Map and localization fusion method and system for multi-cooperative robots

Abstract

This invention relates to the field of robot localization technology and discloses a map and localization fusion method and system for multiple collaborative robots. The method includes: acquiring localization error data and first local map data of multiple robots in a target cleaning area; compensating the first local map data based on the localization error data to obtain second local map data; performing error analysis based on the second local map data and the localization error data to obtain error compensation parameters; and fusing the second local map data according to the error compensation parameters to obtain fused global map data. Compared with traditional linear compensation methods, this method can more accurately handle complex spatial error distributions, thereby improving the speed and accuracy of error compensation.

Description

Technical Field

[0001] This invention relates to the field of robot localization technology, and in particular to a map and localization fusion method and system for multi-cooperative robots. Background Technology

[0002] Multi-robot collaborative systems are widely used in large warehouse cleaning. During collaborative operations, each robot needs to build a local environmental map and achieve precise localization. These local maps are then merged into a global map to support the execution of collaborative tasks. However, during long-term operations, multi-robot systems face a serious problem of cumulative error propagation. Traditional methods typically treat the localization errors of each robot as independent, neglecting the error coupling and propagation effects between robots through visual overlap and relative positional relationships. When a robot experiences a large localization error, this error propagates to other robots through spatial relationships, causing a rapid deterioration in the overall system's localization accuracy. The cumulative error grows exponentially, ultimately preventing the multi-robot collaborative task from being executed correctly. Summary of the Invention

[0003] This invention provides a map and localization fusion method and system for multi-cooperative robots. Compared with traditional linear compensation methods, this method can more accurately handle complex spatial error distributions, thereby improving the speed and accuracy of error compensation.

[0004] In a first aspect, the present invention provides a map and localization fusion method for multiple cooperative robots, the map and localization fusion method for multiple cooperative robots comprising: Acquire positioning error data and first local map data of multiple robots in the target cleaning area; The first local map data is compensated based on the positioning error data to obtain the second local map data. Error analysis is performed based on the second local map data and the positioning error data to obtain error compensation parameters; The second local map data is fused according to the error compensation parameters to obtain fused global map data.

[0005] In conjunction with the first aspect, in a first implementation of the first aspect of the present invention, acquiring positioning error data of multiple robots in the target cleaning area and first local map data includes: Data acquisition is performed by calling multiple LiDAR sensors, wheel encoder sensors, and inertial measurement unit sensors configured on the robot to obtain a multi-sensor dataset for each robot; Based on the multi-sensor dataset, Kalman filter fusion localization is performed in the target cleaning area to obtain the first pose data of each robot; Error calculation is performed based on the first pose data and the preset reference trajectory data to obtain the positioning error data of each robot; Based on the first pose data, the environmental point cloud data collected by the lidar sensor is processed into a rasterized SLAM mapping system to obtain the first local map data for each robot.

[0006] In conjunction with the first aspect, in a second implementation of the first aspect of the present invention, the step of compensating the first local map data based on the positioning error data to obtain the second local map data includes: The relative distance between robots and the area of ​​visual overlap are calculated based on the real-time position coordinates of each robot. Based on the relative distance and the area of ​​the visual overlap region, the exponential decay function is used to calculate the influence intensity of error propagation among multiple robots to obtain the error coupling coefficient matrix. The positioning error data and the error coupling coefficient matrix are input into the recursive error propagation equation to perform dynamic modeling of the cumulative error, thereby obtaining the initial error propagation model. Based on the initial error propagation model, a cumulative error propagation model is constructed, which includes the system state matrix, control input matrix, and process noise matrix. The first local map data is compensated according to the cumulative error propagation model to obtain the second local map data.

[0007] In conjunction with the first aspect, in a third implementation of the first aspect of the present invention, the step of compensating the first local map data according to the cumulative error propagation model to obtain the second local map data includes: The positioning error data is input into the cumulative error propagation model for recursive solution to obtain the cumulative positioning error vector of each robot, and the cumulative positioning error vector is subjected to time difference operation to obtain the error feature dataset. The error feature dataset is input into a radial basis function neural network for nonlinear compensation to obtain map error compensation data for each robot. Based on the map error compensation data, the grid coordinates in the first local map data are corrected to obtain the corrected local map data. The spatial distribution of occupied, free, and unknown grid cells is updated based on the corrected local map data to obtain the second local map data.

[0008] In conjunction with the first aspect, in a fourth implementation of the first aspect of the present invention, the step of correcting the grid coordinates in the first local map data based on the map error compensation data to obtain corrected local map data includes: The position correction, attitude correction, and rotation transformation parameters of each robot are analyzed from the map error compensation data to obtain the grid coordinate transformation parameters. Based on the grid coordinate transformation parameters, a rotation matrix transformation and translation vector correction are performed on each grid coordinate in the first local map data to obtain grid coordinate data. Based on the spatial relationship of landmarks in the raster coordinate data, graph optimization is performed to obtain landmark coordinate data. The landmark coordinate data is input into a semantic segmentation network to identify clean areas, obstacle areas, and navigation path areas, and corresponding error compensation strategies are executed to obtain corrected local map data.

