Indoor and outdoor logistics robot scheduling method

Abstract

The invention relates to the technical field of logistics robot scheduling, and discloses an indoor and outdoor logistics robot scheduling method. According to the method, robot state and environment information is collected in real time through a multi-source sensor, and a standardized data set is formed through data verification and alignment. An interaction mode between robots is mined by using a graph neural network, dynamic characteristics are extracted in combination with time sequence analysis, and cluster state representation is constructed. And based on the representation, performing task and path planning by adopting a multi-objective optimization algorithm, and generating a scheduling instruction to control the robot to move. In the execution process, the system uses an incremental learning strategy to adjust and optimize algorithm parameters online through real-time performance index data, and system iteration updating is achieved. According to the method, cluster cooperation efficiency and scheduling precision are improved through graph neural network modeling, and long-term self-adaption and robustness of the system are enhanced through incremental learning.

Description

Technical Field

[0001] This invention relates to the field of logistics robot scheduling technology, specifically to an indoor and outdoor logistics robot scheduling method. Background Technology

[0002] Currently, in mixed scenarios such as warehousing and logistics, and industrial parks, the scheduling system for logistics robots is crucial. Existing scheduling methods mostly rely on predefined rule bases or basic optimization algorithms. For example, they use global path planning algorithms to assign static task sequences to individual robots, or use simple auction mechanisms for task allocation. These methods are effective in handling independent tasks and structured indoor environments. However, when robot swarms work collaboratively in dynamic, open indoor and outdoor environments, there are numerous nonlinear interactions between robots and between robots and complex environmental elements such as moving obstacles and changes in lighting. These traditional methods struggle to effectively characterize such complex spatiotemporal relationships.

[0003] Existing technical solutions have significant shortcomings. On the one hand, conventional systems have a superficial understanding of the robot swarm's state, typically only aggregating the independent state information of individual robots (such as position and battery level), while neglecting the dynamic interactive network formed by the robot swarm during movement. This neglect prevents the system from anticipating potential conflicts or opportunities for cooperation; for example, at narrow intersections, scheduling commands lacking interactive awareness can easily lead to deadlocks or congestion. On the other hand, the optimization algorithm parameters of most systems are fixed or require manual adjustment, lacking self-learning capabilities. In actual operation, factors such as task load, ambient lighting, and ground conditions are constantly changing, making it difficult for fixed parameters to maintain optimal performance continuously, resulting in long-term efficiency degradation and insufficient adaptability. Summary of the Invention

[0004] The purpose of this invention is to provide an indoor and outdoor logistics robot scheduling method to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides an indoor / outdoor logistics robot scheduling method, the method comprising:

[0006] By using indoor and outdoor sensor networks and robot body sensors, the robot's position coordinates, movement speed, energy status, obstacle distribution and light intensity information in the environment are acquired in real time. Data verification and time sequence alignment operations are performed to generate a standardized multi-source data set.

[0007] Spatiotemporal feature mining is performed on the standardized multi-source data set, graph neural networks are used to capture the interaction patterns between robots, and dynamic behavioral features are extracted by combining time series analysis to form a robot cluster state representation.

[0008] Based on the robot cluster state representation, a multi-objective optimization algorithm is used to plan task sequences and calculate path trajectories to generate individual robot scheduling instruction sets.

[0009] The robot's movement is controlled according to the scheduling instruction set, and the task execution progress is monitored and performance index data is collected through real-time data stream.

[0010] Using the performance index data, an incremental learning strategy is employed to adjust and optimize the algorithm parameters, thereby completing the iterative update of the scheduling system.

[0011] Preferably, the robot's position coordinates, movement speed, energy status, and information on obstacle distribution and light intensity in the environment are acquired in real time through indoor and outdoor sensor networks and robot body sensors. Data verification and timing alignment operations are then performed to generate a standardized multi-source data set, including:

[0012] The system receives real-time position, speed, and battery voltage data reported by the robot via a wireless communication module, as well as obstacle outlines captured by a vision sensor and distance information measured by an infrared sensor.

[0013] The received raw data stream is subjected to noise filtering. Wavelet transform algorithm is used to remove high-frequency noise components, outlier processing is performed, and box plot method is used to identify and correct abnormal values.

[0014] For multi-source asynchronous data, a dynamic time warping algorithm is applied to align the time series of different sensors, and interpolation synchronization is performed based on a reference time axis to eliminate timing inconsistencies.

[0015] The synchronized data stream is then subjected to dimensional unification and scaling, converted into a standard numerical format, and a standardized multi-source data set is generated.

[0016] Preferably, the step of noise filtering of the received raw data stream, using wavelet transform algorithm to remove high-frequency noise components, performing outlier processing, and using box plot method to identify and correct abnormal values ​​includes:

[0017] Wavelet decomposition is applied independently to each sensor data channel, and wavelet basis functions are selected for multi-scale analysis to separate signal and noise components.

[0018] The wavelet coefficients corresponding to noise are filtered out by thresholding, the main signal components are retained, and wavelet reconstruction is performed to obtain the denoised data.

[0019] Calculate the statistical characteristics of the denoised data, including the mean and quartile range, and set outoma boundaries based on the box plot rule;

[0020] Data points that exceed the boundary are considered outliers and replaced using the nearest neighbor interpolation method to ensure data continuity.

[0021] Preferably, the step of performing spatiotemporal feature mining on the standardized multi-source data set, using graph neural networks to capture interaction patterns between robots, and combining time series analysis to extract dynamic behavioral features to form a robot cluster state representation includes:

[0022] The robot is treated as a node in a graph neural network. A dynamic graph structure is constructed based on the physical location and communication relationships. The node attributes contain robot state data.

[0023] Graph attention networks are used to calculate the influence weights between nodes, and neighbor node information is aggregated through multi-layer graph convolution operations to extract spatial dependency features.

[0024] Long Short-Term Memory Network is used to model the time series data of each robot to capture long-term time dependencies and output time feature vectors.

[0025] By fusing spatial dependency features and temporal feature vectors, and performing dimensionality reduction and integration through a fully connected layer, a state representation of the robot cluster is generated.

[0026] Preferably, the robot is treated as a node in a graph neural network, and a dynamic graph structure is constructed based on its physical location and communication relationships. The node attributes include robot state data, including:

[0027] Using the robot's current position coordinates as a reference, Euclidean distances are calculated to establish a proximity graph, and edge weights reflect the interaction strength between robots;

[0028] Dynamically adjust the graph connections based on wireless signal quality, adding or deleting edges to match the actual communication topology;

[0029] Each node's attribute vector includes data on the robot's real-time speed, battery level, and task load.

