An indoor and outdoor logistics robot multi-machine cooperation method

Abstract

The application relates to the technical field of intelligent robot control, and discloses a multi-machine cooperation method for indoor and outdoor logistics robots. The method comprises the following steps: constructing a multi-source perception data stream, fusing laser radar, visual sensor and inertial measurement unit data, and generating time-space synchronous cooperative perception data; based on the same, a communication topology network among robots is established, and a cooperative relationship graph with weights is generated; a time-space graph convolution network is applied to analyze the graph, extract multi-robot motion mode features, and identify behavior consistency; a multi-agent reinforcement learning algorithm is adopted to generate a multi-machine cooperation strategy suitable for a dynamic environment according to the consistency features; the strategy is executed in real time, and the effect is monitored; performance indicators are collected; based on the indicators, strategy parameters are updated through an online learning mechanism, and adaptive optimization of the method is realized. The application effectively improves the adaptability, stability and efficiency of multi-robot cooperative transportation in indoor and outdoor environments.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent robot control, in particular to a multi-machine cooperation method for indoor and outdoor logistics robots. BACKGROUND

[0002] In the hospital logistics scenario, the accurate and efficient delivery of medicines, specimens and medical supplies is a key link to ensure the smooth operation of the medical process. The existing technology usually uses pre-programmed paths or central dispatch systems to realize the single indoor area transportation of robots. For example, some solutions rely on laying magnetic strips or two-dimensional codes in the corridors, and the robots travel along fixed routes and receive instructions from the central server through a wireless network. For complex tasks that need to cross floors and connect indoor and outdoor transition areas such as inpatient and outpatient departments, the existing technology usually uses a segmented dispatch, that is, robots in different areas work independently, and the transfer of goods at the handover point is completed by manual or special devices. These existing technical solutions have obvious defects. The pre-programmed path robot lacks environmental adaptability. When the hospital environment changes temporarily, such as temporary construction and crowded crowds, the robot cannot adjust the path autonomously, resulting in task interruption or delay. Although the central dispatch system can achieve a certain degree of cooperation, its performance is severely dependent on a high-quality and uninterrupted communication network. The hospital building structure is complex, and the indoor and outdoor signals are attenuated or blind. Once the central server is delayed or fails, it can easily cause the dispatch system to be chaotic or even paralyzed. The segmented dispatch destroys the continuity of the transportation process, and the goods transfer link not only reduces the efficiency, but also increases the cross-contamination risk caused by human contact. In particular, conventional multi-robot systems are difficult to handle the dynamic changes of the cooperation relationship between robots in the hospital scenario. The relative position and communication link state of the robot team change all the time due to factors such as task allocation and obstacle avoidance. The fixed cooperation mode cannot adapt to this dynamics. In addition, the hospital environment requires multiple robots to maintain high consistency in their movement behavior when cooperating to transport large or important goods, in order to avoid situations that may cause damage to precision medical equipment or fragile medicines, such as shaking and collision. The existing methods lack the ability to perceive and analyze the movement mode of the robot group in detail, and it is difficult to form a cooperative control strategy that adapts to the real-time environment. SUMMARY

[0003] The present application aims to provide a multi-machine cooperation method for indoor and outdoor logistics robots to solve the problems raised in the background.

[0004] To achieve the above-mentioned purpose, the present application provides a multi-machine cooperation method for indoor and outdoor logistics robots, which comprises:

[0005] A multi-source perception data stream is constructed, robot pose information is collected through a laser radar, a visual sensor and an inertial measurement unit, and three-dimensional point cloud data of an environment is fused to generate spatiotemporally synchronized cooperative perception data;

[0006] Based on the cooperative perception data, a robot intercommunication topology network is established, a relative position relationship of the robots and data transmission quality are analyzed, and a cooperative relationship graph with connection weights is generated;

[0007] 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 a collaborative transportation process;

[0008] A multi-agent reinforcement learning algorithm is used to generate a multi-robot cooperation strategy that adapts to a dynamic environment according to the behavior consistency features;

[0009] The multi-robot cooperation strategy is executed in real time, and a distributed sensor network is used to monitor the execution effect of the strategy and collect cooperation performance indicators;

[0010] Based on the cooperation performance indicators, an online learning mechanism is used to update cooperation strategy parameters to achieve adaptive optimization of the multi-robot cooperation method.

[0011] Preferably, the construction of the multi-source perception data stream comprises:

[0012] A centimeter-level position coordinate of the robot is obtained through a super wideband positioning base station deployed in an indoor or outdoor environment, and a moving speed of the robot is measured using a robot-mounted odometer;

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

[0014] Timestamps are applied to various types of sensor data, and a Kalman filtering algorithm is used for data fusion to eliminate measurement errors;

[0015] A unified coordinate system is established to convert multi-source data to the same reference frame to generate spatiotemporally synchronized cooperative perception data.

[0016] Preferably, the data fusion using the Kalman filtering algorithm comprises:

[0017] A robot motion state equation is established, and position, velocity and acceleration are used as state variables;

[0018] An observation equation is designed, and super wideband positioning data and visual odometer readings are used as observation inputs;

[0019] A state estimation value is updated through a prediction-correction cycle iteration, and a covariance matrix is calculated to reflect estimation uncertainty;

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

[0021] Preferably, the establishment of the inter-robot communication topology network comprises:

[0022] According to the inter-robot distance threshold, an initial communication connection is established, and the signal transmission delay and packet loss rate are calculated;

[0023] Analyzing network connectivity, identifying key relay nodes and network weak links;

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

[0025] Assigning a quality weight to each communication link, generating a weighted collaboration relationship graph.

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

[0027] Applying the minimum cut set algorithm in graph theory to identify key connection points in the network;

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

[0029] When network partitioning risks are detected, actively adjust the robot position to maintain network connectivity;

[0030] Establishing network redundancy paths to improve communication reliability.