[0009] In conjunction with the first aspect, in a fifth implementation of the first aspect of the present invention, the step of performing error analysis based on the second local map data and the positioning error data to obtain error compensation parameters includes: Zero-error expected trajectory is generated by parsing the clean path constraints and environmental boundary constraints from the second local map data; Calculate the deviation between the positioning error data and the zero-error expected trajectory, and design a sliding surface based on the deviation to generate a sliding surface parameter combination; Pole placement is performed based on the sliding surface parameter combination to obtain the convergence parameter configuration; The convergence parameter configuration and the error coupling coefficient matrix are combined to form the state equation coefficients, and the state equation coefficients are substituted into the Riccati differential equation for recursive solution to obtain the positive definite symmetric matrix solution. The optimal feedback gain matrix is ​​generated based on the solution of the positive definite symmetric matrix, and the optimal tracking controller is constructed by combining the optimal feedback gain matrix with the sliding surface parameters. The positioning error data is hierarchically controlled based on the optimal tracking controller to obtain error compensation parameters.

[0010] In conjunction with the first aspect, in a sixth implementation of the first aspect of the present invention, the step of performing hierarchical control on the positioning error data based on the optimal tracking controller to obtain error compensation parameters includes: The positioning error data is multiplied by the optimal feedback gain matrix in the optimal tracking controller to perform a linear quadratic optimal control operation, thereby generating the optimal feedback control component; When the positioning error data deviates from the sliding surface, the deviation distance from the error state to the sliding surface is calculated and a sliding mode control component is generated; When an error coupling propagation effect is detected among multiple robots, the historical error influence of each robot and the mutual propagation error are calculated, and an integral compensation control component is generated. The optimal feedback control component, the sliding mode control component, and the integral compensation control component are weighted and superimposed to form the error compensation parameter.

[0011] In conjunction with the first aspect, in a seventh implementation of the first aspect of the present invention, the step of calculating the deviation distance from the error state to the sliding surface and generating a sliding mode control component when the positioning error data deviates from the sliding surface includes: The positioning error data is substituted into the sliding surface function for calculation to obtain the calculation result. When the calculation result is not equal to zero, it is determined that the error state deviates from the sliding surface and the deviation distance from the error state to the sliding surface is output. Calculate the fractional power of the deviation distance and combine it with the sign function to generate the basic sliding control quantity that drives the error trajectory to converge rapidly to the sliding surface; The sliding gain coefficient is generated based on the real-time rate of change of the deviation distance; The basic sliding control quantity is multiplied by the sliding gain coefficient, and a boundary layer compensation term composed of a hyperbolic tangent function is superimposed to eliminate high-frequency chattering, thus forming a sliding mode control component.

[0012] In conjunction with the first aspect, in the eighth implementation of the first aspect of the present invention, the step of fusing the second local map data according to the error compensation parameters to obtain fused global map data includes: The error compensation parameters are applied to the positioning system of each robot to correct the current position coordinates and attitude angles of each robot, thereby obtaining the second pose data; Based on the second pose data, the positioning accuracy of each robot and the spatial consistency index of the second local map data are recalculated, and a reliability score is generated based on the positioning accuracy and the spatial consistency index. The target weights of each robot in map fusion are dynamically assigned based on the reliability scores. The target weights are applied to the second local map data to perform fusion, while simultaneously handling coordinate alignment and merging of overlapping areas between different robot maps to generate fused global map data.

[0013] Secondly, the present invention provides a map and localization fusion system for multiple cooperative robots, the map and localization fusion system for multiple cooperative robots comprising: The acquisition module is used to acquire positioning error data and first local map data of multiple robots in the target cleaning area; The compensation module is used to compensate the first local map data based on the positioning error data to obtain the second local map data; The error analysis module is used to perform error analysis based on the second local map data and the positioning error data to obtain error compensation parameters; The fusion module is used to fuse the second local map data according to the error compensation parameters to obtain fused global map data.

