormal collaboration, triggering an emergency update of the collaboration strategy.

[0047] Preferably, the method of updating the collaboration strategy parameters using an online learning mechanism includes:

[0048] The difference between the collected collaborative performance indicators and the expected target is used as the loss function of the policy network.

[0049] An asynchronous advantage actor-critic algorithm is adopted, in which each robot uses locally acquired new experience data to update the gradient of the global policy network in parallel;

[0050] The updated parameters of the global policy network are periodically synchronized to all robots to achieve continuous adaptive optimization of the system.

[0051] Compared with the prior art, the beneficial effects of the present invention are:

[0052] Applying a spatiotemporal graph convolutional network to dynamically analyze collaborative relationship graphs with connection weights enables the modeling of spatiotemporal relationships between robots as a graph structure, thereby capturing the motion dependencies of the group in continuous time series. This network aggregates neighboring robot state information through graph convolution operations, extracting group motion pattern features beyond the individual level and accurately identifying the consistency of robot behavior during collaborative transportation. This method allows the system to quantitatively evaluate collaboration quality, providing a data-driven and refined basis for subsequent policy optimization, overcoming the limitations of traditional methods that rely solely on individual states or simple rules for collaborative judgment. Utilizing a multi-agent reinforcement learning algorithm, collaborative policies are generated based on the identified behavioral consistency features, enabling each robot agent to make distributed decisions based on local observations and group consistency goals. Through continuous trial and error learning through interaction with the environment, the final policy can adaptively handle uncertainties common in hospital environments, such as dynamic obstacles and communication fluctuations. The robots learn to flexibly adjust their actions while maintaining overall formation and transportation stability, achieving collaborative transportation that satisfies both collaborative constraints and individual flexibility in complex scenarios, avoiding the drawbacks of centralized planning, such as high computational load and difficulty in real-time response to emergencies. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the working principle of the multi-machine collaborative indoor and outdoor logistics robot system described in this invention.

[0054] Figure 2 A flowchart for constructing a multi-source sensing data stream;

[0055] Figure 3 A flowchart for establishing an inter-robot communication topology network;

[0056] Figure 4 A trend chart of communication topology network performance indicators;

[0057] Figure 5 A comparison chart of performance metrics before and after optimization of the multi-machine collaboration strategy. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figure 1This invention provides a multi-robot collaborative system for indoor and outdoor logistics. The system includes: acquiring robot pose information via LiDAR, vision sensors, and inertial measurement units, while simultaneously fusing 3D point cloud data of the environment to generate spatiotemporally synchronized collaborative perception data; this process ensures all robots operate based on a unified temporal and spatial reference frame. A communication topology network between robots is established based on the collaborative perception data, analyzing relative positional relationships and data transmission quality to generate a collaborative relationship graph with connection weights to quantify the interaction strength between robots. A spatiotemporal graph convolutional network is applied to dynamically analyze the collaborative relationship graph, extracting multi-robot motion pattern features and identifying behavioral consistency during collaborative transportation, thereby capturing the degree of coordination in group movement. A multi-agent reinforcement learning algorithm is used to generate a multi-robot collaborative strategy adapted to the dynamic environment based on behavioral consistency features, enabling robots to autonomously adjust their actions to cope with changes. During the strategy execution phase, the multi-robot collaborative strategy is deployed in real time, and the execution effect is monitored through a distributed sensor network, collecting key collaborative performance indicators. Based on these performance indicators, an online learning mechanism is used to update the collaborative strategy parameters, achieving continuous adaptive optimization of the system and ensuring that collaborative efficiency continuously improves during long-term operation.

[0060] Example 1: See Figure 2In practical implementation, a multi-source sensing data stream is constructed. Centimeter-level position coordinates of the robot are acquired through ultra-wideband positioning base stations deployed in indoor and outdoor environments. Simultaneously, the robot's onboard odometry measures its movement speed. The ultra-wideband positioning base stations calculate the time difference of arrival using high-frequency pulse signals to achieve high-precision positioning, while the odometry derives displacement through encoder pulse counting, supplementing short-term motion information. A stereo vision camera captures depth information of surrounding obstacles. Based on the principle of binocular parallax, the stereo vision camera calculates pixel depth and combines this with an infrared sensor to detect ambient lighting conditions. The infrared sensor emits a modulated beam and receives reflections, quantifying the ambient light intensity to adapt the exposure parameters of the visual algorithm. Timestamps are applied to various sensor data, generated synchronously by a hardware clock. A Kalman filter algorithm is used for data fusion to eliminate measurement errors. The Kalman filter algorithm establishes the robot's motion state equation, using position, velocity, and acceleration as state variables, which represent the system dynamics in vector form. An observation equation is designed, using ultra-wideband positioning data and visual odometry readings as observation inputs. The observation equation linearizes the nonlinear sensor model. The state estimate is updated through a predictive-corrective iterative cycle. The prediction step derives the prior estimate using the state equation, while the correction step fuses observation data to calculate the posterior estimate. The covariance matrix reflects the estimation uncertainty and quantifies the statistical characteristics of the estimation error. When anomalies are detected in sensor data, the observation noise parameters are adaptively adjusted to improve fusion robustness. Anomaly detection is achieved based on the residual chi-square test. A unified coordinate system is established to transform multi-source data to the same reference frame. This unified coordinate system uses the world coordinate system as a reference and aligns local sensor data through rigid body transformation. Spatiotemporally synchronized collaborative sensing data is generated, with spatiotemporal synchronization ensuring data temporal consistency through interpolation and extrapolation methods.