[0031] Preferably, the dynamic analysis using the spatio-temporal graph convolution network comprises:

[0032] Regarding robots as graph nodes and communication links as edges, a spatio-temporal graph data structure is constructed;

[0033] Designing multi-layer graph convolution operations to extract spatial features, and combining gate recurrent units to capture time evolution rules;

[0034] Through attention mechanism, neighbor node information is weighted and aggregated to highlight important collaboration relationships;

[0035] Outputting the behavior consistency score of each robot to identify abnormal collaboration patterns.

[0036] Preferably, the use of multi-agent reinforcement learning algorithm to generate a multi-robot collaboration strategy that adapts to the dynamic environment according to the behavior consistency features comprises:

[0037] Defining an agent for each robot, with its state space being the set of behavior consistency features and its own sensor data, and its action space being the speed, steering angle, and communication switch instructions;

[0038] A cooperative reward function is designed, which considers the overall task progress, anti-collision distance between robots and energy consumption efficiency at the same time;

[0039] Through the framework of centralized training and decentralized execution, each agent learns the strategy network that maximizes the cooperative reward.

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

[0041] Generating formation-maintaining speed instructions for a robot cluster transporting the same object together;

[0042] Generating dynamic avoidance instructions based on priority rules for a robot group moving towards each other in a channel;

[0043] And generating patrol instructions for robots performing area patrol tasks, which cover the maximum range and avoid repeated paths.

[0044] Preferably, the monitoring strategy execution effect collects cooperation performance indicators, including:

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

[0046] When it is detected that the task completion rate continuously falls below a threshold or the average distance deviation exceeds a safety range, it is determined that the cooperation is abnormal, triggering an emergency update of the cooperation strategy.

[0047] Preferably, the online learning mechanism is used to update the cooperation strategy parameters, including:

[0048] The difference between the collected cooperation performance indicators and the expected target is used as the loss function of the strategy network;

[0049] The asynchronous advantage actor-critic algorithm is used, and each robot uses the newly collected experience data to perform gradient update on the global strategy network in parallel;

[0050] The updated parameters of the global strategy network are periodically synchronized to all robots, realizing continuous self-adaptive optimization of the method.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The dynamic analysis of the cooperation relationship graph with connection weight by the space-time graph convolution network can model the space-time correlation between robots as a graph structure, thereby capturing the motion dependency of the group in the continuous time sequence. The network aggregates the neighbor robot state information through the graph convolution operation, extracts the group motion pattern features beyond the individual level, and accurately identifies the consistency degree of the behavior of each robot in the collaborative transportation. This method can quantitatively evaluate the cooperation quality and provide a data-driven and refined discrimination basis for subsequent strategy optimization, overcoming the limitations of traditional methods that rely only on individual state or simple rules for collaboration judgment. By using the multi-agent reinforcement learning algorithm, the cooperation strategy is generated according to the identified behavior consistency features, so that each robot agent can make distributed decisions based on local observations and group consistency goals. Through interaction with the environment, the algorithm constantly tries and learns, and the final strategy can adaptively handle common dynamic obstacles, communication fluctuations and other uncertain factors in the hospital environment. The robot learns to flexibly adjust its own action under the premise of maintaining the overall formation and transportation stability, realizes the cooperative transportation that meets the collaborative constraints and has individual flexibility in complex scenarios, and avoids the shortcomings of centralized planning with large computational load and difficulty in real-time response to unexpected situations. BRIEF DESCRIPTION OF DRAWINGS

[0053] Figure 1 The working principle diagram of the indoor and outdoor logistics robot multi-machine cooperation method is described.

[0054] Figure 2 The flowchart for constructing a multi-source perception data stream is described.

[0055] Figure 3 The flowchart for establishing a communication topology network between robots is described.

[0056] Figure 4 The performance index trend chart of the communication topology network is described.

[0057] Figure 5 The performance index comparison chart before and after the multi-machine cooperation strategy optimization is described. DETAILED DESCRIPTION

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

[0059] Please refer to Figure 1The application provides an indoor and outdoor logistics robot multi-machine cooperation method, which comprises the following steps: collecting the pose information of the robot through a laser radar, a visual sensor and an inertial measurement unit, and simultaneously fusing three-dimensional point cloud data of an environment to generate spatiotemporally synchronized cooperative perception data; this process ensures that all robots operate based on a unified time and space reference frame. A communication topology network between the robots is established based on the cooperative perception data, the relative position relationship and data transmission quality are analyzed, a cooperative relationship graph with connection weights is generated to quantify the interaction intensity between the robots. A spatiotemporal graph convolution network is applied to dynamically analyze the cooperative relationship graph, extract multi-robot motion mode features, and identify the behavior consistency in the cooperative transportation process, so as to capture the cooperation degree of group motion. A multi-agent reinforcement learning algorithm is used to generate a multi-machine cooperation strategy that adapts to the dynamic environment according to the behavior consistency features, so that the robot can autonomously adjust actions to respond to changes. In the strategy execution phase, the multi-machine cooperation strategy is deployed in real time, and the execution effect is monitored through a distributed sensor network to collect key cooperation performance indicators. Based on these performance indicators, an online learning mechanism is used to update the cooperation strategy parameters, realize continuous adaptive optimization of the method, and ensure that the cooperation efficiency is continuously improved in long-term operation.