[0014] The technical solution provided by this invention establishes a mathematical model that considers the error coupling and propagation between robots, thereby accurately describing the error propagation mechanism of a multi-robot system and overcoming the technical deficiency of traditional methods that ignore the mutual influence between robots. A nonlinear map error compensation function is constructed, combined with a three-tiered architecture of geometric transformation, topology optimization, and semantic compensation to comprehensively eliminate various errors in map data. A finite-time sliding optimal tracking controller is designed, overcoming the technical bottleneck of traditional control methods that can only converge asymptotically, achieving accurate convergence of positioning errors within a preset time. A three-tiered control architecture including optimal feedback, sliding mode, and integral compensation is constructed to handle deterministic errors, modeling uncertainties, and coupling propagation errors respectively, achieving an optimized combination of control performance. A dynamic weight allocation mechanism based on real-time error compensation effect is established, adaptively adjusting the map fusion weights according to the error suppression effect, solving the technical problem that traditional fixed-weight fusion cannot adapt to dynamic error changes, and significantly improving the positioning accuracy and map fusion quality of multi-robot collaborative systems.

[0015] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0016] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of an embodiment of the map and localization fusion method for multiple collaborative robots in this invention. Figure 2 This is a schematic diagram of one embodiment of the map and localization fusion system for multiple collaborative robots in this invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0019] The terms "comprising" and "having," and any variations thereof, used in the embodiments of this invention are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the steps or units listed, but may optionally include other steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0020] To facilitate understanding of this embodiment, a map and localization fusion method for multi-cooperative robots disclosed in this invention will first be described in detail. For example... Figure 1 As shown, the map and localization fusion method for multi-cooperative robots includes the following steps: 101. Acquire positioning error data and first local map data of multiple robots in the target cleaning area; Specifically, within the target cleaning area, multiple LiDAR sensors, wheel encoder sensors, and inertial measurement unit sensors configured on the robots are utilized to continuously and synchronously collect data during robot movement, forming a multi-sensor dataset containing distance information, cumulative wheel speed information, and acceleration and angular velocity data. A Kalman filter algorithm is used to fuse the multi-sensor dataset, eliminating the instability and random drift effects of single-sensor data through recursive updates and noise suppression mechanisms. During the fusion process, the sensor error covariance matrix is ​​used to assign appropriate weights to data from different sources, obtaining the first pose data for each robot within the target cleaning area. This first pose data is compared and analyzed with preset baseline trajectory data. By calculating position and attitude deviations, positioning error data reflecting the difference between the robot's running trajectory and the ideal trajectory is obtained. The first pose data is used to perform rasterized SLAM mapping processing on the environmental point cloud data collected by the LiDAR sensors. By mapping continuously acquired point cloud frames to a local coordinate system and performing probabilistic occupancy grid modeling, a first local map data containing obstacle distribution, free areas, and unknown areas is constructed.

[0021] 102. Based on the positioning error data, the first local map data is compensated to obtain the second local map data; Specifically, during robot operation, the relative distance between robots is calculated based on the real-time position coordinates of each robot. The area of ​​the visual overlap region is determined by identifying overlapping areas in the field of view of multiple robot sensors, providing fundamental data describing the spatial relationships and observation redundancy of the robots. Based on the relative distance and the area of ​​the visual overlap region, the intensity of error propagation between multiple robots is calculated using an exponential decay function. This results in higher error coupling strength between robots that are closer together and have greater visual overlap, while the error coupling strength between robots that are farther apart or have less overlap is significantly reduced. This generates an error coupling coefficient matrix that comprehensively characterizes the degree of mutual influence between robots. The positioning error data and the error coupling coefficient matrix are input into the recursive error propagation equation to establish a time-series-sensitive cumulative error dynamic model, resulting in an initial error propagation model that reflects the error evolution trend. Based on the initial error propagation model, a cumulative error propagation model is constructed, including a system state matrix, a control input matrix, and a process noise matrix. This allows error propagation to not only characterize the positioning drift effect of individual robots but also reflect the error amplification or attenuation effect caused by the interaction between robots. The process noise matrix reflects the impact of environmental complexity and sensor uncertainty on error dynamics. The cumulative error propagation model is used to perform multi-level error compensation on the first local map data. By means of attitude correction, coordinate transformation and topology optimization, the ghosting and misalignment caused by the cumulative error are gradually eliminated, and the second local map data that is more consistent with the distribution of the real environment is generated.