[0061] In some embodiments, the deployment of ultra-wideband (UWB) positioning base stations follows a grid layout principle, with base station spacing optimized according to environmental dimensions to cover the entire operational area. The robot carries an UWB tag to receive base station signals and calculates its position coordinates using a polygonal positioning method. When measuring speed using odometry, data from a wheel encoder and an inertial measurement unit (IMU) are fused; encoder pulses are converted into displacement, and the IMU provides angular velocity compensation for heading drift. The stereo vision camera captures depth information of surrounding obstacles, using a calibration board to pre-calibrate the camera's intrinsic and extrinsic parameters. Depth calculation is based on epipolar geometry principles, generating a dense depth map. Infrared sensors detect ambient lighting conditions, outputting light intensity values ​​and triggering visual algorithm mode switching, such as enabling image enhancement under low-light conditions. Timestamp marking uses a network time protocol to synchronize all sensor nodes, ensuring microsecond-level accuracy. The state equation of the Kalman filter algorithm is discretized based on a uniform motion model, while the observation equation uses UWB data as position observations and visual odometry as velocity observations. In the prediction-correction loop, the prediction step calculates the state transition, and the correction step applies Kalman gain to minimize the mean square error. The covariance matrix is ​​updated using the Lyapunov equation, and anomaly detection is achieved by monitoring the innovation sequence, increasing the observation noise when the innovation exceeds a threshold. Coordinate system transformation uses Euler angles and a rotation matrix to map the sensor's local coordinate system to the world coordinate system.

[0062] Optionally, the generation of multi-source sensing data streams is implemented in a pipeline manner. Sensor data is first pre-processed and filtered to remove noise before being input into the fusion module. Data packets from the ultra-wideband positioning base station include timestamps and coordinate values; odometry data is output in frequency form; the stereo vision camera generates depth frame sequences; and the infrared sensor provides scalar illumination values. Timestamp marking is triggered by FPGA hardware during data acquisition to ensure global synchronization. The Kalman filter algorithm is implemented in an extended form to handle nonlinear problems. State variables include three-dimensional components of position, velocity, and acceleration. The observation equations linearize the observation models of the ultra-wideband and visual odometry. The iteration frequency of the prediction-correction loop is matched to the sensor sampling rate, typically 100Hz, and the covariance matrix is ​​initialized based on sensor accuracy parameters. When abnormally adjusting observation noise parameters, an exponentially weighted moving average is used to smooth the changes and avoid abrupt changes. A unified coordinate system is established through ground control point calibration, with the world coordinate system origin set at the center of the environment. The spatiotemporally synchronized collaborative sensing data output is a time-aligned multi-dimensional vector for use by subsequent modules.

[0063] It is understandable that constructing a multi-source sensing data stream is the foundation of a collaborative system. Ultra-wideband positioning (UWB) provides a global reference, reducing accumulated errors; odometry supplements local motion details; stereo vision cameras enhance environmental perception; and infrared sensors ensure robustness. Data fusion optimizes estimation through the Kalman filter algorithm, and timestamps and coordinate systems ensure data consistency. In specific implementations, sensor data acquisition is performed in parallel by distributed nodes and aggregated to the central processing unit via Ethernet or wireless networks. When deploying UWB base stations, multipath suppression is considered, and time inversion technology is used to improve accuracy. When the odometry measures movement speed, wheel speed and gyroscope data are fused, and complementary filtering reduces drift. After the stereo vision camera captures depth information, a filtering algorithm is applied to remove outliers and generate a smooth point cloud. Infrared sensors detect lighting conditions and output digital signals to trigger visual parameter adjustments. Timestamp marking adopts the IEEE 1588 protocol to achieve nanosecond-level synchronization. The state equation of the Kalman filter algorithm considers motion constraints, such as maximum acceleration limits, and the observation equation fuses the weights of multi-source observations based on the signal-to-noise ratio. In the prediction-correction loop, the prediction step uses the state transition matrix, and the correction step calculates the Kalman gain to update the state. The covariance matrix update reflects the estimated confidence level, and anomaly detection is achieved by monitoring the Mahalanobis distance. Coordinate system transformation uses quaternions to represent rotations to avoid gimbal lock. Spatiotemporally synchronized collaborative sensing data is published in the form of a message queue for downstream modules to subscribe to.

[0064] In practical implementation, the construction of multi-source sensing data streams emphasizes modular design. Ultra-wideband (UWB) positioning base stations, odometers, stereo vision cameras, and infrared sensors are treated as independent modules, outputting standardized data packets. The UWB base station calculates distance using the TOA algorithm, the odometer integrates velocity to obtain position, the stereo vision camera performs stereo matching to generate a disparity map, and the infrared sensor converts analog signals using an ADC. A clock count is embedded in the data packet header using a timestamp, the Kalman filter algorithm is implemented as a recursive filter, state variables are stored using vectors, and the observation equations are linearized using a Jacobian matrix. The prediction-correction loop runs in real-time on the embedded system, and the covariance matrix uses Cholesky decomposition to ensure numerical stability. When abnormally adjusting observation noise parameters, noise statistics are learned based on historical data. A unified coordinate system aligns the sensor frame through hand-eye calibration, and spatiotemporal synchronization uses interpolation algorithms to fill in data gaps. After generating collaborative sensing data, data integrity is verified through a checksum mechanism.