[0060] Example 1: see Figure 2In a specific implementation, a multi-source perception data stream is constructed by deploying ultra-wideband positioning base stations in indoor and outdoor environments to obtain centimeter-level position coordinates of the robot, and a robot-mounted odometer is used to measure the moving speed. The ultra-wideband positioning base station calculates the time difference of arrival of high-frequency pulse signals to achieve high-precision positioning, and the odometer derives the displacement through encoder pulse counting to supplement short-term motion information. A stereo vision camera is used to capture the depth information of surrounding obstacles. The stereo vision camera calculates the pixel depth based on the binocular parallax principle, and combines with an infrared sensor to detect the environmental lighting conditions. The infrared sensor emits a modulated light beam and receives the reflection, and quantifies the environmental light intensity to adapt the exposure parameters of the vision algorithm. Time stamps are applied to the data of various sensors, which are generated synchronously by a hardware clock. Kalman filtering algorithm is used for data fusion to eliminate measurement errors. The Kalman filtering algorithm establishes the motion state equation of the robot, taking position, velocity and acceleration as state variables, and represents the system dynamics in vector form. The observation equation is designed, and the ultra-wideband positioning data and vision odometer readings are used as observation inputs. The observation equation linearizes the nonlinear sensor model. The state estimation value is updated through a prediction-correction iterative cycle. The prediction step uses the state equation to derive the prior estimate, and the correction step fuses the observation data to calculate the posterior estimate. The covariance matrix is calculated to reflect the estimation uncertainty, and the covariance matrix quantifies the statistical properties of the estimation error. When abnormal sensor data is detected, the observation noise parameters are adaptively adjusted to improve the robustness of the fusion. Abnormality detection is realized based on residual chi-square test. A unified coordinate system is established to convert multi-source data to the same reference frame. The unified coordinate system takes the world coordinate system as the reference, and aligns the local sensor data through rigid body transformation. Time and space synchronized collaborative perception data is generated. Time and space synchronization ensures data time consistency through interpolation and extrapolation methods.

[0061] In some embodiments, the deployment of ultra-wideband positioning base stations follows the principle of grid layout, base stations are spaced according to the size of the environment to cover the entire work area, robots carry ultra-wideband tag receivers to receive base station signals, and the position coordinates are calculated by the multilateration method. The odometer measures the moving speed, and the data of the wheeled encoder and the inertial measurement unit are fused. The encoder pulse is converted into displacement, and the inertial measurement unit provides angular velocity to compensate for the drift of the heading. The process of capturing the depth information of the surrounding obstacles by the stereo vision camera, the camera intrinsic and extrinsic parameters are pre-corrected using the calibration board, and the depth calculation is based on the epipolar geometry principle to generate a dense depth map. The infrared sensor detects the environmental lighting conditions and outputs the light intensity value, triggering the visual algorithm to switch modes, such as enabling image enhancement in low light conditions. The network time protocol is used to synchronize all sensor nodes with timestamp markers, ensuring sub-microsecond accuracy. The state equation of the Kalman filter algorithm is discretized based on the uniform motion model, and the observation equation takes the ultra-wideband data as the position observation and the visual odometry as the speed observation. In the prediction-correction cycle, the prediction step calculates the state transition, and the correction step applies the Kalman gain to minimize the mean square error. The covariance matrix update uses the Lyapunov equation, and the anomaly detection is achieved by monitoring the innovation sequence, and the observation noise is increased when the innovation exceeds the threshold. The coordinate system conversion applies Euler angles and rotation matrices to map the sensor local coordinate system to the world coordinate system.

[0062] Optionally, the generation of multi-source perception data streams is implemented in a pipeline manner, and the sensor data is first pre-processed and filtered to remove noise, and then input into the fusion module. The data packet of the ultra-wideband positioning base station contains the timestamp and coordinate value, the odometer data is output in the form of frequency, the stereo vision camera generates a sequence of depth frames, and the infrared sensor provides a scalar light value. The timestamp marker is triggered by FPGA hardware during data acquisition, ensuring global synchronization. The implementation of the Kalman filter algorithm uses an extended form to handle nonlinear problems, and the state variables include three-dimensional components of position, velocity, and acceleration. The observation equation linearizes the observation model of ultra-wideband and visual odometry. The iteration frequency of the prediction-correction cycle matches the sensor sampling rate, which is usually 100Hz, and the covariance matrix initialization is based on the sensor accuracy parameters. When adjusting the observation noise parameter, an exponentially weighted moving average is used to smooth the changes and avoid sudden changes. The establishment of the unified coordinate system system is realized by ground control point calibration, and the origin of the world coordinate system is set at the center of the environment. The spatiotemporal synchronized collaborative perception data output is a time-aligned multi-dimensional vector for subsequent modules.

[0063] It can be understood that the construction of multi-source perception data stream is the basis of the collaborative system, the ultra-wideband positioning provides a global reference, reduces the cumulative error, the odometer supplements the local motion details, the stereo vision camera enhances the environmental perception, and the infrared sensor guarantees the robustness. Data fusion optimizes estimation through Kalman filtering algorithm, and time stamp and coordinate system ensure data consistency. In specific implementation, sensor data acquisition is performed in parallel by distributed nodes, and is converged to a central processing unit through Ethernet or wireless network. When deploying the ultra-wideband positioning base station, the multipath effect suppression is considered, and the time reversal technology is used to improve the accuracy. When the odometer measures the moving speed, the wheel speed and gyroscope data are fused, and the drift is reduced through complementary filtering. After the stereo vision camera captures the depth information, a filtering algorithm is applied to remove outliers and generate a smooth point cloud. The infrared sensor detects the lighting conditions and outputs a digital signal to trigger the vision parameter adjustment. The time stamp label uses IEEE1588 protocol to achieve nanosecond-level synchronization. The state equation of the Kalman filtering algorithm considers the motion constraints such as the maximum acceleration limit, and the weight of the observation equation based on the signal-to-noise ratio is fused. In the prediction-correction cycle, 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 estimation confidence, and the anomaly detection is realized by monitoring the Mahalanobis distance. The coordinate system conversion applies quaternion to represent rotation and avoids gimbal lock. The spatiotemporally synchronized collaborative perception data is published in the form of message queue for downstream modules to subscribe.

[0064] In specific implementation, the construction of multi-source perception data stream emphasizes modular design, and the ultra-wideband positioning base station, the odometer, the stereo vision camera and the infrared sensor are used as independent modules to output standardized data packets. The ultra-wideband positioning base station calculates the distance through the TOA algorithm, the odometer integrates the speed to obtain the position, the stereo vision camera performs stereo matching to generate a parallax map, and the infrared sensor converts the analog signal through ADC. The time stamp label is embedded in the clock count in the data packet header, the Kalman filtering algorithm is realized as a recursive filter, the state variable is stored in a vector, and the observation equation is linearized by constructing a Jacobian matrix. The prediction-correction cycle runs in real time on an embedded system, and the covariance matrix uses Cholesky decomposition to ensure numerical stability. When adjusting the observation noise parameter, the noise statistics are learned based on historical data. The unified coordinate system aligns the sensor frame through hand-eye calibration, and the spatiotemporal synchronization uses an interpolation algorithm to fill in the missing data. After generating the collaborative perception data, the data integrity is verified through the checksum mechanism.