[0022] 103. Based on the second local map data and positioning error data, perform error analysis to obtain error compensation parameters; Specifically, the clean path constraints and environmental boundary constraints are analyzed from the second local map data. A zero-error expected trajectory is generated by combining topological connectivity, obstacle expansion radius, and upper limits of curvature and velocity. The positioning error data is registered with the zero-error expected trajectory time-by-time, and position and attitude deviations are calculated. A finite-time sliding surface with integral and power-law arrival terms is constructed around the deviations to enhance low-frequency drift suppression and arrival velocity, thus forming a sliding surface parameter combination. Based on this sliding surface parameter combination, the closed-loop error is dynamically configured using the pole placement method, ensuring that the closed-loop poles are grounded in the stable region. The convergence velocity is quantitatively set in conjunction with task time constraints, with the upper bound of the convergence time constrained to no more than 300 seconds according to operational requirements. If necessary, parameter search or intelligent optimization is used to select better arrival and convergence coefficients, resulting in a convergence parameter configuration that satisfies the trade-off between time and robustness. The convergence parameter configuration and the error coupling coefficient matrix are combined to form the state equation coefficient set. These coefficients are then substituted into the improved Riccati differential equation for recursive solution, yielding a positive definite symmetric matrix solution. The solution process allows for time-varying design of the weight matrix to accommodate varying error strengths and environmental complexity at different stages, improving adaptability to coupling errors and external disturbances. Based on the positive definite symmetric matrix solution, an optimal feedback gain matrix is ​​generated and integrated with the sliding surface parameters at the control law level. This constructs an optimal tracking controller framework incorporating optimal feedback suppression, sliding mode robustness enhancement, and integral steady-state elimination. Boundary layers and adaptive gains mitigate high-frequency chattering and enhance uncertainty resistance. After the controller is operational, hierarchical control is implemented for the positioning error data: the optimal feedback layer rapidly suppresses deterministic error components; the sliding mode layer resists modeling uncertainties and transient disturbances; and the integral compensation layer eliminates steady-state biases and reduces residual errors introduced by coupling between robots. The output includes error compensation parameters such as attitude correction increments, translation correction vectors, fusion weight adjustment factors, and boundary layer thickness.

[0023] 104. The second local map data is fused according to the error compensation parameters to obtain the fused global map data.

[0024] Specifically, error compensation parameters are applied to the localization systems of each robot, enabling them to correct their real-time position coordinates and attitude angles during operation. This eliminates deviations caused by sensor drift, cumulative error propagation, and inter-robot coupling effects, resulting in second pose data. Based on this second pose data, the localization accuracy of each robot is recalculated. The convergence of position and attitude is quantified by comparing the difference between the expected and actual trajectories. Simultaneously, the spatial consistency index of the second local map data is used to evaluate the matching degree between local maps. This spatial consistency index is obtained by analyzing dimensions such as landmark overlap rate, boundary contour fitting degree, and grid occupancy consistency rate. The localization accuracy and spatial consistency index are weighted to generate a comprehensive score reflecting the current map reliability and credibility of each robot. Based on the reliability score, a dynamic allocation mechanism is used to assign target weights to different robots during the global map fusion process, ensuring high localization accuracy and reliability. Figure 1 Robots with high consistency are given higher weights, while those with larger errors or lower local map quality have their weights reduced accordingly, to avoid unreliable data interfering with the global map. The assigned target weights are applied to the fusion process of the second local map data. Coordinate alignment, pose rotation, and translation transformations are used to achieve map coordinate system consistency. Then, an overlapping area merging strategy is used to remove redundant information and retain unique environmental details, generating fused global map data.

[0025] In one specific embodiment, the process of performing step 101 may specifically include the following steps: Data acquisition is performed by calling multiple LiDAR sensors, wheel encoder sensors, and inertial measurement unit sensors configured on the robot to obtain a multi-sensor dataset for each robot; Based on the multi-sensor dataset, Kalman filter fusion localization is performed in the target cleaning area to obtain the first pose data of each robot; Error calculation is performed based on the first pose data and the preset reference trajectory data to obtain the positioning error data of each robot; Based on the first pose data, the environmental point cloud data collected by the lidar sensor is processed into a rasterized SLAM map to obtain the first local map data for each robot.