[0030] The graph structure is updated over time to ensure it reflects the latest state of the robot swarm.

[0031] Preferably, the step of generating individual robot scheduling instruction sets based on the robot cluster state representation and using a multi-objective optimization algorithm for task sequence planning and path trajectory calculation includes:

[0032] The task allocation problem is modeled as a multi-objective optimization problem, with objective functions including minimizing total travel time, balancing robot load, and maximizing task completion rate.

[0033] A non-dominated sorting genetic algorithm is used to solve for the Pareto optimal solution set, and the solution population is evolved through selection, crossover, and mutation operations.

[0034] For each candidate solution, evaluate its path feasibility, use the A* algorithm for obstacle avoidance path planning, and calculate the trajectory cost;

[0035] Based on the evaluation results, the optimal scheduling scheme is selected, and the task sequence and waypoint instructions for each robot are decomposed to generate a scheduling instruction set.

[0036] Preferably, the task allocation problem is modeled as a multi-objective optimization problem, and the objective function includes minimizing the total travel time, balancing the robot load, and maximizing the task completion rate, including:

[0037] The decision variables are defined as the task-to-robot assignment matrix and the path point sequence, and the constraints include robot capacity limitations and environmental obstacle constraints.

[0038] The total travel time target is the sum of all robot travel times, the load balancing target is the variance of robot task volume, and the task completion rate target is the proportion of tasks completed on time.

[0039] A weighted sum method is used to transform multiple objectives into a single objective, with the weights dynamically adjusted based on the real-time system status.

[0040] The objective function value is minimized by iteratively optimizing the search allocation matrix and path sequence.

[0041] Preferably, the step of controlling the robot's movement according to the scheduling instruction set, while simultaneously monitoring the task execution progress and collecting performance index data through real-time data streams, includes:

[0042] The scheduling instruction set is sent to each robot controller to drive the robot to move along the planned path and perform grasping, transporting and placing operations;

[0043] The robot's position and posture are tracked in real time using LiDAR and cameras, and the planned trajectory is compared with the actual trajectory to calculate the deviation value.

[0044] Record the start time, end time, and resource consumption data for each task, and generate a task log stream;

[0045] Extract performance metrics from the log stream, including task latency, energy efficiency, and path deviation.

[0046] Preferably, the step of tracking the robot's position and posture in real time using lidar and cameras, comparing the planned trajectory with the actual trajectory, and calculating the deviation value includes:

[0047] Extended Kalman filter is used to fuse multi-sensor data to estimate the robot's current position and velocity state;

[0048] The estimated state is matched with the planned trajectory points to calculate the position error and angle error vectors;

[0049] The robot control parameters, such as speed gain and steering angle, are dynamically adjusted based on the error vector to reduce tracking deviation.

[0050] Historical deviation data is recorded in real time for subsequent analysis and model updates.

[0051] Preferably, the step of using the performance index data to adjust and optimize the algorithm parameters using an incremental learning strategy to complete the iterative update of the scheduling system includes:

[0052] The performance metrics data are used as training samples and input into the online learning model, which has a neural network regressor structure.

[0053] Calculate the gap between the current scheduling effect and the expected goal, and generate the loss function value;

[0054] The neural network weights are updated using the stochastic gradient descent algorithm to adjust the task allocation and path planning strategies.

[0055] Regularly validate model performance and retrain the model when the error exceeds a threshold to ensure that the scheduling system adapts to changes in the environment.

[0056] Compared with the prior art, the beneficial effects of the present invention are:

[0057] By employing graph neural networks to capture interaction patterns among robots and integrating time series analysis to extract dynamic behavioral features, robot swarms can be modeled as a complex, dynamically evolving system. This approach not only perceives the state of individual robots but also profoundly reveals the spatiotemporal relationships and mutual influences within the group, such as accurately predicting the propagation effect of speed fluctuations caused by obstacle avoidance. The resulting swarm state representation provides context-rich decision-making support for subsequent planning, enabling task sequence planning and path trajectory calculation to proactively avoid group conflicts and significantly improve the smoothness and overall efficiency of multi-robot collaborative operations.

[0058] By utilizing real-time monitored performance metrics and employing an incremental learning strategy to dynamically adjust and optimize algorithm parameters, the scheduling system gains the ability to continuously learn from practical operational experience. The system no longer relies on a static, unchanging model but can automatically fine-tune its decision-making logic based on feedback from task execution performance to adapt to dynamic environmental changes and long-term performance drift. This mechanism effectively overcomes the performance degradation drawbacks of fixed-parameter systems when facing unforeseen scenarios, ensuring that the scheduling strategy always approaches the optimal solution under current conditions, thereby achieving self-maintenance of long-term system stability and continuous performance improvement. Attached Figure Description

[0059] Figure 1 This is a schematic diagram illustrating the working principle of the indoor and outdoor logistics robot scheduling method described in this invention.

[0060] Figure 2 A flowchart for generating standardized multi-source datasets;

[0061] Figure 3 A flowchart for forming the state representation of a robot swarm;

[0062] Figure 4 Performance comparison chart of optimization algorithms for task allocation and scheduling for logistics robots;

[0063] Figure 5 A time distribution chart showing the task completion rate and number of tasks for logistics robots. Detailed Implementation

[0064] 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.

[0065] Please see Figure 1 This invention provides an indoor and outdoor logistics robot scheduling method. The method includes: acquiring the robot's position coordinates, movement speed, energy status, and environmental obstacle distribution and light intensity information in real time through indoor and outdoor sensor networks and robot body sensors; performing data verification and time sequence alignment operations to generate a standardized multi-source data set; performing spatiotemporal feature mining on the standardized multi-source data set, using graph neural networks to capture interaction patterns between robots, and combining time series analysis to extract dynamic behavioral features to form a robot cluster state representation; based on the robot cluster state representation, using a multi-objective optimization algorithm to plan task sequences and calculate path trajectories to generate individual robot scheduling instruction sets; controlling robot movement according to the scheduling instruction sets, while simultaneously monitoring task execution progress through real-time data streams and collecting performance index data; and using the performance index data, employing an incremental learning strategy to adjust and optimize algorithm parameters to complete the iterative update of the scheduling system.