[0065] Example 2: See Figure 3In practical implementation, establishing an inter-robot communication topology involves establishing initial communication connections based on a distance threshold between robots. This distance threshold is set based on the effective transmission range of the wireless communication module. For example, in an open environment, a free-space path loss model is used to calculate the maximum connectable distance. When the distance between robots is less than the threshold, a communication link is attempted to be established. Signal transmission delay and packet loss rate are calculated. Delay measurement is achieved by sending synchronization timestamp messages and calculating the round-trip time difference. Packet loss rate statistics are obtained by continuously sending test data packets and calculating the loss ratio. Network connectivity is analyzed to identify key relay nodes and vulnerable links. The minimum cut set algorithm in graph theory is applied to identify key network connections. The minimum cut set algorithm traverses all possible edge cut sets in the graph to find the cut with the minimum weight. The system uses a set of parameters to determine which edge removals would lead to network segmentation. It calculates the node betweenness centrality index, which assesses the robot's importance in the communication network based on the proportion of nodes whose shortest paths between all node pairs pass through that node. It monitors communication quality changes in real time, detecting real-time fluctuations in latency and packet loss rate through periodic broadcast heartbeat messages. The system dynamically adjusts the network topology; for example, it automatically disconnects a link and searches for an alternative path when the packet loss rate of a link consistently exceeds a threshold. It assigns quality weights to each communication link, calculating weights that consider latency, packet loss rate, and bandwidth factors, and uses a weighted average formula to generate normalized weight values. This generates a weighted cooperative relationship graph, which is stored as an adjacency matrix or adjacency list for subsequent graph analysis algorithms. In some embodiments, the distance threshold setting considers the impact of environmental obstacles, uses ray tracing to simulate signal attenuation, and adjusts the threshold to adapt to non-line-of-sight conditions. The latency and packet loss rate calculation uses a sliding window averaging method to smooth short-term fluctuations and improve measurement stability. The heartbeat message interval is dynamically adjusted based on network stability. Optionally, communication link establishment supports multiple wireless protocols, selecting the optimal protocol based on the scenario. Latency measurement compensates for clock drift using NTP synchronization. Packet loss statistics distinguish between temporary and permanent faults. The minimum cut set algorithm considers edge weights when processing weighted graphs. Node betweenness centrality is normalized to the 0-1 range for easy comparison. High-weight links are prioritized during dynamic topology adjustments. The weight allocation formula is configurable to adapt to different application requirements. In essence, establishing a communication topology network is fundamental to a collaborative system. Distance thresholds ensure connection reliability, latency and packet loss quantify link quality, the minimum cut set algorithm and node betweenness centrality provide network robustness analysis, real-time monitoring and dynamic adjustment respond to environmental changes, and weighted collaborative relationship graphs support refined collaborative decision-making.

[0066] In practical implementation, the process of analyzing network connectivity deeply applies graph theory algorithms. When the minimum cut set algorithm is executed, it first constructs a graph model of the communication network, where nodes represent robots and edges represent communication links. Edge weights are initialized based on quality weights. The minimum cut set algorithm iteratively merges nodes until the graph is reduced to two super nodes, recording the minimum cut value and its corresponding edge set during the merging process. Key connection points are identified, corresponding to the robots connected by these edges. The number of times each node acts as an intermediary node is counted; a higher node betweenness centrality value indicates that the robot undertakes more relay tasks in the network. When a network segmentation risk is detected, risk detection is achieved by monitoring changes in the minimum cut value. If the minimum cut value is below a threshold, it indicates network fragility, and the robot's position is proactively adjusted to maintain network connectivity. Position adjustment instructions are generated based on network connectivity optimization algorithms, such as moving key robots to reduce the impact of cut sets, establishing redundant paths in the network, and using the k-shortest path algorithm to find multiple non-intersecting paths to improve communication reliability. In some embodiments, the minimum cut set algorithm supports incremental updates when processing dynamic graphs, avoiding full graph recalculation. Node betweenness centrality calculation optimization uses approximation algorithms to reduce computational complexity, making it suitable for large-scale networks. Risk detection sets multi-level thresholds to distinguish between warnings and emergencies. Position adjustment incorporates robot motion constraints, such as maximum speed and safe distance. Redundant path establishment considers minimizing the sum of path weights. Optionally, the minimum cut set algorithm can be extended to handle directed graphs. Node betweenness centrality distinguishes between in-degree and out-degree centrality. Risk detection integrates machine learning models to predict splitting probabilities. Position adjustment coordinates multi-robot collaborative movement. Redundant path dynamic activation is based on real-time load balancing. It can be understood that network connectivity analysis ensures the collaborative system can still operate during partial failures, the minimum cut set algorithm identifies weak points, node betweenness centrality guides resource allocation, proactive adjustments prevent network outages, and redundant paths provide a fault recovery mechanism.

[0067] In practical implementation, real-time monitoring of communication quality changes is achieved by deploying a distributed monitoring agent. Each robot runs a monitoring process, periodically sending probe messages to neighboring robots. These messages include a sequence number and a sending timestamp. The receiver records the arrival time and calculates the latency, statistically analyzes message loss, and calculates the packet loss rate. Monitoring data is aggregated to a central server or shared via a distributed consensus protocol. Dynamic adjustments to the network topology are triggered by monitoring data. Adjustment strategies include adding new links, removing degraded links, or modifying the routing table. Adjustment decisions utilize the principle of minimizing distance in graph editing to reduce oscillations caused by topology changes. The process of assigning quality weights to each communication link involves inputting latency, packet loss rate, and available bandwidth into the weight calculation function. Latency and packet loss rate are processed using an exponentially weighted moving average filter to process historical data. Bandwidth is measured through throughput testing. Weight values ​​are normalized to the 0-1 range. When generating a weighted collaboration graph, the graph structure is stored in a graph database, supporting fast querying and updating. The collaboration graph is output in a standard graph format for use by downstream modules. In some embodiments, real-time monitoring supports multi-granularity monitoring: coarse-grained monitoring for routine maintenance and fine-grained monitoring for fault diagnosis; dynamic adjustment introduces a lag mechanism to avoid frequent switching; weight allocation supports custom weight formulas to adapt to specific scenarios; and a collaboration graph visualization tool assists in debugging. Optionally, monitoring data is transmitted with encryption to ensure security; the dynamic adjustment strategy can be configured to manual or automatic mode; weight calculation integrates user feedback adjustment parameters; and the collaboration graph export interface supports various analysis tools. In essence, real-time monitoring provides system status awareness, dynamic adjustment maintains optimal network performance, weight allocation quantifies link value, and the collaboration graph is the fundamental input for the collaboration strategy.