[0026] Specifically, during the process of multiple collaborative robots performing target cleaning tasks, each robot is equipped with a high-precision LiDAR sensor, wheel encoder sensor, and inertial measurement unit (IMU) sensor on its own hardware platform. The LiDAR sensor acquires two-dimensional or three-dimensional point cloud information of the robot's surrounding environment, reflecting the spatial distribution of obstacles, walls, and passable areas. The wheel encoder sensor records the rotation of the robot's wheels during movement, and calculates the robot's relative displacement on the plane by accumulating rotation angle and speed data. The IMU sensor provides the robot's angular velocity, acceleration, and attitude changes in three-dimensional space. Through the continuous collaborative operation of these three types of sensors, each robot collects a set of multi-source raw data containing spatial point clouds, wheel motion trajectories, and motion states such as acceleration and angular velocity at different time points, forming a multi-sensor dataset. Based on the multi-sensor dataset, Kalman filter fusion localization is performed in the target cleaning area. In the prediction phase, the robot's state is predicted a priori based on the displacement calculated by the wheel encoder and the attitude estimated by the IMU. In the update phase, the prior prediction is corrected using the environmental feature observation results provided by the LiDAR sensor, thereby continuously reducing noise interference and random drift in the filtering loop. The Kalman filter dynamically adjusts the prediction error covariance matrix and the observation error covariance matrix during execution to ensure the convergence and robustness of the fused localization. Through this process, each robot generates its first pose data within the target clean area, including the robot's translational position in the coordinate system and its current pose angle. This first pose data is compared with preset reference trajectory data. The reference trajectory data is pre-set during the planning phase to describe the robot's ideal motion path and pose evolution. By matching the actual first pose data with the reference trajectory point-by-point or over continuous time, the difference between the robot's current position and desired position, as well as the difference between the actual pose angle and desired pose angle, are calculated. These differences constitute the localization error data, reflecting the robot's deviation during operation due to various factors such as sensor drift, uneven ground, tire slippage, and environmental occlusion. The first pose data is then used to perform rasterized SLAM mapping on the environmental point cloud data acquired by the LiDAR. LiDAR continuously scans the environment to generate point cloud frames during robot movement. The SLAM algorithm projects these point cloud frames onto a global coordinate system according to the coordinates of the robot's first pose data and discretizes the environment space using an occupancy grid model. During gridding, the environment space is divided into numerous small grid cells, and a probabilistic method is used to determine whether each cell is occupied by an obstacle. Simultaneously, by combining motion estimation and loop closure detection mechanisms, map ghosting and misalignment caused by accumulated errors are reduced, improving the geometric consistency of the map.Through the above processing, each robot generates a first local map data reflecting the environmental structure within its perception range. The local map contains the spatial layout of the cleaning area and also marks the distribution of obstacles and passable areas.

[0027] In one specific embodiment, the process of performing step 102 may specifically include the following steps: The relative distance between robots and the area of ​​visual overlap are calculated based on the real-time position coordinates of each robot. Based on the relative distance and the area of ​​visual overlap, the exponential decay function is used to calculate the influence intensity of error propagation among multiple robots, and the error coupling coefficient matrix is ​​obtained. The positioning error data and the error coupling coefficient matrix are input into the recursive error propagation equation to perform dynamic modeling of the cumulative error, thus obtaining the initial error propagation model. A cumulative error propagation model, comprising the system state matrix, control input matrix, and process noise matrix, is constructed based on the initial error propagation model. The first local map data is compensated using the cumulative error propagation model to obtain the second local map data.

[0028] Specifically, during task execution, the real-time position coordinates of multiple collaborative robots within the target cleaning area are continuously acquired. These coordinates are provided by the robot's local positioning module and filtered to ensure numerical stability. Based on each robot's real-time position coordinates, the relative distance between robots is determined through geometric calculations. The distance directly reflects the proximity of different robots in spatial distribution. Simultaneously, using LiDAR scanning results and camera visual perception results, overlapping areas in the observation areas of multiple robots are identified, yielding the area of ​​the visual overlap region. Relative distance and the area of ​​the visual overlap region together characterize the mutual perception relationship between robots; the former reflects spatial proximity, while the latter reflects the redundancy and consistency of environmental observation information. An exponential decay function is used to calculate the impact of relative distance and the area of ​​the visual overlap region on error propagation. Since error propagation between robots is not linear but decays rapidly with increasing distance, introducing an exponential decay form effectively captures this physical characteristic. When two robots are close together and the visual overlap region is large, the error coupling effect is enhanced, and their errors are superimposed and amplified; conversely, when the distance is far or there is almost no visual overlap, the possibility and intensity of error propagation decrease. By employing a nonlinear mapping relationship, the influence intensity of error propagation between any two individuals in a robot swarm is calculated, forming a symmetric error coupling coefficient matrix. The localization error data and the error coupling coefficient matrix are input into the recursive error propagation equation to construct a dynamic modeling process for cumulative error. The recursive equation, through iteration over time steps, combines the error state of the previous moment with the error increment of the current moment and the coupling effect between robots to obtain the cumulative error of the robot population at each moment, thus yielding the initial error propagation model. This initial error propagation model is then extended to a cumulative error propagation model that includes a system state matrix, a control input matrix, and a process noise matrix. The system state matrix describes the self-perpetuation and mutual propagation of errors over time, the control input matrix reflects the ability of the compensation strategy to adjust the error state, and the process noise matrix models the uncertainties caused by the complexity of the external environment and sensor noise. The cumulative error propagation model is used to compensate for the first local map data, including geometric coordinate translation and attitude angle correction, as well as topological optimization of landmark relationships and probabilistic reassessment of grid occupancy states. Each grid point in the map has its coordinates adjusted based on the robot's current position and the corrected error model, thereby eliminating misalignment and ghosting caused by positioning errors. Simultaneously, leveraging error coupling, data consistency correction is performed in overlapping areas of multiple robot local maps, ensuring seamless spatial stitching of maps from different sources. Through compensation operations, a second local map data is obtained.