[0066] Example 1: See Figure 2In practice, the robot's position coordinates, movement speed, energy status, obstacle distribution, and light intensity information in the environment are acquired in real time through indoor and outdoor sensor networks and robot body sensors. Data verification and time sequence alignment operations are performed to generate a standardized multi-source data set. The implementation process begins with the wireless communication module receiving real-time position data, real-time speed data, battery voltage data reported by the robot, as well as obstacle contour information captured by the vision sensor and distance information measured by the infrared sensor. These raw data streams are transmitted asynchronously, which may result in inconsistent timestamps or data loss. Therefore, a systematic processing flow is required to ensure data quality. In the specific implementation, noise filtering is performed on the received raw data stream. A wavelet transform algorithm is used to remove high-frequency noise components. This algorithm employs multi-scale analysis by selecting wavelet basis functions, independently applying wavelet decomposition to each sensor data channel to separate signal and noise components. The Daubechies wavelet is chosen as the wavelet basis function because it exhibits tight support and regularity when processing sensor time-series data, effectively capturing local signal features. Multi-scale analysis is achieved through discrete wavelet transform, performing a three-level decomposition on each sensor data channel independently. The first level extracts high-frequency detail coefficients, while the second and third levels progressively decompose low-frequency approximation coefficients, thus separating signal and noise components. Then, thresholding is used to filter out wavelet coefficients corresponding to noise, retaining the main signal components, and wavelet reconstruction is performed to obtain the denoised data. Simultaneously, outlier processing is performed, using box plots to identify and correct abnormal values. The statistical characteristics of the denoised data, including the mean and quartile range, are calculated. Anomaly boundaries are set based on the box plot rule, and data points exceeding these boundaries are considered outliers and replaced using nearest-neighbor interpolation to ensure data continuity. In some embodiments, for multi-source asynchronous data, a dynamic time warping algorithm is applied to align the time series of different sensors, and interpolation synchronization is performed based on a reference time axis to eliminate temporal inconsistencies. The dynamic time warping algorithm calculates the minimum path distance between different time series to align multiple sensor data streams to a unified time base, and uses linear interpolation or spline interpolation methods to fill in missing data points, ensuring the continuity of the data stream in the time dimension. Optionally, the synchronized data stream is subjected to dimensionality unification and scaling processing, and converted into a standard numerical format to generate a standardized multi-source dataset. Dimensionality unification involves normalizing the data units of different sensors, such as converting position coordinates to metric units, velocity to meters per second, and energy status to percentages. Scaling processing uses minimum-maximum scaling or Z-score normalization methods to ensure that all data fall within the same numerical range, facilitating use by the subsequent feature extraction module.

[0067] In specific implementations, the wireless communication module uses IEEE 802.11 series protocols or 5G communication technology to achieve data transmission. The data packets reported by the robot contain timestamps, sensor identifiers, and numerical fields. The system first parses the data packets and verifies their integrity. For data packets that fail verification, it requests retransmission or discards them. Obstacle contour information captured by the visual sensor is extracted using an edge detection algorithm, and distance information measured by the infrared sensor is calculated based on the time-of-flight principle. These data are buffered and stored at the receiving end, awaiting subsequent processing. In some embodiments, the wavelet transform algorithm in the noise filtering stage dynamically adjusts the number of multi-scale decomposition layers according to the signal characteristics. For example, more decomposition layers are used for velocity data rich in high-frequency noise. Thresholding uses soft or hard thresholding rules, and wavelet reconstruction is achieved through inverse wavelet transform. The box plot method for outlier processing calculates the upper quartile, lower quartile, and interquartile range for each data channel. The outlier boundary is set as the upper quartile plus 1.5 times the interquartile range and the lower quartile minus 1.5 times the interquartile range. The neighbor data interpolation method uses linear interpolation or time-series-based moving average interpolation to replace outliers to maintain data smoothness. It is understandable that the implementation of the dynamic time warping algorithm requires a predefined reference time axis, based on the system clock or the master sensor timestamp. The alignment process uses a dynamic programming algorithm to calculate the alignment path between sequences, and polynomial interpolation is used to ensure uniform distribution of data points during interpolation synchronization. In dimensionality unification and scaling, the standard numerical format is defined as a floating-point format, and the scaling parameters are dynamically updated based on historical data statistics; for example, the mean and standard deviation are calculated in real time from the sliding window data. Optionally, data verification operations include cyclic redundancy check and parity check. The time-series alignment operation uses multi-threaded parallel processing to improve efficiency, and the standardized multi-source data set is output as a structured array or database record for subsequent spatiotemporal feature mining module calls.

[0068] In practice, the entire data preprocessing process is executed in a pipeline manner, with each processing step sequentially connected. Intermediate results are cached in memory or storage devices to avoid data loss. The calculation of box plot rules uses descriptive statistical functions, the dynamic time warping algorithm uses dynamic programming tables to store intermediate results, and interpolation synchronization is completed through a numerical analysis library. In some embodiments, noise filtering and outlier processing can be applied in parallel to different sensor channels, utilizing multi-core processors to accelerate computation. In the time alignment stage, for high-frequency sensor data such as LiDAR point clouds, downsampling or aggregation methods are used to reduce the data volume to ensure real-time performance. It can be understood that standardization processing not only includes numerical transformation but also involves data format standardization, such as encapsulating all data into JSON or Protocol Buffers format for easy network transmission and parsing. After generating a standardized multi-source data set, the system records data processing logs, including processing time, data quality, and error reports, for monitoring and debugging. Optionally, resource constraints are considered during implementation. For embedded devices, algorithm complexity is optimized, for example, by using a simplified wavelet transform or approximate dynamic time warping. Data verification and timing alignment operations are managed through a state machine model to ensure the system can degrade in abnormal situations such as sensor failures. In specific implementations, the data receiving frequency of the wireless communication module is adaptively adjusted according to network conditions, and the data acquisition rates of the visual and infrared sensors are set synchronously to avoid data overload. Parameters of the wavelet transform algorithm, such as the wavelet basis function and the number of decomposition levels, are specified through a configuration file and support online updates. In some embodiments, the box plot rule for outlier processing can be extended to machine learning-based methods, such as using the isolated forest algorithm to detect anomalies, but this embodiment strictly follows the box plot rule. The reference time axis selection for the dynamic time warping algorithm prioritizes high-precision clock sources, and the interpolation method is selected according to the data type: linear interpolation is used for continuous data, and nearest neighbor interpolation is used for discrete data. It is understood that the implementation of dimensionality unification and scaling requires maintaining a metadata dictionary to record the original unit and transformation rules for each data dimension, and periodically calibrating scaling parameters. The output interface of the standardized multi-source data set is designed as an API, allowing subsequent modules to subscribe to the data stream.