[0068] In practical implementation, the overall process of establishing the inter-robot communication topology network operates in an automated manner. During the initialization phase, robots broadcast their own position and capability information, establish an initial connection matrix based on distance thresholds, record reachability between robots, and calculate initial latency and packet loss rate baselines. During network connectivity analysis, the minimum cut set algorithm and node betweenness centrality calculation are periodically executed, with results cached for quick retrieval. During real-time monitoring, background tasks continuously run to collect network metrics. During dynamic adjustment, event-driven processes are implemented, invoking adjustment algorithms when performance degradation is detected. During weight allocation, weight values ​​are periodically recalculated, and the collaboration graph is updated in real-time to ensure data freshness. In some embodiments, process optimization uses parallel computing to accelerate graph algorithms, distributed storage of network state data, fault tolerance mechanisms to handle robot joining or leaving, and integrated process logging for auditing and optimization. Optionally, the process supports pause and resume operations, weight allocation allows external input priority settings, and collaboration graph version management tracks historical changes. In essence, this process achieves fully automated network management, reduces manual intervention, and improves system reliability.

[0069] In practical implementation, the application scenarios of the communication topology network include indoor warehouses and outdoor sites. Indoor environments consider multipath effects and have lower distance thresholds, while outdoor environments consider weather impacts. Delay measurement compensates for propagation delay, the minimum cut set algorithm handles sparse graphs, node betweenness centrality focuses on hub nodes, real-time monitoring adapts to dynamic obstacles, dynamic adjustment rapidly responds to terrain changes, weight allocation balances latency and reliability, and the collaborative relationship graph supports large-scale robot swarms. In some embodiments, scenario-specific parameters are loaded via configuration files, and network optimization objectives are adjusted according to different tasks, such as prioritizing real-time performance or energy consumption. Optionally, scenario simulation uses a simulation environment to verify the algorithm, and stress testing is performed before actual deployment. It is understood that scenario adaptability ensures the method's broad applicability.

[0070] In practical implementation, the software architecture for the communication topology network adopts a microservice design, with each functional module deployed independently, such as a connection management module, monitoring module, analysis module, and tuning module. Modules communicate via message queues, data flow standardization uses Protocol Buffers serialization, hardware supports various communication chips, and the software runs cross-platform on Linux or ROS systems. In some embodiments, the architecture supports horizontal scaling, modules can be deployed in a distributed manner, data consistency is guaranteed using distributed transactions, and performance monitoring integrates Prometheus metric collection. Optionally, the architecture provides a REST API for external integration, and configuration management uses environment variables or configuration files. It can be understood that the software architecture improves system maintainability and scalability. See also... Figure 4 This figure is a dual-axis line graph, with runtime on the horizontal axis and communication latency and packet loss rate on the left and right axes respectively. It visually presents the performance evolution of the communication topology network in a multi-robot collaborative logistics system. The graph verifies the effectiveness of the technical solution of real-time monitoring of communication quality changes and dynamic adjustment of the network topology. By continuously analyzing indicators such as signal transmission latency and packet loss rate, the system dynamically optimizes the network topology, ultimately achieving a significant improvement in communication performance. This provides low-latency, highly reliable communication support for multi-robot collaboration, ensuring the real-time performance and stability of data interaction between robots, and is a key communication foundation for the efficient operation of multi-robot collaborative systems.

[0071] Example 3: In specific implementation, the process of dynamically analyzing the collaborative relationship graph using a spatiotemporal graph convolutional network is described. Robots are treated as graph nodes, and communication links as edges. A spatiotemporal graph data structure is constructed, which is a composite data structure containing node feature matrices, edge connection matrices, and time series tensors. The node feature matrix stores the real-time state information of each robot, including position coordinates, velocity vectors, and attitude angles. The edge connection matrix represents the existence and quality weights of communication links between robots in the form of an adjacency matrix. The time series tensor records historical node features and edge connection states at fixed time steps, thus comprehensively describing the evolution of the system in the spatiotemporal dimension. Multi-layer graph convolutional operations are designed to extract spatial features. Each layer of graph convolutional operation updates the representation of the central node by aggregating neighbor node information. In the specific implementation, Chebyshev polynomial approximation graph convolution kernels are used to reduce computational complexity. Gated recurrent units are combined to capture the temporal evolution patterns. A gated recurrent unit is a variant of a recurrent neural network that controls the retention and forgetting of historical information through update and reset gates, enabling it to effectively learn temporal dependencies. The model employs an attention mechanism to weighted aggregate neighbor node information, highlighting important collaborative relationships. This mechanism calculates relevance scores between nodes and dynamically assigns aggregation weights based on node feature similarity and edge weights, enabling the model to focus on collaborative interactions that have a greater impact on the current task. Each robot outputs a behavior consistency score, a scalar value. A fully connected layer maps the final node representation to the score space, identifying abnormal collaborative patterns. Anomaly detection is achieved by setting a score threshold or using statistical process control methods; an alarm is triggered when a robot's behavior consistency score consistently deviates from the group mean.