[0029] In one specific embodiment, the process of compensating the first local map data according to the cumulative error propagation model to obtain the second local map data may specifically include the following steps: The positioning error data is input into the cumulative error propagation model for recursive solution to obtain the cumulative positioning error vector of each robot. Then, the cumulative positioning error vector is subjected to time difference operation to obtain the error feature dataset. The error feature dataset is input into a radial basis function neural network for nonlinear compensation to obtain map error compensation data for each robot. Based on map error compensation data, the grid coordinates in the first local map data are corrected to obtain the corrected local map data. The second local map data is obtained by updating the spatial distribution of occupied, free, and unknown grid cells based on the corrected local map data.

[0030] Specifically, the positioning error data is input into the cumulative error propagation model. The model is then recursively solved using the state matrix, control input matrix, and error coupling coefficients to calculate the current cumulative positioning error vector for each robot at each time step. The cumulative positioning error vector describes the growth trend of the error over time and reflects the error coupling effect between robots due to spatial proximity and perceptual overlap. To extract dynamic features, a time difference operation is performed on the cumulative positioning error vector. This involves calculating the difference between adjacent time steps to capture the error change rate and error fluctuation characteristics, thus forming an error feature dataset. This error feature dataset is then input into a radial basis function neural network (RBN) for nonlinear compensation. RBNs are suitable for handling nonlinear mapping problems. They utilize radial basis functions such as Gaussian kernel functions to map the input space to a high-dimensional feature space, thereby approximating complex nonlinear relationships. During training and operation, the network continuously adjusts its weight parameters to minimize the deviation between the input error features and the target compensation value in its output. Because robot localization errors exhibit nonlinear, time-varying, and random characteristics in practical applications, radial basis function neural networks (RBNs) utilize their function approximation capabilities to generate corresponding map error compensation data based on the input error feature dataset. This data includes corrections for position coordinates and attitude angles, with differentiated outputs for each robot. Based on the map error compensation data, the grid coordinates in the first local map data are corrected. The spatial coordinates of each grid point in the original local map are combined with the corresponding compensation data, and map misalignment and distortion caused by accumulated localization errors are corrected through coordinate translation and angle rotation. The correction operation is not limited to the overall coordinate system offset but also involves detailed adjustments to areas with significant errors at the local level to ensure that the corrected map's geometric structure better matches the distribution of the real environment. Based on the corrected local map data, the spatial distribution of the grid is updated to generate the second local map data. The corrected coordinate information is remapped into the occupied grid representation, and the state of each grid cell is remarked through probability updates: if a region is covered by new sensor observation data and the presence of an obstacle is confirmed, the grid corresponding to that region is marked as an occupied grid; if a region remains empty in multiple observations, its corresponding grid is marked as an idle grid; and for regions that still lack observation information, the grid state remains unknown.

[0031] In one specific embodiment, the process of performing error correction on the grid coordinates in the first local map data based on map error compensation data to obtain the corrected local map data can specifically include the following steps: The position correction, attitude correction, and rotation transformation parameters of each robot are analyzed from the map error compensation data to obtain the grid coordinate transformation parameters; Based on the raster coordinate transformation parameters, a rotation matrix transformation and translation vector correction are performed on each raster coordinate in the first local map data to obtain raster coordinate data; Graph optimization is performed based on the spatial relationships of landmarks in raster coordinate data to obtain landmark coordinate data. The landmark coordinate data is input into the semantic segmentation network to identify the clean area, obstacle area and navigation path area, and the corresponding error compensation strategy is executed to obtain the corrected local map data.