[0069] Example 2: See Figure 3In practical implementation, robots are treated as nodes in a graph neural network, and a dynamic graph structure is constructed based on their physical location and communication relationships. Node attributes include robot state data. When constructing the dynamic graph structure, Euclidean distance is calculated based on the robot's current position coordinates to establish a proximity graph. Edge weights reflect the interaction strength between robots. Simultaneously, the graph connectivity is dynamically adjusted according to the wireless signal quality to add or delete edges to match the actual communication topology. Each node attribute vector contains multi-dimensional data such as the robot's real-time speed, battery level, and task load, and the graph structure is continuously updated over time to ensure it reflects the latest spatial distribution and connection status of the robot swarm. In practical implementation, a graph attention network is applied to calculate the influence weights between nodes, and neighbor node information is aggregated through multi-layer graph convolution operations to extract spatial dependency features. The graph attention network assigns a learnable attention coefficient to each node in the graph to measure the contribution of neighboring nodes to the feature update of the central node. Multi-layer graph convolution operations use an iterative message passing mechanism to make node features include multi-hop neighborhood information, thereby capturing complex spatial interaction patterns in the robot swarm, such as the state coupling relationships caused by cooperative avoidance or task passing. In some embodiments, a Long Short-Term Memory (LSTM) network is used to model the time-series data of each robot to capture long-term temporal dependencies and output a temporal feature vector. The LSM network selectively remembers and forgets historical information through its internal gating mechanism, processing continuous time series of parameters such as robot speed, position, and battery level to identify robot motion trends, periodic behavior patterns, or abnormal state evolutions. The temporal feature vector encodes the dynamic behavior patterns of individual robots within a specific time window. It can be understood that fusing spatial dependency features and temporal feature vectors involves dimensionality reduction and integration through fully connected layers, ultimately generating a robot cluster state representation. Spatial dependency features are obtained from the graph neural network output, representing the spatial association between robots at a specific moment. Temporal feature vectors are obtained from the LSM network output, representing the historical behavior patterns of individual robots. The fully connected layer concatenates or weights these two types of heterogeneous features and compresses them into a low-dimensional vector space through nonlinear transformation. The generated cluster state representation is a compact numerical vector that comprehensively summarizes the joint state of the entire robot system in the spatiotemporal dimensions.

[0070] In practical implementation, the construction process of the dynamic graph structure needs to be executed periodically, with the update frequency consistent with the sensor data acquisition frequency or dynamically adjusted according to the system load. The Euclidean distance is calculated based on the robot's planar coordinates in the world coordinate system. The edge connection threshold of the proximity graph is preset according to the robot's communication radius and task interaction requirements. The initialization of edge weights can be based on the reciprocal of the distance or the communication signal strength value. The wireless signal quality is evaluated by the received signal strength indicator or the signal-to-noise ratio measurement value. When the signal quality is lower than the threshold, the corresponding edge is disconnected; otherwise, the connection is established. In some embodiments, the implementation of the graph attention network adopts a multi-head attention mechanism to enhance the model's expressive power. Each attention head learns the correlation between different aspects between nodes, and the final attention weights are merged through concatenation or averaging operations. The graph convolution operation can choose a graph convolutional network or a variant graph neural network architecture such as graph sampling aggregation. The number of layers is usually set to 2 to 3 to avoid oversmoothing. Optionally, the input to the Long Short-Term Memory (LSTM) network is a fixed-length time-series window. The window size is determined based on the robot's motion dynamics; for example, a smaller window is used for short-term dynamics, and a larger window is used for long-term trends. The network can have a single or multiple layers, and the hidden state dimension is configured according to the feature complexity. The extraction of temporal feature vectors typically involves taking the hidden state of the last time step or performing pooling operations on the hidden states of all time steps. It can be understood that the dimensionality reduction and integration operations of the fully connected layers may include multiple sequentially connected linear transformation layers and nonlinear activation functions. The dimensionality of intermediate layers is gradually reduced to eliminate redundant information, and the dimension of the final output layer is fixed according to the requirements of the downstream optimization task. The robot cluster state representation, as the output of the entire feature extraction module, is written to shared memory or sent to a message queue for real-time reading and use by the scheduling decision module.

[0071] In practical implementation, the construction of node attribute vectors requires real-time selection of corresponding fields from a standardized multi-source dataset. For example, extracting the current coordinates from location coordinate data, the battery percentage from energy status data, and the current task load from task allocation logs. Attribute vectors typically need to be normalized before graph neural network processing to ensure consistent numerical ranges across dimensions. In some embodiments, the dynamic graph structure can be stored using an adjacency list or adjacency matrix data structure, with a sparse or dense matrix operation library selected based on the graph size. The calculation of graph attention coefficients involves linear transformations and activation function processing of node feature vectors. The extraction process of spatially dependent features can be represented as a weighted summation of neighboring node features, with the weighting coefficients dynamically generated by the attention mechanism. Optionally, the processing of the Long Short-Term Memory network is performed independently for each robot in the cluster. Parallel computing frameworks can be used to process the time-series data of all robots simultaneously to improve efficiency. In the fusion stage of time feature vectors and spatially dependent features, if dimensional mismatches exist, adjustments can be made through a projection layer. The fused features may be subject to Dropout regularization before being input into the fully connected layer to mitigate overfitting risks. It is understandable that the entire spatiotemporal feature mining process is deployed on a computing server with GPU acceleration capabilities. The model parameters are obtained through pre-training on historical data and forward inference is performed with fixed parameters during the online phase. In specific implementation, the update cycle of the robot cluster state representation is decoupled from the scheduling decision cycle. That is, the feature extraction module runs independently at a fixed frequency to ensure the freshness of the state information. For newly added robots to the cluster, the system dynamically initializes their corresponding nodes and connects them to the graph structure. For offline robots, their node updates are temporarily frozen. In some embodiments, the node feature update mechanism of the graph neural network can be extended to a model that considers edge features, such as using communication latency or link quality as edge attribute inputs. Long Short-Term Memory networks can be replaced with alternative structures such as gated recurrent units or temporal convolutional networks, as long as the basic requirement of capturing long-term temporal dependencies is met. Optionally, in addition to fully connected layers, more complex mechanisms such as attention fusion or outer product fusion can be tried for feature fusion, but the core purpose is to generate a joint representation with rich information. The generated robot cluster state representation will be accompanied by a timestamp and version number for multi-objective optimization algorithms to determine its validity. If the data expires, recalculation is triggered.