[0072] In some embodiments, the spatiotemporal graph data structure employs a sliding window mechanism to maintain recent data. The window size is set based on the task cycle and computational resource balance. The dimension of the node feature matrix is ​​the number of robots multiplied by the feature dimension, which includes basic motion states and extended context information. The edge connection matrix supports dynamic updates to reflect topological changes. The time series tensor is organized as a three-dimensional array for easy batch processing. The number of layers in the multi-layer graph convolution operation is configured according to the network size and complexity, typically 2 to 3 layers. The output feature dimension of each layer is progressively reduced to extract higher-order features. Chebyshev polynomial approximation expands the graph convolution kernel into a K-order polynomial, avoiding feature decomposition and improving computational efficiency. The time step of the gated recurrent unit is synchronized with the data sampling interval. The hidden state dimension matches the graph convolution output features. The update gate controls the proportion of historical states incorporated, and the reset gate determines the degree of influence of the current input on the state update. The attention mechanism is implemented as a multi-head self-attention mechanism. Multiple attention heads compute in parallel and then concatenate the results to enhance the model's expressive power. The correlation score is calculated using additive attention or dot product attention methods. After the behavioral consistency score is output, it is normalized to the 0-1 range using the Sigmoid function. Abnormal collaboration pattern recognition is combined with time consistency checks to avoid false alarms due to instantaneous fluctuations.

[0073] In some embodiments, the training process of the spatiotemporal graph convolutional network uses supervised learning. Training data comes from historical task records or simulation generation, and label data consists of expert-annotated normal or abnormal collaborative patterns. The loss function combines behavioral consistency scoring error and topology preservation constraints, and the optimization algorithm uses Adam or stochastic gradient descent. The model inference stage is deployed on edge computing devices, inputting real-time spatiotemporal graph data streams. The output cycle matches the control system's decision cycle, and behavioral consistency scores are pushed to the decision module in real time. Optionally, the model supports an online learning mechanism, adapting to environmental changes through incremental updates. Modular design facilitates the integration of different sensor modal data. The key operations for spatial feature extraction in the spatiotemporal graph convolutional network are described by the following formula:

[0074]

[0075] in: Indicates the first The feature matrix of all nodes in the layer, This represents the feature matrix of all nodes in the (l+1)th layer of the neural network, which is the output obtained after computation by the current layer. It is a normalized graph Laplace matrix. It is the first Chebyshev polynomials It is the maximum order of the polynomial expansion. It is the first Layer A trainable parameter matrix of order 1. This represents a non-linear activation function.

[0076] It is understandable that the method of applying spatiotemporal graph convolutional networks for dynamic analysis, by combining the spatial modeling capabilities of graph convolution with the temporal modeling capabilities of gated recurrent units, can effectively extract collaborative behavior features from complex multi-robot interaction data. The attention mechanism enhances the model's sensitivity to key collaborative relationships, and the behavior consistency score provides a quantitative basis for subsequent decision-making. The entire architecture balances computational efficiency and model expressive power, making it suitable for real-time collaborative systems. In specific implementation, the implementation of multi-layer graph convolution operations involves the construction and normalization of the graph Laplacian matrix. The graph Laplacian matrix is ​​obtained by subtracting the adjacency matrix from the degree matrix. The normalization process uses symmetric normalization to enhance numerical stability, and Chebyshev polynomial recursive calculation avoids explicit eigenvalue decomposition. The polynomial coefficients are obtained through training. The temporal evolution modeling of the gated recurrent unit takes the graph convolution output sequence as input, initializes the hidden state as a zero vector, and updates the hidden state iteratively through time steps. Finally, the hidden state encodes the temporal evolution law. During the weighted aggregation of the attention mechanism, the attention coefficients between each pair of nodes are calculated. The coefficients are normalized using the Softmax function, and the weighted summation generates attention-enhanced node representations. The behavioral consistency scoring output layer adopts a fully connected structure, and the weights are optimized through training. The abnormal collaboration pattern recognition module continuously monitors the scoring sequence and uses a sliding window statistical method to detect outliers.

[0077] In some embodiments, model hyperparameters such as the number of layers, feature dimensions, and number of attention heads are determined through cross-validation. Training data augmentation employs random dropout and noise injection to improve generalization ability, while model compression techniques such as pruning and quantization are used for deployment in resource-constrained scenarios. Real-time inference optimization reduces latency through operator fusion and memory reuse, and distributed inference supports multi-robot collaborative computation. It is understood that this implementation prioritizes practical deployment feasibility, balancing model complexity and computational overhead to ensure real-time system responsiveness. The integration of the spatiotemporal graph convolutional network with the multi-robot system is achieved through a standardized interface. Input data preprocessing includes feature normalization and missing value imputation, while output postprocessing includes score smoothing and anomaly detection. The system monitoring module records model performance metrics and periodically retrains to maintain model accuracy. Model updates support hot-swapping to avoid service interruptions, and version management tracks model iteration history.