[0065] Embodiment 2: see Figure 3In specific implementations, the robot-to-robot communication topology network is established by establishing initial communication connections based on a distance threshold between robots, the distance threshold is set based on the effective transmission range of the wireless communication module, for example, the maximum connectable distance is calculated using the free space path loss model in an open environment, a communication link is attempted to be established when the distance between robots is less than the threshold, the signal transmission delay and packet loss rate are calculated, the delay measurement is achieved by sending a synchronization timestamp message and calculating the round-trip time difference, the packet loss rate is obtained by continuously sending test data packets and calculating the loss ratio, the network connectivity is analyzed, the key relay nodes and the network vulnerable links are identified, the minimum cut set algorithm in graph theory is applied to identify the key connection points of the network, the minimum cut set algorithm traverses all possible edge cut sets in the graph to find the cut set with the smallest weight, thereby determining which edges to remove to cause the network to split, the node betweenness centrality index is calculated, the node betweenness centrality is based on the proportion of the number of shortest paths between all pairs of nodes in the graph that pass through the node to evaluate the importance of the robot in the communication network, the communication quality changes are monitored in real time, the real-time fluctuations of the delay and the packet loss rate are detected by periodically broadcasting heartbeat messages, the network topology structure is dynamically adjusted, for example, when the packet loss rate of a certain link continuously exceeds the threshold, it is automatically disconnected and an alternative path is found, a quality weight is assigned to each communication link, the weight calculation considers the factors of delay, packet loss rate and bandwidth, a weighted average formula is used to generate a normalized weight value, a weighted collaboration relationship graph is generated, the collaboration relationship graph is stored in an adjacency matrix or adjacency list data structure, which is convenient for subsequent graph analysis algorithm processing. In some embodiments, the setting of the distance threshold takes into account the influence of environmental obstacles, ray tracing technology is used to simulate signal attenuation, the threshold is adjusted to adapt to non-line-of-sight conditions, the delay and packet loss rate calculation uses a sliding window average method to smooth short-term fluctuations, and the measurement stability is improved. The heartbeat message interval is dynamically adjusted according to the network stability. Optionally, the communication link establishment supports multiple wireless protocols, the optimal protocol is selected according to the scene, the clock drift is compensated in the delay measurement using NTP synchronization, the packet loss rate is distinguished between temporary faults and permanent faults, the minimum cut set algorithm considers edge weights when processing weighted graphs, the node betweenness centrality index is normalized to the range of 0-1 for easy comparison, and when dynamically adjusting the topology structure, high-weight links are prioritized, and the weight allocation formula has configurable parameters to adapt to different application requirements. It can be understood that the establishment of the communication topology network is the basis of the collaboration system, the distance threshold ensures the connection reliability, the delay and packet loss rate quantify the link quality, the minimum cut set algorithm and the node betweenness centrality provide network robustness analysis, real-time monitoring and dynamic adjustment to respond to environmental changes, and the weighted collaboration relationship graph supports refined collaboration decisions.

[0066] In a specific implementation, the process of analyzing network connectivity applies graph theory algorithm in depth, the min-cut set algorithm performs first constructing a graph model of the communication network, nodes represent robots, edges represent communication links, edge weights are initialized based on quality weights, the min-cut set algorithm iteratively merges nodes until the graph is reduced to two supernodes, records the minimum cut value and its corresponding edge set in the merging process, the identified key connection points correspond to the robots connected by these edges, counts the number of times each node serves as an intermediate node, the higher the node betweenness centrality value indicates that the robot undertakes more relay tasks in the network. When detecting network partition risk, risk detection is achieved by monitoring the change of the minimum cut value, if the minimum cut value is below the threshold value, it indicates that the network is vulnerable, actively adjusting the robot position to maintain network connectivity, the position adjustment instruction is generated based on the network connectivity optimization algorithm, for example, moving key robots to reduce the impact of cut sets, establishing network redundant paths, redundant path planning uses k shortest path algorithm to find multiple disjoint paths to improve communication reliability. In some embodiments, the min-cut set algorithm supports incremental updates when processing dynamic graphs, avoiding full graph recalculation, node betweenness centrality calculation optimizes the use of approximation algorithms to reduce computational complexity, suitable for large-scale networks, risk detection sets multiple thresholds to distinguish between warning and emergency state, position adjustment combines robot motion constraints such as maximum speed and safety distance, redundant path establishment considers the minimization of the sum of path weights. Optionally, the min-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 partition probability, 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 that the collaborative system can still operate when part of it fails, the min-cut set algorithm identifies weak links, node betweenness centrality guides resource allocation, active adjustment prevents network interruption, and redundant paths provide fault recovery mechanism.

[0067] In specific implementations, real-time monitoring of communication quality changes is achieved by deploying distributed monitoring agents, each robot running a monitoring process that periodically sends probe messages to neighbor robots, the probe messages containing a sequence number and a sending timestamp, the receiver recording the arrival time and calculating the latency, the statistics of message loss calculating the packet loss rate, the monitoring data being aggregated to a central server or shared through a distributed consensus protocol, the dynamic adjustment of network topology being triggered based on the monitoring data, the adjustment strategies including adding new links, removing degraded links, or modifying routing tables, the adjustment decisions using the principle of graph edit distance minimization to reduce the shock caused by topology changes. For each communication link, a quality weight is assigned, the weight calculation function inputting the latency, packet loss rate, and available bandwidth, the latency and packet loss rate using exponential weighted moving average filtering to process historical data, the bandwidth measurement being obtained through throughput testing, the weight value being normalized to the 0-1 interval, when generating the weighted collaboration graph, the graph structure being stored using a graph database to support fast querying and updating, the collaboration graph being output in a standard graph format for use by downstream modules. In some embodiments, real-time monitoring supports multi-granularity monitoring, coarse-grained for routine maintenance and fine-grained for fault diagnosis, a hysteresis mechanism is introduced to avoid frequent switching in dynamic adjustment, the weight assignment supports custom weight formulas to adapt to specific scenarios, and a collaboration graph visualization tool assists in debugging. Optionally, monitoring data is encrypted for transmission to ensure security, dynamic adjustment strategies can be configured in manual or automatic mode, weight calculation integrates user feedback to adjust parameters, and collaboration graph export interfaces support multiple analysis tools. It can be understood that real-time monitoring provides system state awareness, dynamic adjustment maintains optimal network performance, weight assignment quantifies link value, and collaboration graph is the basis for collaboration strategy input.