[0072] Example 3: In a specific implementation, the objective function of the multi-objective optimization problem includes minimizing the total travel time, balancing the robot load, and maximizing the task completion rate. The decision variables are defined as the task-to-robot assignment matrix and the path point sequence. The constraints include the robot's own load and capacity limitations, as well as constraints imposed by static and dynamic obstacles in the environment. In this implementation, a non-dominated sorting genetic algorithm is used to solve the Pareto optimal solution set for this multi-objective optimization problem. The non-dominated sorting genetic algorithm begins the iterative optimization process by initializing a random population of candidate solutions. Each candidate solution represents a complete task assignment and path planning scheme. The evolutionary process repeatedly performs selection, crossover, and mutation operations. The selection operation filters high-quality individuals based on their non-dominated level and crowding distance. The crossover operation exchanges some task assignment information or path point sequences between different parent individuals to generate offspring. The mutation operation randomly changes the task assignment or path point of a certain robot in an individual with a certain probability to maintain population diversity. In some embodiments, for each candidate solution in the non-dominated sorting genetic algorithm population, a path feasibility assessment is required, and the A* algorithm is used for detailed obstacle avoidance path planning and trajectory cost calculation. The path feasibility assessment checks whether the path point sequence meets the robot's physical constraints, such as minimum turning radius and maximum climbing angle. The A* algorithm performs a heuristic search on a grid map containing obstacle information to connect path points. The trajectory cost is calculated by comprehensively considering factors such as path length, estimated travel time, and safety risks. It can be understood that an optimal scheduling scheme is selected from the Pareto optimal solution set of the final generation based on the evaluation results of all candidate solutions, and decomposed into a specific task sequence and path point instructions for each robot to generate a scheduling instruction set. The selection of the optimal scheme is based on the system's real-time set preference strategy. For example, when it is necessary to efficiently complete a batch of urgent tasks, the scheme with the shortest total travel time is prioritized; when it is necessary to extend the overall system working time, the scheme with the most balanced load is prioritized. The decomposition process maps the global, abstract task allocation into a sequence of precise position and action instructions that each robot can directly execute.

[0073] In practical implementation, the objective function of the multi-objective optimization problem needs to be specifically quantified. Minimizing the total travel time is calculated as the sum of the movement times taken by all robots from their current positions, sequentially visiting all assigned task points, and finally returning to their waiting points. Balancing robot load is calculated as the variance of the workload undertaken by all robots (measured by the number of tasks, total weight of goods, or estimated operation time); a smaller variance indicates a more balanced cluster load. Maximizing the task completion rate is calculated as the ratio of the number of tasks expected to be completed before the specified deadline to the total number of tasks. To enable the non-dominated sorting genetic algorithm to handle these conflicting objectives with different dimensions, a comprehensive objective function needs to be constructed. One implementation method is to normalize each objective and then perform a weighted sum, the mathematical expression of which is as follows:

[0074]

[0075] in: This represents the value of the overall objective function, which needs to be minimized. The total estimated time for all robot movement paths; It is a baseline value for the total travel time, used for normalization, such as the average total time from historical scheduling records; The variance representing the task load of the robot swarm; It is a baseline value for the load variance, used for normalization. This represents the estimated task completion rate; It is a baseline value for task completion rate, usually set to 1 (i.e., 100%). , , These are weighting coefficients, representing the relative importance of the three objectives: total travel time, load balancing, and task completion rate. These weighting coefficients can be dynamically adjusted based on the real-time system status. In some embodiments, the parameters of the non-dominated sorting genetic algorithm need to be carefully configured. The population size is typically set between 50 and 200, the maximum number of generations is set between 100 and 500 generations depending on the problem complexity, the selection operation uses a binary tournament selection method, the crossover operation uses two-point crossover for the task allocation matrix and sequential crossover for the path sequence, and the mutation operation uses exchange mutation or reverse mutation. Optionally, the heuristic function of the A* algorithm can be either Manhattan distance or diagonal distance to adapt to different movement patterns. The environment map is stored in the form of a two-dimensional grid, with each grid containing passage cost information. Obstacle information is updated periodically from the real-time sensor data fusion layer. In some embodiments, the scheduling instruction set is generated using a structured data format. Each robot's instruction package contains a unique identifier, a priority-ordered task list, a trajectory consisting of a series of path point coordinates, and the expected speed and orientation information at each path point. The instruction set is distributed to the controller of each robot body via a wireless communication network. It is understandable that the entire optimization scheduling module runs as an independent service on the central scheduling server. The iterative calculation of the non-dominated sorting genetic algorithm and the path planning of the A* algorithm can be deployed on high-performance computing units or accelerated using parallel computing frameworks to meet real-time requirements. The robot cluster state representation serves as the main input to the optimization algorithm, and its update will trigger a new round of scheduling calculation cycle. However, the system will set a minimum scheduling interval to prevent excessively frequent rescheduling due to minor state fluctuations.

[0076] In practical implementation, the task allocation matrix is ​​a binary matrix or probability matrix, where rows correspond to tasks to be assigned and columns correspond to robots in the cluster. The values ​​of the matrix elements represent the allocation relationship between tasks and robots. The path point sequence is an ordered list containing the spatial coordinates of the locations the robots need to reach sequentially to complete the task. The fitness evaluation phase of the non-dominated sorting genetic algorithm requires calculating the above-mentioned comprehensive objective function value for each individual in the population, then performing fast non-dominated sorting to determine the Pareto rank of each individual, and calculating its crowding distance in the target space to promote the distribution of the solution set. In some embodiments, the dynamic adjustment of weight coefficients α, β, and γ can be based on a rule system. For example, when a robot in the cluster is found to have severely low battery power, the weight of the load balancing objective β is automatically increased to encourage tasks to be transferred to robots with sufficient battery power; when the system receives a large number of high-priority tasks, the weight of the task completion rate objective γ is increased. The implementation of the A* algorithm requires the integration of a real-time obstacle map. For dynamic obstacles, strategies such as reserving time windows or performing local replanning can be adopted. It is understandable that the issuance of scheduling instruction sets requires reliable communication protocols. Typically, instruction packets include serial numbers and checksums. After receiving an instruction, the robot controller needs to send an acknowledgment signal. If the central scheduler does not receive acknowledgment within the timeout period, it will trigger an instruction retransmission mechanism or initiate an exception handling process for that robot. In specific implementations, the optimized scheduling service described in Example 3 can be designed as an independent module within a microservice architecture, communicating asynchronously with the feature extraction module and motion control module through a message middleware to improve the decoupling and scalability of the entire scheduling system.