[0078] Example 4: In specific implementation, when generating multi-machine cooperative strategies using multi-agent reinforcement learning algorithms, an agent is defined for each robot. An agent is an independent decision-making entity whose state space is a set of behavioral consistency features and its own sensor data. Behavioral consistency features, derived from the output of a spatiotemporal graph convolutional network, represent the degree of cooperation between the robot and the group. Its own sensor data includes LiDAR point clouds, visual recognition results, inertial measurement unit readings, and current position and velocity information. The state space is a high-dimensional vector, normalized to accelerate training. Its action space includes velocity, steering angle, and communication switch commands. Velocity commands include set values ​​for linear and angular velocities. The steering angle controls the robot's direction of travel. Communication switch commands manage the activation and shutdown of the wireless communication module to save energy. A cooperative reward function is designed, which is a scalar value that considers overall task progress, inter-robot collision avoidance distance, and energy consumption efficiency. Overall task progress is calculated based on the proximity of the target point or the completion rate of item transportation. The inter-robot collision avoidance distance is penalized based on the minimum Euclidean distance between robots. Energy consumption efficiency is modeled based on motor power consumption and communication energy consumption. Through a framework of centralized training and decentralized execution, each agent learns a policy network that maximizes collaborative rewards. During the centralized training phase, the experience data of all agents is stored in a shared replay buffer for jointly updating a global policy network and value network. During the decentralized execution phase, each agent makes independent decisions based on local observations and generates actions using the trained policy network.

[0079] Referring to Table 1, the multi-machine cooperation strategy is specified as motion coordination instructions between robots. These instructions are setpoints of the low-level controller, including: generating speed instructions to maintain formation for a swarm of robots jointly transporting the same object (the formation can be linear, circular, or a custom configuration), with the speed instructions generated by a PID controller tracking the desired relative position and orientation; generating dynamic avoidance instructions based on priority rules for a group of robots moving towards each other in a channel (priority rules are dynamically assigned according to the robots' task urgency, ID sequence, or distance from the target point), with avoidance instructions including temporary stopping, deceleration, or moving to the side; and generating patrol instructions for robots performing area patrol tasks that cover the maximum area and avoid repeated paths (patrol instructions are generated based on grid maps or graph theory path planning algorithms to ensure efficient area coverage). The policy network is typically represented by a deep neural network, with the input being a state vector and the output being a probability distribution or deterministic action value in the action space. The training algorithm can be a deep deterministic policy gradient, a multi-agent deep deterministic policy gradient, or a variant thereof.

[0080] Table 1: Components of the Collaborative Reward Function

[0081]

[0082] In some embodiments, the state space is constructed with temporal information, such as using a stack of states from the most recent time steps as input to the policy network to capture dynamic characteristics. The action space implementation may include a mixture of continuous and discrete actions; for example, speed commands are continuous values, while communication switches are discrete binary commands. The weight coefficients of each part of the cooperative reward function need to be fine-tuned through extensive simulation experiments to achieve the optimal balance between task completion efficiency, safety, and energy consumption. The experience replay buffer in the centralized training and distributed execution framework employs a priority sampling mechanism to replay samples with high learning value more frequently. The policy network for each agent can be designed as a homogeneous network with shared parameters to reduce model complexity and training costs, especially in large-scale robot swarms. The cooperative reward function can introduce additional reward terms, such as rewards to encourage exploration of unknown regions or rewards to penalize drastic actions to improve motion smoothness. The priority rules for dynamic avoidance commands can be designed in a learnable form, with priorities dynamically evaluated and assigned through an auxiliary network module.

[0083] Optionally, to improve the generalization ability of the strategy, domain randomization techniques can be introduced during training, such as randomly changing the friction coefficient of the environment, the robot's dynamic parameters, or the noise model of the sensors. The generation of patrol instructions can be combined with topological methods to divide the patrol area into multiple sub-regions and solve for the optimal access sequence based on the traveling salesman problem. A method for generating multi-machine cooperative strategies using multi-agent reinforcement learning algorithms defines each robot as an agent and designs a cooperative reward function that comprehensively considers global objectives. This allows each robot to learn an efficient cooperative strategy through centralized training, relying only on local observations. The framework of centralized training and distributed execution effectively solves the learning challenges in non-stationary environments. Concretizing the policy output into motion coordination instructions for various scenarios gives the method practicality and flexibility for handling complex indoor and outdoor logistics tasks. The reward function explicitly listed in the table provides a clear and quantifiable optimization objective for policy learning.

[0084] Example 5: In specific implementation, the process of monitoring the execution effect of the strategy and collecting collaborative performance indicators is achieved through a distributed monitoring framework. Each robot, acting as a monitoring agent, continuously collects local data and uploads it to the central analysis node. Real-time calculations are performed on task completion rate, average robot distance traveled per unit time, average distance deviation between robots, and communication load rate. The task completion rate is defined as the ratio of the number of completed sub-tasks to the total number of sub-tasks within the evaluation period. The average robot distance traveled per unit time is obtained by summing the odometer readings of all robots and dividing by the number of robots and the running time. The average distance deviation between robots is calculated as the average of the absolute differences between the actual distance between all robot pairs and the expected formation distance. The communication load rate is statistically calculated as the percentage of the actual data transmitted by the communication module per unit time to the maximum theoretical bandwidth. When the task completion rate is detected to be consistently below the threshold or the average distance deviation exceeds the safe range, it is determined to be a collaborative anomaly, triggering an emergency update of the collaborative strategy. The threshold is dynamically set based on historical performance data and task requirements; for example, the task completion rate threshold may be appropriately lowered in complex environments to avoid misjudgment. The safe range is determined based on the robot's physical dimensions and braking distance. The emergency update mechanism of the collaborative strategy adopts a high-priority interruption method, pausing the execution of the current strategy and loading a backup strategy.