[0068] In specific implementations, the overall process of establishing an inter-robot communication topology network is run in an automated manner, in the initialization stage, robots broadcast their own position and capability information, an initial connection matrix is established based on distance thresholds, the connection matrix records the reachability between robots, initial latency and packet loss rate baseline values are calculated, in the network connectivity analysis stage, the minimum cut set algorithm and node betweenness centrality calculation are executed periodically, the results are cached for fast querying, in the real-time monitoring stage, background tasks are continuously running, network indicators are collected, in the dynamic adjustment stage, events drive the adjustment algorithm when performance degradation is detected, in the weight assignment stage, weight values are recalculated periodically, and the collaboration graph is updated in real time to ensure data freshness. In some embodiments, the process is optimized using parallel computing to accelerate graph algorithm execution, distributed storage of network state data, fault tolerance mechanisms handle robot joining or leaving, and the process integrates logging for auditing and optimization. Optionally, the process supports pause and resume operations, weight assignment allows external input priority settings, and collaboration graph version management tracks historical changes. It can be understood that this process achieves fully automated network management, reduces human intervention, and improves system reliability.

[0069] In specific implementations, application scenarios of the communication topology network include indoor warehouses and outdoor sites. The indoor environment considers multipath effects, and the distance threshold is set low. The outdoor environment considers weather effects, and the time delay measurement compensates for propagation delay. The minimum cut set algorithm processes sparse graphs, the node betweenness centrality focuses on hub nodes, real-time monitoring adapts to dynamic obstacles, dynamic adjustment quickly responds to changes in the terrain, weight distribution balances delay and reliability, and the collaboration graph supports large-scale robot clusters. In some embodiments, scenario-specific parameters are loaded through configuration files, and network optimization targets are adjusted according to different tasks, such as real-time priority or energy consumption priority. Optionally, scenario simulation uses a simulation environment to verify the algorithm, and stress testing is performed before actual deployment. It can be understood that scenario adaptability ensures the wide applicability of the method.

[0070] In specific implementations, the software architecture that implements the communication topology network adopts a microservice design, with each functional module deployed independently, such as the connection management module, the monitoring module, the analysis module, and the adjustment module. Modules communicate through a message queue, data streams are standardized using ProtocolBuffers serialization, hardware supports multiple communication chips, and software runs cross-platform on Linux or ROS systems. In some embodiments, the architecture supports horizontal expansion, modules can be distributedly deployed, data consistency uses distributed transactions to ensure, and performance monitoring integrates Prometheus index collection. Optionally, the architecture provides RESTAPI 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. Referring to Figure 4 The figure is a double-axis line graph, with running time as the horizontal axis and communication delay and packet loss rate as the left and right vertical axes, respectively, intuitively presenting the performance evolution process of the communication topology network in the multi-robot collaboration system. The figure verifies the effectiveness of the technical solution of real-time monitoring of communication quality changes and dynamic adjustment of network topology. The system continuously analyzes indicators such as signal transmission delay and packet loss rate, dynamically optimizes the network topology, and ultimately significantly improves the communication performance, providing low-latency and high-reliability communication support for multi-robot collaboration, ensuring the real-time and stability of data interaction between robots, and is the key communication foundation for efficient operation of the multi-robot collaboration system.

[0071] In a specific implementation, the process of dynamically analyzing the collaboration graph using a spatio-temporal graph convolutional network, the robots are treated as graph nodes and the communication links as edges, a spatio-temporal graph data structure is constructed, which is a composite data structure containing a node feature matrix, an edge connection matrix and a time series tensor, 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 weight of the communication link between robots in the form of an adjacency matrix, and the time series tensor records the historical node features and edge connection states at fixed time steps, thereby completely describing the evolution of the system in the spatio-temporal dimension. A multi-layer graph convolution operation is designed to extract spatial features, each layer of graph convolution operation updates the representation of the central node by aggregating neighbor node information, in a specific implementation, Chebyshev polynomial is used to approximate the graph convolution kernel to reduce the computational complexity, and a gated recurrent unit is combined to capture the time evolution law, which is a recurrent neural network variant that controls the retention and forgetting of historical information through update and reset gates, enabling it to effectively learn time-dependent relationships. The neighbor node information is aggregated by weighting through an attention mechanism, highlighting important collaboration relationships, the attention mechanism calculates the correlation score between nodes, and the aggregation weight is dynamically allocated based on node feature similarity and edge weight, enabling the model to focus on collaboration interactions that have a greater impact on the current task. The behavior consistency score of each robot is output, which is a scalar value, the final node representation is mapped to the score space through a fully connected layer to identify abnormal collaboration patterns, and abnormal detection is achieved by setting a score threshold or using statistical process control methods, an alarm is triggered when the behavior consistency score of a certain robot deviates from the group mean for a sustained period of time.