[0077] See Figure 4 This chart uses bar charts to compare the performance of three schemes—traditional methods, genetic algorithms, and multi-objective optimization algorithms—in three core metrics: total travel time, load balancing, and task completion rate. Blue bars represent total travel time, orange bars represent load balancing, and green bars represent task completion rate. From the data distribution, the multi-objective optimization algorithm performs best in all three dimensions: total travel time, load balancing, and task completion rate; the genetic algorithm is second best; the traditional method has the longest total travel time, poorest load balancing, and the lowest task completion rate. This is a visualization of the effect of the multi-objective optimization algorithm after solving the Pareto optimal solution set. By modeling the task allocation problem as a multi-objective optimization problem, this chart intuitively verifies the algorithm's superiority, provides a basis for selecting the optimal scheduling scheme, and serves as key decision support for generating individual robot scheduling instruction sets. Ultimately, it achieves synergistic optimization of efficient task allocation, load balancing, and high completion rate, driving a leap in the performance of logistics robot scheduling systems in complex scenarios.

[0078] Example 4: In a specific implementation, the scheduling instruction set includes task sequences and path point instructions independently generated for each robot. After receiving the instructions, the robot controller drives the robot to move along the planned path and controls end effectors such as robotic arms and grippers to complete specific operations such as grasping, transporting, and placing. The robot's motion control is based on a low-level closed-loop control law, such as a proportional-integral-derivative controller, to achieve accurate tracking of the target path. In a specific implementation, the robot's position and attitude are tracked in real time by LiDAR and cameras. The LiDAR acquires point cloud data by scanning the environment to calculate the robot's pose relative to environmental features. The camera assists in positioning through visual odometry or marker recognition technology. The acquired position and attitude data are compared with the planned trajectory to calculate deviation values ​​such as lateral deviation, longitudinal deviation, and angular deviation. It is understandable that recording the start time, end time, and energy consumption data of each task and generating a task log stream is the core of performance data acquisition. The task start time is recorded when the robot arrives at the task starting point and issues a ready signal, and the task end time is recorded when the robot completes the operation and receives success feedback. Energy consumption data is read in real time from the robot's battery management system. All data is accompanied by a timestamp and robot identification identifier and is continuously uploaded to the central monitoring system in the form of a structured data stream. Performance indicators are extracted from the task log stream, including task delay time, energy efficiency, and path deviation. Task delay time is calculated as the difference between the actual completion time and the planned completion time. Energy efficiency is calculated as the ratio of energy consumed to task difficulty. The path deviation is obtained by integrating or statistically averaging the position deviation sequence throughout the entire movement process.

[0079] In some embodiments, referring to Table 1, an extended Kalman filter (EPF) is used to fuse multi-sensor data to estimate the robot's current position and velocity state. The EPF linearizes the robot's motion and observation models, recursively fusing data from wheel odometry, inertial measurement units, lidar, and cameras in two stages: prediction and update, to generate the optimal robot state estimate. The state vector typically includes planar coordinates, orientation angle, and corresponding velocity. The estimated state is matched with planned trajectory points to calculate position and angle error vectors. The position error vector represents the Euclidean distance and direction between the estimated position and the nearest planned path point, while the angle error represents the difference between the robot's actual orientation and the planned orientation. The matching process is achieved by finding the planned point with the closest timestamp or by interpolating to obtain the ideal state at the corresponding moment. Based on the error vector, robot control parameters, such as velocity gain and steering angle, are dynamically adjusted to reduce tracking deviation. The adjustment strategy typically employs feedback control principles: when the position error is large, the forward speed is appropriately reduced and the steering angular velocity gain is increased; when the angle error is large, in-situ rotation correction is prioritized. The adjustment amount of the control parameters is determined by looking up tables or simple proportional relationships. Historical deviation data is recorded in real time for subsequent analysis and model updates. This historical deviation data includes timestamps, position error magnitudes, and absolute angle errors. This data is appended to the task log stream as raw data for evaluating path tracking performance and learning to improve the motion model.

[0080] Table 1: Performance Index Data Record Table Record timestamp robot identifier Task Identifier Current location X Current position Y Current speed Remaining battery power Current task status Deviation from planned path 2023-10-2710:05:30.123 Robot_05 Task_A012 1250.34 876.51 1.2 85.5 In progress (transportation) 0.15 2023-10-2710:05:31.456 Robot_05 Task_A012 1251.28 875.89 1.1 85.4 In progress (transportation) 0.18 2023-10-2710:06:00.789 Robot_12 Task_B007 980.15 1120.67 0.0 92.1 Placement complete. 0.02

[0081] In practical implementation, the robot controller's instruction execution unit parses scheduling instructions, converting them into underlying motor drive signals and I / O control signals. The drive signals control the robot's moving platform, while the I / O control signals manipulate the end effector's movements. The deployment locations of the LiDAR and cameras need to cover the entire work area to avoid blind spots. The data fusion algorithm runs on the robot's own computing unit or edge computing nodes. Task log stream generation adopts an asynchronous, non-blocking approach to avoid affecting the real-time performance of the robot's main control loop. Log data is first cached locally and then transmitted to the server in batches. The performance metric extraction module periodically scans the task log stream database, calculates various metric values ​​according to predefined algorithms, and stores them in the performance metric database for use during system iteration. Optionally, the calculation of deviation values ​​can incorporate filtering to smooth fluctuations caused by sensor noise, such as by performing a moving average of deviation values ​​over several consecutive time points. The robot's control parameters can be adjusted based on more complex strategies, such as fuzzy logic control, dynamically adjusting the gain according to the magnitude and trend of the error. It is understandable that the entire motion control and monitoring process constitutes a closed-loop system. The scheduling command is the input, the robot's actual motion is the output, the sensor monitoring is the feedback, and the performance index is the basis for evaluating the performance of the closed-loop system. This closed loop provides a data foundation for the adaptive optimization of the scheduling system.

[0082] Optionally, the task log stream data structure can be serialized using JSON or Protocol Buffers format to improve transmission and parsing efficiency. The actual storage of performance metric data records in the database may include more fields, such as ambient light intensity and communication signal strength, for more in-depth analysis. Extended Kalman filter parameters, such as the process noise covariance matrix and observation noise covariance matrix, need to be calibrated beforehand based on the accuracy of the sensors used. The robot's control parameter adjustment strategy can be designed as a configurable rule set, allowing maintenance personnel to fine-tune it based on practical experience. In implementation, ensuring time synchronization of all data records is crucial, typically achieved by using a network time protocol between the robot cluster and the server. Calibration parameters for LiDAR and cameras need to be periodically verified to ensure the accuracy of pose estimation. In addition to position information, task status judgment can be comprehensively assessed by combining feedback information from force sensors or vision sensors in the end effector, improving the reliability of status determination.