[0032] Specifically, the map error compensation data is analyzed to extract the position correction, attitude correction, and rotation transformation parameters for each robot. The position correction adjusts the robot's translation coordinates, eliminating displacement deviations caused by sensor drift or accumulated errors. The attitude correction corrects the robot's current orientation angle, ensuring the map coordinate system aligns with the actual environment. The rotation transformation parameters construct a standard rotation matrix, performing a rigid body transformation of the map coordinates at the geometric level. After analysis, the grid coordinate transformation parameters are obtained. Based on these parameters, rotation matrix transformation and translation vector correction are performed on each grid coordinate in the first local map data. The rotation matrix transformation multiplies the original coordinates of each grid with the rotation matrix to correct the coordinate system's angle, aligning the local map's orientation with the real environment. Translation vector correction adds the position correction to the spatial coordinates of each grid, eliminating overall positional offset. Through rotation matrix transformation and translation vector correction, the spatial coordinates of each grid are redefined, generating new grid coordinate data. Graph optimization is then performed based on the spatial relationships of landmarks contained in the grid coordinate data. Landmarks are key feature points in a map, representing stable and repeatable structural information in the environment, such as corners, door frames, or boundaries of characteristic obstacles. By analyzing the relative positional relationships between landmarks in raster coordinate data, a constraint graph model is constructed. In this model, nodes represent landmark coordinates, and edges represent spatial constraints between nodes. A graph optimization method is used to solve the constraint graph model, minimizing the error function of all constraint edges to obtain globally consistent landmark coordinate data. This landmark coordinate data is then input into a semantic segmentation network for scene parsing. The semantic segmentation network uses a deep learning model to classify the input map data pixel-by-pixel or raster-by-raster, automatically identifying clean areas, obstacle areas, and navigation path areas. Clean areas are identified as passable free space, obstacle areas correspond to impassable occupied space, and navigation path areas are the key corridors and passages upon which the planning algorithm relies. Based on the recognition results for different regions, corresponding error compensation strategies are implemented: For clean areas, the focus is on maintaining the continuity of grid boundaries and the uniformity of spatial distribution to avoid fragmentation of free areas due to errors; for obstacle areas, the geometric consistency of boundary points is strengthened to ensure that the positions of obstacles on the map are aligned with the real environment to avoid robot collisions; for navigation path areas, the path width, direction, and connectivity are corrected by combining the global optimization results of landmarks to ensure the reliability and accessibility of path planning. Through differentiated compensation for different semantic regions, corrected local map data is generated.

[0033] In one specific embodiment, the process of performing step 103 may specifically include the following steps: Zero-error expected trajectory is generated by parsing clean path constraints and environmental boundary constraints from the second local map data; Calculate the deviation between the positioning error data and the zero-error expected trajectory, and design a sliding surface based on the deviation to generate a combination of sliding surface parameters; Pole placement is performed based on the sliding surface parameter combination to obtain the convergence parameter configuration; The convergence parameter configuration and the error coupling coefficient matrix are combined to form the state equation coefficients, and the state equation coefficients are substituted into the Riccati differential equation for recursive solution to obtain the positive definite symmetric matrix solution. The optimal feedback gain matrix is ​​generated based on the solution of the positive definite symmetric matrix, and the optimal tracking controller is constructed by combining the optimal feedback gain matrix with the sliding surface parameters. Based on the optimal tracking controller, the positioning error data is controlled in layers to obtain error compensation parameters.

[0034] Specifically, the cleaning path constraints and environmental boundary constraints are analyzed in the second local map data. The analysis process includes generating boundary bands along the occupied grid contours, expanding obstacles according to the minimum safety distance, filtering passable corridors based on bandwidth constraints and minimum turning radius constraints, and generating a coverage sequence by combining coverage targets and operation priorities. Subsequently, continuity and curvature smoothing are performed on the corridor skeleton lines to ensure that the desired path meets the three requirements of collision-free, passable, and complete coverage, forming a zero-error desired trajectory. The positioning error data and the zero-error desired trajectory are registered temporally and spatially. Temporal registration is completed by unifying timestamps and interpolation, while spatial registration is completed by rigid body alignment and outlier removal. Position deviation and attitude deviation are calculated to form a deviation sequence, and a finite-time sliding surface is designed based on the deviation sequence. The sliding surface includes an integral channel sensitive to low-frequency drift and a nonlinear arrival channel that responds quickly to transient deviations. Boundary layer thickness, arrival index, and integral weight together constitute the sliding surface parameter combination. After obtaining the sliding surface parameter combination, pole placement is performed around the linearized error dynamics. The goal of the placement is to obtain sufficiently large damping and convergence margin under given actuator capabilities and operation time limits. The pole positions are obtained by inverse calculation using the expected rise time, allowable overshoot, and upper bound of steady-state error, forming a convergence parameter configuration that satisfies the upper limit of convergence time and chattering suppression requirements. The convergence parameter configuration and the error coupling coefficient matrix are combined to form the state equation coefficients. The state equation simultaneously covers the single-robot self-propagation term and the multi-robot propagation term. The state equation coefficients are then substituted into the improved Riccati differential equation for recursive solution. The solution process employs discrete-time iteration and numerical stabilization. The weight matrix is ​​adaptively adjusted according to the operation stage and environmental complexity. The convergence criterion is based on maintaining positive definiteness and reducing the objective cost function, resulting in a positive definite symmetric matrix solution. The optimal feedback gain matrix is ​​generated based on the solution of the positive definite symmetric matrix. This optimal feedback gain matrix is ​​then structurally fused with the sliding surface parameters at the control law level to form the optimal tracking controller. The feedback channel suppresses deterministic error components, the sliding channel provides robust enhancement against modeling uncertainties and external disturbances, the integral channel eliminates steady-state bias and compensates for coupling residuals, and the boundary layer and adaptive gain are used to reduce high-frequency chattering and avoid overloading the actuator. Based on the optimal tracking controller, hierarchical control is implemented on the real-time positioning error data. The first layer directly calculates attitude and position corrections according to the optimal feedback gain. The second layer adjusts the nonlinear arrival intensity and boundary layer thickness based on the sliding surface state. The third layer generates cross-robot coupling compensation quantities by integrating the error coupling intensity, and the outputs of the three layers are synthesized into executable correction commands under the action of an anti-saturation mechanism. The hierarchical control outputs are mapped to error compensation parameters after safety limiting and rate limiting. These error compensation parameters include position correction, attitude correction, rotation transformation parameters, boundary layer thickness, and fusion weight adjustment factors.