[0072] In some embodiments, the construction of the spatio-temporal graph data structure maintains recent data using a sliding window mechanism, the window size is set according to the task period and the balance of computing resources, the dimension of the node feature matrix is the number of robots multiplied by the feature dimension, the feature dimension includes basic motion state and extended context information, the edge connection matrix supports dynamic update to reflect the topology change, and the time series tensor is organized in the form of a three-dimensional array for batch processing. The number of layers of the multi-layer graph convolution operation is configured according to the network size and complexity, and is usually 2 to 3 layers, the output feature dimension of each layer is reduced layer by layer to extract high-order features, and the Chebyshev polynomial approximation expands the graph convolution kernel into a K-order polynomial to avoid feature decomposition and improve 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 feature, the update gate controls the history state, and the reset gate determines the influence degree of the current input on the state update. The attention mechanism is implemented in the form of multi-head self-attention, multiple attention heads are calculated in parallel and then spliced, the model expression ability is enhanced, and the correlation score is calculated using additive attention or dot product attention method. The behavior consistency score is normalized to the 0-1 interval after applying the Sigmoid function, and the abnormal cooperation mode recognition is combined with the time consistency check to avoid false positives due to transient fluctuations.

[0073] In some embodiments, the training process of the spatio-temporal graph convolution network uses a supervised learning method, the training data comes from historical task records or simulation generation, the label data is the normal or abnormal cooperation mode annotated by experts, the loss function combines the behavior consistency score error and the topology preservation constraint, and the optimization algorithm uses Adam or stochastic gradient descent method. The model inference stage is deployed on an edge computing device, the input is a real-time spatio-temporal graph data stream, the output period matches the decision-making period of the control system, and the behavior consistency score is pushed to the decision module in real time. Optionally, the model supports online learning mechanism to adapt to environmental changes through incremental update, and the modular design facilitates the integration of different sensor modal data. The key operation of spatial feature extraction in the spatio-temporal graph convolution network is described by the following formula:

[0074]

[0075] Wherein: represents the feature matrix of all nodes in the l-th layer, represents the feature matrix of all nodes in the l+1-th layer, which is the output obtained after the current layer calculation, is the normalized graph Laplacian matrix, is the Chebyshev polynomial of the l-th layer, is the Chebyshev polynomial of the l-th layer, is the maximum order of polynomial expansion, is the trainable parameter matrix of the l-th layer, ​​​denotes a nonlinear activation function.

[0076] It can be understood that the method of applying the spatio-temporal graph convolution network for dynamic analysis can effectively extract collaborative behavior features from complex multi-robot interaction data by combining the spatial modeling capability of graph convolution and the time modeling capability of the gating recurrent unit. The attention mechanism enhances the sensitivity of the model to key collaborative relationships. The behavior consistency score provides a quantitative basis for subsequent decision-making. The entire architecture takes into account both computational efficiency and model expressiveness, making it suitable for real-time collaboration systems. In specific implementation, the specific implementation of multi-layer graph convolution operation involves the construction and normalization processing of the graph Laplacian matrix, which is obtained by subtracting the adjacency matrix from the degree matrix. The normalization processing uses the symmetric normalization method to enhance numerical stability, and the Chebyshev polynomial recursive calculation avoids explicit feature decomposition, and the polynomial coefficients are obtained through training and learning. The time evolution modeling of the gating recurrent unit takes the graph convolution output sequence as input, initializes the hidden state as a zero vector, iteratively updates the hidden state through time steps, and the final hidden state encodes the time evolution law. When the attention mechanism is weighted and aggregated, the attention coefficient between each pair of nodes is calculated, the coefficient is normalized using the Softmax function, and the weighted sum generates the attention-enhanced node representation. The behavior consistency scoring output layer uses a fully connected structure, and the weights are optimized through training. The abnormal collaboration pattern recognition module continuously monitors the score sequence and uses a sliding window statistical method to detect abnormal points.

[0077] In some embodiments, model hyperparameters such as the number of layers, feature dimensions, and the number of attention heads are determined through cross-validation. Random dropout and noise injection are used for training data enhancement to improve generalization ability. Model compression techniques such as pruning and quantization are used for resource-constrained scenario deployment. Real-time inference optimization reduces latency through operator fusion and memory reuse, and distributed inference supports multi-robot collaboration computing. It can be understood that this implementation focuses on practical deployment feasibility, balances model complexity and computational overhead, and ensures system real-time response capability. The integration of the spatio-temporal graph convolution network and the multi-robot system is realized through a standardized interface. Input data preprocessing includes feature normalization and missing value filling, output post-processing includes score smoothing and abnormality confirmation, the system monitoring module records model performance indicators, and periodically re-trains to maintain model accuracy. Model updates support hot switching to avoid service interruption, and version management tracks model iteration history.

[0078] In a specific implementation, when generating multi-robot coordination strategies using multi-agent reinforcement learning algorithms, define an agent for each robot, which is an independent decision-making entity. The state space of the agent is a set of behavioral consistency features and its own sensor data. The behavioral consistency features come from the output of the spatio-temporal graph convolutional network, representing the degree of coordination between the robot and the group. The own sensor data includes lidar point cloud, visual recognition results, inertial measurement unit readings, and current position and speed information. The state space is a high-dimensional vector, which is normalized to speed up training. The action space is the speed, steering angle, and communication switch instructions. The speed instruction includes the set values of linear speed and angular speed. The steering angle controls the direction of the robot's travel. The communication switch instruction manages the activation and sleep of the wireless communication module to save energy. Design a collaborative reward function, which is a scalar value that considers the overall task progress, inter-robot collision avoidance distance, and energy consumption efficiency. The overall task progress is calculated based on the proximity of the target point or the completion of the item transportation. The inter-robot collision avoidance distance is penalized based on the minimum Euclidean distance between robots. The energy consumption efficiency is modeled based on motor power consumption and communication energy consumption. Through a centralized training and decentralized execution framework, each agent learns a strategy network that maximizes the collaborative reward. In the centralized training phase, the experience data of all agents is stored in a shared replay buffer for joint updating of a global policy network and value network. In the decentralized execution phase, each agent makes independent decisions based on local observations using the trained strategy network to generate actions.