[0085] In some embodiments, performance metrics are calculated using a sliding time window method. The window size is configured according to the system response speed requirements, typically set to several minutes to tens of minutes. Metric data is filtered to eliminate instantaneous fluctuations. Anomaly detection algorithms combine threshold comparison and trend analysis; a task completion rate consistently below the threshold indicates that the metric fails to meet the standard for multiple consecutive calculation windows, with the average deviation exceeding the safe range, taking into account both the duration and magnitude of the deviation. Emergency update processes include a rollback mechanism, allowing for rapid reversion to the previous stable version if performance does not improve after the update.

[0086] When updating the collaboration strategy parameters using an online learning mechanism, the difference between the collected collaboration performance indicators and the expected goals is used as the loss function of the strategy network. The expected goals are pre-set according to the task planning, such as a task completion rate target of 98% and an average distance deviation target of 0.5 meters. The difference is calculated using mean squared error or absolute error. An asynchronous dominant actor-critic algorithm is employed, where each robot uses locally collected new experience data to update the gradient of the global strategy network in parallel. The asynchronous dominant actor-critic algorithm includes the actor network generating actions and the critic network evaluating values; the advantage function estimate is used to reduce variance. The updated parameters of the global strategy network are periodically synchronized to all robots to achieve continuous adaptive optimization of the system. The synchronization period is adjusted according to network bandwidth and computing resources, typically set within the range of several hours to several days.

[0087] Optionally, performance metric collection supports custom expansion, allowing users to add new monitoring metrics based on specific scenarios, such as battery power balance or task queue length. Anomaly detection rules can be configured with multiple conditions, such as triggering an emergency update only when both low task completion rate and high communication load are met simultaneously. The loss function design can incorporate a weighted mechanism, assigning different weights to different performance metrics to reflect priority. The asynchronous advantage actor-critic algorithm can be implemented using a multi-worker architecture, with each worker independently interacting with the environment and calculating gradients. Parameter synchronization supports differential synchronization, transmitting only changed parameters to reduce network load. Monitoring data storage uses a time-series database for easy long-term trend analysis and fault diagnosis. After an emergency update is triggered, the system can enter a safe mode, limiting the robot's maximum speed until strategy optimization is complete. The online learning process includes exploring strategies, encouraging strategy diversity by adding noise to actions or using random sampling. Global policy network parameter synchronization employs version control, ensuring all robots within the cluster use the same policy version.

[0088] Optionally, the system provides a visual dashboard to display various performance indicators in real time, assisting maintenance personnel in understanding the system status. Detailed logs are recorded for collaborative anomalies, including the time of occurrence, associated robot identifiers, and snapshots of key indicators. The gradient update process includes a gradient pruning mechanism to prevent gradient explosion that could lead to training instability. The parameter synchronization protocol is fault-tolerant, supporting continued synchronization even when some robots are offline. Monitoring the effectiveness of the strategy execution and using an online learning mechanism to update collaborative strategy parameters constitute a complete adaptive optimization closed loop. By continuously monitoring key performance indicators and promptly detecting collaborative anomalies, the system can dynamically adjust its collaborative strategy to cope with environmental changes and internal state fluctuations. Using the difference between performance indicators and the target as a loss function provides a clear direction for strategy optimization. The asynchronous advantage actor-critic algorithm fully utilizes distributed computing resources to accelerate the learning process, and periodic parameter synchronization ensures the consistency of the entire robot cluster's behavior. This implementation approach enables the multi-machine collaborative system to possess self-optimization and fault recovery capabilities during long-term operation.

[0089] In practical implementation, consider a specific example: a multi-robot system performs cargo handling tasks in a warehouse environment. The monitoring module calculates the task completion rate in real time, which is the proportion of boxes successfully delivered to their destination out of the total task volume. The average distance traveled by the robots per unit time is obtained by analyzing the odometer data of each robot, reflecting the overall operational efficiency. The average distance deviation between robots is particularly important when robots are moving in formation, used to evaluate the accuracy of formation maintenance. The communication load rate monitors the usage of the wireless network. When the system detects that the task completion rate is below the 85% threshold for three consecutive sampling periods due to newly added obstacles, or that the average distance deviation of the robot formation is continuously exceeded within the safe range of 0.8 meters due to slippery ground, the monitoring module immediately marks the collaboration as abnormal. This event triggers a policy update process, the system suspends the current task, and activates the emergency response mechanism. The online learning mechanism is then activated. The loss function is calculated based on the difference between the recently collected performance data and the preset target value. Each robot runs the asynchronous advantage actor-critic algorithm in parallel and calculates the policy gradient using locally stored interaction experience. These gradients are aggregated to update the global policy network parameters. The updated parameters are distributed to the entire robot cluster in the next synchronization cycle, thereby optimizing subsequent collaborative behaviors, such as adjusting path planning strategies or improving formation control algorithms.