[0083] See Figure 5The graph, with time on the horizontal axis, shows the task completion rate (blue curve) on the left vertical axis and the number of tasks (green bars) on the right vertical axis, visually illustrating the dynamic relationship between completion rate and task volume during the execution of logistics robot tasks. The data trend shows that the number of tasks peaks at 2 PM, at which point the task completion rate, while fluctuating, remains at a relatively high level. The task completion rate reaches its highest point at 6 PM, while the number of tasks remains in a moderate range. This distribution reflects the task load and execution efficiency of the scheduling system at different times: during peak task load periods, the system maintains a high completion rate, demonstrating the effectiveness of the scheduling strategy; when the task load decreases, the completion rate further optimizes, indicating the dynamic adaptability of the system's resource allocation. This visualization is the result of real-time tracking of robot task execution progress using sensors such as LiDAR and cameras, combined with performance metrics extracted from task log streams. This data is used for subsequent incremental learning to adjust and optimize algorithm parameters, serving as a key basis for the iterative updates of the scheduling system. It provides data support for balancing task load, reducing task latency, and improving overall service quality, ultimately achieving efficient and stable logistics robot scheduling in dynamic task scenarios.

[0084] Example 5: In the specific implementation, the performance index data comes from the task delay time, energy efficiency, path deviation, and other data monitored and recorded in Example 4. After cleaning and labeling, this data forms a supervised learning dataset. The structure of the online learning model is chosen as a neural network regressor. The number of nodes in its input layer corresponds to the dimension of the performance index vector, and the number of nodes in its output layer corresponds to the number of optimization algorithm parameters that need to be adjusted. In the specific implementation, calculating the gap between the current scheduling effect and the expected goal and generating the loss function value is the key step driving the model update. The current scheduling effect is represented by the statistical value of the latest batch of performance index data, the expected goal is defined by the system's preset performance benchmark value, and the gap is calculated as the difference between the actual effect and the expected goal. The loss function is usually chosen to quantify this gap using mean squared error or mean absolute error. The weights of the neural network regressor are updated using the stochastic gradient descent algorithm to adjust the task allocation and path planning strategies. The stochastic gradient descent algorithm extracts a small batch of samples from the newly arrived performance index data each time to calculate the gradient of the loss function with respect to the network weights, and then updates the weights in the opposite direction of the gradient. The weight update makes the parameter suggestions output by the neural network regressor more inclined to generate a scheduling strategy that meets the expected goal. The system periodically verifies model performance and retrains the model when the error exceeds a threshold to ensure that the scheduling system adapts to changes in the environment. Model verification uses a reserved dataset independent of the training data to calculate the prediction error. When the error exceeds a preset threshold for multiple consecutive verification cycles, the model retraining process is triggered. Retraining uses a larger dataset, including historical data, for full training to reset the model foundation.

[0085] In some embodiments, the specific architecture of the neural network regressor includes an input layer, two hidden layers, and an output layer. The input layer receives a performance metric vector including several-dimensional features such as the average task latency, total energy consumption, and average path tracking deviation over nearly ten time windows. The first hidden layer contains a large number of neurons for feature extraction, the second hidden layer performs feature compression, and the output layer generates a set of standardized parameter adjustment suggestions, such as the adjustment amounts for the weight coefficients α, β, and γ used in multi-objective optimization algorithms. The loss function is understood to be a mean squared error function, calculated as the sum of squares of the differences between the predicted scheduling effect and the expected scheduling effect corresponding to the parameter adjustment suggestions predicted by the neural network regressor. Minimizing this loss function means guiding the system to evolve towards the expected performance goal. The implementation of the stochastic gradient descent algorithm requires setting hyperparameters such as the learning rate and batch size. The learning rate uses a decay strategy, employing a larger value in the early stages of training to accelerate convergence and a smaller value in the later stages to improve accuracy. The batch size is chosen in a balanced manner based on the data inflow rate and computational resources.

[0086] Optionally, performance metrics data need to be preprocessed before being input into the neural network regressor, including normalization and feature encoding. Normalization scales metrics with different dimensions to a similar numerical range, while feature encoding converts discrete states, such as task type, into numerical vectors. The update frequency of the online learning model can be synchronized with the scheduling batches, i.e., an incremental learning update is performed after each batch of task scheduling is completed and new performance data is collected. The retained dataset for model validation is randomly sampled and fixed from historical data to fairly evaluate the model's generalization ability. Optionally, when model retraining is triggered, the training process can be executed asynchronously in the background without interrupting the current online scheduling system. The system can then smoothly switch to the online service after the new model has been trained and validated. The scale of historical data used for retraining can be set within a time window based on storage resources and computing power; for example, only data from the most recent three months can be retained for training.

[0087] In practical implementation, incremental learning strategies enable the scheduling system to learn from actual operational experience. For example, if the system continuously observes a significant increase in task delays during the midday peak period, by analyzing the corresponding performance index data, the incremental learning model may automatically learn that the weight coefficient α of the total travel time target needs to be increased during this period. This allows for the generation of more efficiency-oriented task allocation schemes in subsequent scheduling decisions. Another concrete example is that when warehouse layout changes or a new robot model is introduced, leading to a change in the motion model, the initial path planning parameters may no longer be optimal, causing the path deviation index to rise. By continuously monitoring this index and inputting it into the incremental learning model, the model will gradually adjust the internal parameters or cost function weights of the path planning algorithm, making the planned path more adaptable to the new robot dynamics or environmental structure.

[0088] 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.

[0089] 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 scheduling indoor and outdoor logistics robots, characterized in that, The implementation process is as follows: By using indoor and outdoor sensor networks and robot body sensors, the robot's position coordinates, movement speed, energy status, obstacle distribution and light intensity information in the environment are acquired in real time. Data verification and time sequence alignment operations are performed to generate a standardized multi-source data set. Spatiotemporal feature mining is performed on the standardized multi-source data set, graph neural networks are used to capture the interaction patterns between robots, and dynamic behavioral features are extracted by combining time series analysis to form a robot cluster state representation. Based on the robot cluster state representation, a multi-objective optimization algorithm is used to plan task sequences and calculate path trajectories to generate individual robot scheduling instruction sets. The robot's movement is controlled according to the scheduling instruction set, and the task execution progress is monitored and performance index data is collected through real-time data stream. Using the performance index data, an incremental learning strategy is employed to adjust and optimize the algorithm parameters, thereby completing the iterative update of the scheduling system.