[0079] Referring to Table 1, the multi-robot coordination strategy is embodied in motion coordination instructions between robots. The motion coordination instructions are set values for low-level controllers, including generating speed instructions for robot clusters transporting the same object to maintain formation. The formation can be linear, circular, or custom configuration. The speed instructions track the desired relative position and orientation through a PID controller to generate; generating dynamic avoidance instructions based on priority rules for robot groups moving towards each other in the channel. The priority rules dynamically allocate based on the task urgency, ID sequence, or distance from the target point. The avoidance instructions include temporary stop, slow down, or travel on the side; and generating patrol instructions for robots performing area patrol tasks that cover the maximum range and avoid repeated paths. The patrol instructions are generated based on grid maps or graph theory path planning algorithms to ensure efficient coverage of the area. The strategy network is usually represented by a deep neural network, with the input being a state vector and the output being a probability distribution or deterministic action value of the action space. The training algorithm can be deep deterministic policy gradient, multi-agent deep deterministic policy gradient, or its variants.

[0080] Table 1: Collaborative Reward Function Component Table

[0081]

[0082] In some embodiments, the construction of the state space can incorporate temporal information, for example, stacking the states of the last few time steps as input to the policy network to capture dynamic characteristics. The implementation of the action space can contain a mixture of continuous actions and discrete actions, for example, the velocity command is a continuous value, while the communication switch is a discrete binary command. The weight coefficients of the parts in the collaborative reward function need to be tuned through a large number of simulation experiments to achieve the best balance between task completion efficiency, safety, and energy consumption. The experience replay buffer in the centralized training and decentralized execution framework adopts a priority sampling mechanism, which more frequently replays samples with high learning value. The policy network of each agent can be designed as a homogeneous network with shared parameters to reduce model complexity and training cost, especially in large-scale robot clusters. The collaborative reward function can introduce additional reward terms, such as a reward for encouraging exploration of unknown areas, or a reward for punishing violent actions to improve motion smoothness. The priority rules of dynamic avoidance instructions can be designed in a learnable form, with an auxiliary network module to dynamically evaluate and assign priorities.

[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 dynamics parameters of the robot, or the noise model of the sensor. The generation of patrol instructions can combine topological methods to divide the patrol area into multiple sub-regions and solve the optimal access sequence based on the traveling salesman problem. The method of generating multi-robot cooperation strategies using multi-agent reinforcement learning algorithms defines each robot as an agent and designs a collaborative reward function that considers global goals, allowing each robot to learn an efficient collaborative strategy through centralized training while relying only on local observations. The centralized training and decentralized execution framework effectively solves the learning challenge in non-stationary environments. The output of the strategy is concretized as motion coordination instructions in various scenarios, making the method practical and flexible in dealing with complex indoor and outdoor logistics tasks. The reward function structure explicitly listed in the table provides a clear and quantifiable optimization goal for policy learning.

[0084] In a specific implementation, the process of monitoring the execution effect of the coordination strategy and collecting the coordination performance indicators is implemented through a distributed monitoring framework, each robot acts as a monitoring agent to continuously collect local data and upload to a central analysis node, the task completion rate, the average travel distance of robots per unit time, the average distance deviation between robots, and the communication load rate are calculated in real time, 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 travel distance of robots per unit time is obtained by aggregating 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 distances of all robot pairs and the expected formation distance, the communication load rate is calculated as the percentage of the actual data transmitted by the communication module per unit time to the maximum theoretical bandwidth. When it is detected that the task completion rate is continuously below the threshold or the average distance deviation exceeds the safety range, it is determined that the coordination is abnormal, triggering an emergency update of the coordination strategy, the threshold is dynamically set according to historical performance data and task requirements, for example, appropriately reducing the task completion rate threshold in complex environments to avoid misjudgment, and the safety range is determined based on the physical size of the robot and the braking distance. The emergency update mechanism of the coordination strategy uses a high-priority interrupt method to suspend the current strategy execution and load a backup strategy.

[0085] In some embodiments, the calculation of performance indicators uses a sliding time window method, the window size is configured according to the system response speed requirement, usually set to several minutes to several tens of minutes, and the indicator data is filtered to eliminate transient fluctuations. The abnormality detection algorithm combines threshold comparison and trend analysis, the task completion rate continuously below the threshold indicates that the indicators in consecutive multiple calculation windows are not up to standard, and the average distance deviation exceeding the safety range considers the duration and amplitude of the deviation. The emergency update process includes a rollback mechanism that can quickly restore to the last stable version when the performance does not improve after updating.

[0086] When updating the coordination strategy parameters using an online learning mechanism, the difference between the collected coordination performance indicators and the expected target is used as the loss function of the strategy network, the expected target is pre-set according to the task planning, for example, the task completion rate target is 98%, the average distance deviation target is 0.5 meters, and the difference calculation uses mean square error or absolute error form. Using the asynchronous advantage actor-critic algorithm, each robot uses the newly collected experience data to perform gradient updates on the global strategy network in parallel, the asynchronous advantage actor-critic algorithm includes an actor network to generate actions and a critic network to evaluate values, and the advantage function estimate is used to reduce variance. The parameters of the updated 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, usually set in the range of several hours to several days.

[0087] Optionally, the performance indicator collection supports custom extension, users can add new monitoring indicators according to specific scenarios, such as battery power balance or task queue length. The abnormality judgment rule can be configured as a multi-condition combination, such as triggering an emergency update when both the task completion rate is low and the communication load rate is high. The loss function design can introduce a weighting mechanism, assigning different weights to different performance indicators to reflect the priority. The implementation of the asynchronous advantage actor-critic algorithm can use a multi-worker architecture, each worker independently interacts with the environment and calculates the gradient. The parameter synchronization process supports differential synchronization, only transmitting the changed parameters to reduce network load. The monitoring data storage uses a time series database, which is convenient for long-term trend analysis and fault diagnosis. After the emergency update is triggered, the system can enter a safe mode, limiting the maximum speed of the robot until the policy optimization is completed. The online learning process contains an exploration strategy, which encourages policy diversity by adding noise to actions or using random sampling. The global policy network parameter synchronization uses version control to ensure that all robots in the cluster use the same policy version.