[0090] See Figure 5 This chart, a grouped bar chart, uses task completion rate, distance traveled per unit time, average distance deviation, and communication load rate on the horizontal axis and performance index values ​​on the vertical axis. By comparing the data before and after optimization, it visually presents the effect of multi-machine collaboration strategy optimization. The chart verifies the effectiveness of the technical solution of monitoring collaboration performance indicators and adopting an online learning update strategy. Through continuous monitoring and strategy iteration of key indicators such as task completion rate, motion efficiency, collaboration accuracy, and communication cost, the multi-machine collaboration system achieves comprehensive performance optimization, forming an adaptive closed loop of monitoring-analysis-optimization-execution, providing strong support for the efficient collaborative operation of logistics robots in complex indoor and outdoor scenarios.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for multi-machine cooperation of indoor and outdoor logistics robots, characterized in that, The implementation steps include: Constructing multi-source perception data streams, collecting robot pose information through laser radar, visual sensor and inertial measurement unit, and fusing environment three-dimensional point cloud data to generate spatiotemporal synchronous cooperative perception data; Based on the cooperative perception data, a robot-to-robot communication topology network is established, the relative position relationship and data transmission quality of the robots are analyzed, and a cooperative relationship graph with connection weights is generated; A spatiotemporal graph convolution network is applied to dynamically analyze the cooperative relationship graph, extract multi-robot motion mode features, and identify behavior consistency in the collaborative transportation process; A multi-agent reinforcement learning algorithm is used to generate a multi-robot cooperation strategy that adapts to the dynamic environment according to the behavior consistency features; The multi-robot cooperation strategy is executed in real time, and the strategy execution effect is monitored through a distributed sensor network to collect cooperation performance indicators; Based on the cooperation performance indicators, an online learning mechanism is used to update the cooperation strategy parameters to achieve adaptive optimization of the multi-robot cooperation system. 2.The method of claim 1, wherein, The construction of multi-source perception data streams includes: Obtain the robot's centimeter-level position coordinates through ultra-wideband positioning base stations deployed in indoor and outdoor environments, and measure the moving speed using the robot-mounted odometer; Capture the depth information of surrounding obstacles using stereo vision cameras, and detect the environmental lighting conditions using infrared sensors; Apply time stamp labels to various sensor data, and use Kalman filtering algorithm for data fusion to eliminate measurement errors; Establish a unified coordinate system to convert multi-source data to the same reference frame to generate spatiotemporal synchronous cooperative perception data. 3.The method of claim 2, wherein, The data fusion using Kalman filtering algorithm includes: Establish the robot motion state equation, taking position, velocity, and acceleration as state variables; Design the observation equation, taking ultra-wideband positioning data and visual odometer readings as observation inputs; Update the state estimate value through the prediction-correction cycle iteration, and calculate the covariance matrix to reflect the estimation uncertainty; When detecting abnormal sensor data, adaptively adjust the observation noise parameters to improve the robustness of fusion. 4.The method of claim 1, wherein, The establishment of the robot-to-robot communication topology network includes: Establish initial communication connections based on the distance threshold between robots, calculate the signal transmission delay and packet loss rate; Analyze network connectivity, identify key relay nodes and network weak links; Real-time monitoring of communication quality changes, dynamic adjustment of network topology structure; Assign quality weights to each communication link to generate a weighted cooperative relationship graph.

5. The multi-machine cooperation method for indoor and outdoor logistics robots according to claim 4, characterized in that, The analysis of network connectivity includes: Apply the minimum cut set algorithm in graph theory to identify key connection points in the network; Calculate the node betweenness centrality index to evaluate the importance of robots in the communication network; When detecting network partitioning risks, actively adjust the robot positions to maintain network connectivity; Establish network redundancy paths to improve communication reliability. 6.The method of claim 1, wherein, The dynamic analysis using spatiotemporal graph convolution network includes: Treat robots as graph nodes and communication links as edges to construct a spatiotemporal graph data structure; Design multi-layer graph convolution operations to extract spatial features, and combine gated recurrent units to capture temporal evolution rules; Weight and aggregate neighbor node information through attention mechanism to highlight important cooperative relationships; Output the behavior consistency score of each robot to identify abnormal cooperative patterns.

7. The multi-machine cooperation method for indoor and outdoor logistics robots according to claim 1, characterized in that, The multi-agent reinforcement learning algorithm is used to generate a multi-robot cooperation strategy adapted to a dynamic environment according to the behavior consistency features, and the method comprises the following steps: An agent is defined for each robot, and the state space of the agent is a set of the behavior consistency features and sensor data of the robot, and the action space of the agent is a speed, a steering angle and a communication switch instruction; A collaborative reward function is designed, and the collaborative reward function simultaneously considers a whole task progress, a robot-to-robot anti-collision distance and an energy consumption efficiency; Through a centralized training and decentralized execution framework, each agent learns a strategy network maximizing the collaborative reward. 8.The method of claim 7, wherein, The multi-robot cooperation strategy is embodied in motion coordination instructions between robots, and the instructions comprise the following steps: A speed instruction is generated for a robot cluster transporting a same object to maintain a formation; A dynamic avoidance instruction based on a priority rule is generated for a robot group moving towards each other in a channel; And a patrol instruction is generated for a robot performing an area patrol task to cover a maximum range and avoid a repeated path. 9.The method of claim 1, wherein, The cooperation performance indicators are collected by monitoring a strategy execution effect, and the indicators comprise the following steps: A task completion rate, an average robot travel distance per unit time, an average distance deviation between robots and a communication load rate are calculated in real time; When it is detected that the task completion rate continuously falls below a threshold value or the average distance deviation exceeds a safety range, it is determined that a cooperation is abnormal, and an emergency update of the cooperation strategy is triggered.

10. The multi-machine cooperation method for indoor and outdoor logistics robots according to claim 9, characterized in that, The cooperation strategy parameters are updated by using an online learning mechanism, and the parameters comprise the following steps: A difference between the collected cooperation performance indicators and expected targets is used as a loss function of the strategy network; An asynchronous advantage actor-critic algorithm is used, and each robot uses newly collected experience data to perform gradient updates on a global strategy network in parallel; Parameters of the global strategy network after the updates are regularly synchronized to all robots, and continuous adaptive optimization of the system is realized.

Images