2. The indoor / outdoor logistics robot scheduling method according to claim 1, characterized in that, The system acquires the robot's position coordinates, speed, energy status, obstacle distribution, and light intensity in real time through indoor and outdoor sensor networks and the robot's own sensors. It then performs data verification and timing alignment operations to generate a standardized multi-source data set, including: The system receives real-time position, speed, and battery voltage data reported by the robot via a wireless communication module, as well as obstacle outlines captured by a vision sensor and distance information measured by an infrared sensor. The received raw data stream is subjected to noise filtering. Wavelet transform algorithm is used to remove high-frequency noise components, outlier processing is performed, and box plot method is used to identify and correct abnormal values. For multi-source asynchronous data, a dynamic time warping algorithm is applied to align the time series of different sensors, and interpolation synchronization is performed based on a reference time axis to eliminate timing inconsistencies. The synchronized data stream is then subjected to dimensional unification and scaling, converted into a standard numerical format, and a standardized multi-source data set is generated.

3. The indoor / outdoor logistics robot scheduling method according to claim 2, characterized in that, The process of noise filtering of the received raw data stream, using wavelet transform algorithm to remove high-frequency noise components, outlier processing, and box plot method to identify and correct abnormal values ​​includes: Wavelet decomposition is applied independently to each sensor data channel, and wavelet basis functions are selected for multi-scale analysis to separate signal and noise components. The wavelet coefficients corresponding to noise are filtered out by thresholding, the main signal components are retained, and wavelet reconstruction is performed to obtain the denoised data. Calculate the statistical characteristics of the denoised data, including the mean and quartile range, and set outoma boundaries based on the box plot rule; Data points that exceed the boundary are considered outliers and replaced using the nearest neighbor interpolation method to ensure data continuity.

4. The indoor / outdoor logistics robot scheduling method according to claim 1, characterized in that, The process of performing spatiotemporal feature mining on the standardized multi-source data set, using graph neural networks to capture interaction patterns between robots, and combining this with time series analysis to extract dynamic behavioral features, forms a robot cluster state representation, including: The robot is treated as a node in a graph neural network. A dynamic graph structure is constructed based on the physical location and communication relationships. The node attributes contain robot state data. Graph attention networks are used to calculate the influence weights between nodes, and neighbor node information is aggregated through multi-layer graph convolution operations to extract spatial dependency features. Long Short-Term Memory Network is used to model the time series data of each robot to capture long-term time dependencies and output time feature vectors. By fusing spatial dependency features and temporal feature vectors, and performing dimensionality reduction and integration through a fully connected layer, a state representation of the robot cluster is generated.

5. The indoor / outdoor logistics robot scheduling method according to claim 4, characterized in that, The robot is treated as a node in a graph neural network, and a dynamic graph structure is constructed based on its physical location and communication relationships. The node attributes include robot state data, including: Using the robot's current position coordinates as a reference, Euclidean distances are calculated to establish a proximity graph, and edge weights reflect the interaction strength between robots; Dynamically adjust the graph connections based on wireless signal quality, adding or deleting edges to match the actual communication topology; Each node's attribute vector includes data on the robot's real-time speed, battery level, and task load. The graph structure is updated over time to ensure it reflects the latest state of the robot swarm.

6. The indoor / outdoor logistics robot scheduling method according to claim 1, characterized in that, Based on the robot cluster state representation, a multi-objective optimization algorithm is used to plan task sequences and calculate path trajectories, generating individual robot scheduling instruction sets, including: The task allocation problem is modeled as a multi-objective optimization problem, with objective functions including minimizing total travel time, balancing robot load, and maximizing task completion rate. A non-dominated sorting genetic algorithm is used to solve for the Pareto optimal solution set, and the solution population is evolved through selection, crossover, and mutation operations. For each candidate solution, evaluate its path feasibility, use the A* algorithm for obstacle avoidance path planning, and calculate the trajectory cost; Based on the evaluation results, the optimal scheduling scheme is selected, and the task sequence and waypoint instructions for each robot are decomposed to generate a scheduling instruction set.

7. The indoor / outdoor logistics robot scheduling method according to claim 6, characterized in that, The task allocation problem is modeled as a multi-objective optimization problem, with objective functions including minimizing total travel time, balancing robot load, and maximizing task completion rate, including: The decision variables are defined as the task-to-robot assignment matrix and the path point sequence, and the constraints include robot capacity limitations and environmental obstacle constraints. The total travel time target is the sum of all robot travel times, the load balancing target is the variance of robot task volume, and the task completion rate target is the proportion of tasks completed on time. A weighted sum method is used to transform multiple objectives into a single objective, with the weights dynamically adjusted based on the real-time system status. The objective function value is minimized by iteratively optimizing the search allocation matrix and path sequence.

8. The indoor / outdoor logistics robot scheduling method according to claim 1, characterized in that, The process of controlling the robot's movement according to the scheduling instruction set, while simultaneously monitoring the task execution progress through real-time data streams and collecting performance index data, includes: The scheduling instruction set is sent to each robot controller to drive the robot to move along the planned path and perform grasping, transporting and placing operations; The robot's position and posture are tracked in real time using LiDAR and cameras, and the planned trajectory is compared with the actual trajectory to calculate the deviation value. Record the start time, end time, and resource consumption data for each task, and generate a task log stream; Extract performance metrics from the log stream, including task latency, energy efficiency, and path deviation.

9. The indoor / outdoor logistics robot scheduling method according to claim 8, characterized in that, The process of tracking the robot's position and posture in real time using lidar and cameras, comparing the planned trajectory with the actual trajectory, and calculating the deviation value includes: Extended Kalman filter is used to fuse multi-sensor data to estimate the robot's current position and velocity state; The estimated state is matched with the planned trajectory points to calculate the position error and angle error vectors; The robot control parameters, such as speed gain and steering angle, are dynamically adjusted based on the error vector to reduce tracking deviation. Historical deviation data is recorded in real time for subsequent analysis and model updates.

10. The indoor / outdoor logistics robot scheduling method according to claim 1, characterized in that, The step of using the performance index data to adjust and optimize algorithm parameters using an incremental learning strategy to complete the iterative update of the scheduling system includes: The performance index data is used as training samples and input into the online learning model, which has a neural network regressor structure. Calculate the gap between the current scheduling effect and the expected goal, and generate the loss function value; The neural network weights are updated using the stochastic gradient descent algorithm to adjust the task allocation and path planning strategies. Regularly validate model performance and retrain the model when the error exceeds a threshold to ensure that the scheduling system adapts to environmental changes.

Images