[0088] Optionally, the system provides a visual dashboard to display real-time performance indicators, assisting operation and maintenance personnel in mastering the system status. Collaborative abnormal event records detailed logs, including abnormal occurrence time, associated robot identifier and key indicator snapshot. The gradient update process contains a gradient clipping mechanism to prevent gradient explosion and unstable training. The parameter synchronization protocol has fault tolerance capability, supporting partial robot offline synchronization operation. Monitoring policy execution effect and using online learning mechanism to update collaborative strategy parameters constitute a complete adaptive optimization closed loop, through continuous monitoring of key performance indicators and timely discovery of collaboration abnormalities, the system can dynamically adjust the collaboration strategy to respond to environmental changes and internal state fluctuations. The difference between the performance indicators and the target as the loss function provides a clear direction for policy optimization, the asynchronous advantage actor-critic algorithm fully utilizes distributed computing resources to accelerate the learning process, and regular parameter synchronization ensures the consistency of the behavior of the entire robot cluster. This implementation makes the multi-robot collaboration system have the ability of self-optimization and fault recovery in long-term operation.

[0089] In a specific implementation, taking a specific example, the multi-robot system performs a cargo transportation task in a warehouse environment, the monitoring module calculates the task completion rate in real time, i.e., the proportion of the number of cargo boxes that have successfully arrived at the destination to the total task quantity, the average distance traveled by the robots per unit time is obtained by analyzing the odometer data of each robot, reflecting the overall operation efficiency, the average distance deviation between the robots is particularly important when the robots travel in formation, and is used to evaluate the formation keeping accuracy, and 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 the addition of obstacles, or that the average distance deviation of the robot formation continuously exceeds the 0.8 meter safety range due to wet ground, the monitoring module immediately marks the cooperative anomaly, and this event triggers the strategy updating process, the system suspends the current task and activates the emergency response mechanism. The online learning mechanism is then started, a loss function is calculated based on the difference between the recent performance data and the preset target value, and each robot runs an asynchronous advantage actor-critic algorithm in parallel, calculates the strategy gradient using the locally stored interaction experience, and 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 period, thereby optimizing subsequent cooperative behavior such as adjusting path planning strategies or improving formation control algorithms.

[0090] Referring to Figure 5 The figure is a grouped column chart, with task completion rate, distance traveled per unit time, average distance deviation, and communication load rate as the horizontal axis, and performance indicator value as the vertical axis. By comparing the two sets of data before and after optimization, the effect of multi-robot cooperation strategy optimization is intuitively presented. This figure verifies the effectiveness of the technical solution of monitoring cooperative performance indicators and updating strategies using online learning. Through continuous monitoring and strategy iteration of key indicators such as task completion rate, motion efficiency, cooperation accuracy, and communication cost, the multi-robot cooperation system achieves comprehensive performance optimization, forming a self-adaptive closed loop of monitoring-analysis-optimization-execution, and provides strong support for efficient collaborative work 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 and the like can be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0092] While embodiments of the application have been shown and described, it is to be understood that the embodiments described are merely exemplary of the principles and application of the present application. Numerous modifications and adaptions can be effected without departing from the spirit and scope of the present application, which is not limited to the exact construction and arrangement described. It is intended, therefore, to cover all modifications and adaptions that fall within the scope of the 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 a multi-source perception data stream, 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 communication topology network between robots 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; Applying a spatiotemporal graph convolution network to dynamically analyze the cooperative relationship graph, extracting multi-robot motion mode features, and identifying behavior consistency in the collaborative transportation process; Using a multi-agent reinforcement learning algorithm, generating a multi-robot cooperation strategy that adapts to the dynamic environment according to the behavior consistency features; Real-time execution of the multi-robot cooperation strategy and monitoring of the strategy execution effect through a distributed sensor network to collect cooperation performance indicators; Based on the cooperation performance indicators, updating the cooperation strategy parameters using an online learning mechanism to achieve adaptive optimization of the multi-robot cooperation method; The application of the spatiotemporal graph convolution network for dynamic analysis includes: Considering the robot as a graph node and the communication link as an edge, a spatiotemporal graph data structure is constructed; Designing multi-layer graph convolution operations to extract spatial features and combining gate recurrent units to capture time evolution rules; Weighted aggregation of neighbor node information through attention mechanism to highlight important cooperative relationships; Outputting the behavior consistency score of each robot to identify abnormal cooperation patterns. 2.The method of claim 1, wherein, The construction of the multi-source perception data stream includes: Obtaining the robot's centimeter-level position coordinates through ultra-wideband positioning base stations deployed in indoor and outdoor environments, and simultaneously measuring the moving speed using the robot-mounted odometer; Capturing the depth information of surrounding obstacles using stereo vision cameras and detecting the environmental lighting conditions using infrared sensors; Applying time stamp labels to various sensor data and using Kalman filtering algorithm for data fusion to eliminate measurement errors; Establishing 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: Establishing the robot motion state equation with position, velocity, and acceleration as state variables; Designing the observation equation with ultra-wideband positioning data and visual odometer readings as observation inputs; Updating the state estimate value through a prediction-correction iterative cycle and calculating the covariance matrix to reflect the estimation uncertainty; When detecting sensor data anomalies, adaptively adjusting the observation noise parameters to improve the robustness of the fusion. 4.The method of claim 1, wherein, The establishment of the communication topology network between robots includes: Establishing initial communication connections based on the distance threshold between robots, calculating signal transmission delay and packet loss rate; Analyzing network connectivity, identifying key relay nodes and network weak links; Real-time monitoring of communication quality changes and dynamically adjusting the network topology structure; Assigning quality weights to each communication link to generate a cooperative relationship graph with weights.

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: Applying the minimum cut set algorithm in graph theory to identify key connection points in the network; Calculating the node betweenness centrality index to assess the importance of robots in the communication network; When detecting network partitioning risks, actively adjusting the robot positions to maintain network connectivity; Establishing network redundancy paths to improve communication reliability. 6.The method of claim 1, wherein, 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.

7. The multi-machine cooperation method for indoor and outdoor logistics robots according to claim 6, characterized in that, 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. 8.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. 9.The method of claim 8, wherein, 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