[0035] In one specific embodiment, the process of performing hierarchical control of positioning error data based on the optimal tracking controller to obtain error compensation parameters can specifically include the following steps: The positioning error data is multiplied by the optimal feedback gain matrix in the optimal tracking controller to perform a linear quadratic optimal control operation, which produces the optimal feedback control component. When the positioning error data deviates from the sliding surface, the deviation distance from the error state to the sliding surface is calculated and the sliding mode control component is generated; When an error coupling propagation effect is detected among multiple robots, the historical error influence of each robot and the mutual propagation error are calculated, and an integral compensation control component is generated. The optimal feedback control component, the sliding mode control component, and the integral compensation control component are weighted and superimposed to form the error compensation parameter.

[0036] Specifically, the positioning error data is multiplied by the optimal feedback gain matrix in the optimal tracking controller to perform a linear quadratic optimal control operation. During the online phase, the optimal feedback control component is output in real time according to the linear quadratic optimal control operation of "error vector × feedback gain matrix". The weight design considers the relative importance of position and attitude, the safety cost increase caused by obstacle distance, and the confidence reduction penalty obtained from noise statistical estimation, reducing the risk of overcorrection without sacrificing response speed. After the linear quadratic optimal component is generated, dual amplitude and slope limits are set, and a noise-suppressed observer with hysteresis is used to shield high-frequency measurement jitter, so that the control output can be stably followed by the drive system. When the positioning error data deviates measurably from the sliding surface, the sliding mode channel calculates the deviation distance from the error state to the sliding surface, and generates the sliding mode control component based on the arrival criterion, boundary layer thickness, and arrival exponent. The larger the deviation distance and the longer the residence time outside the boundary layer, the higher the nonlinear arrival intensity of the sliding mode component; when entering the boundary layer, the control law automatically switches to a continuous asymptotic mode to reduce chattering. The sliding gain employs an adaptive update strategy: it increases the arrival strength when the error energy rises and slowly decreases it when the error energy falls, avoiding overdrive during the disturbance convergence phase. The sliding modal control channel and the linear quadratic optimal channel operate in parallel. The sliding modal control channel provides robust enhancement against modeling uncertainties and external disturbances, while the linear quadratic optimal channel ensures a fast response under the nominal model. Both maintain sampling alignment in the time domain and are decoupled from each other in the frequency domain through boundary layers and filtering bandwidth, preventing mutual noise amplification. In multi-robot collaborative scenarios, error coupling propagation introduces residual bias across individuals; therefore, an integral compensation channel is needed to track the impact of historical errors and mutual propagation errors over a long period. The integral compensation channel continuously reads the coupling strength assessment and historical error records, establishing higher coupling weights for robot pairs that are relatively close and have a large visual overlap area. It accumulates the mutual influence using a time decay kernel to obtain the cross-robot integral compensation control component. To prevent integral drift and drive saturation, the integral channel employs a conditional triggering and anti-saturation mechanism: integration is paused when the error amplitude falls below a threshold, the actuator approaches saturation, or repeated direction switching occurs within the sampling period; a release factor is introduced when the error energy decreases to a set range to gradually release the accumulated amount; and a freezing strategy is implemented to maintain the current compensation value and prevent further accumulation when the coupling strength falls below a threshold. The integral channel is jointly determined by the coupling matrix, historical error convolution, and release parameters, effectively mitigating chronic bias and cross-individual propagation residuals. The optimal feedback control component, sliding mode control component, and integral compensation control component are weighted and superimposed. The fusion weights are determined by error amplitude, error energy change rate, speed level, environmental complexity, and ground conditions. Figure 1The dynamic coefficient set is jointly determined by consistency indicators and drive margin; upper limits are set for weight changes to avoid drastic fluctuations in a short period that could lead to control reversal. The superimposed results are limited in amplitude and rate by a safety monitor, and then checked for monotonicity and feasibility to ensure that the position correction, attitude correction, and rotation parameters are consistent with the kinematic constraints of the machine. Passability constraints, cleaning path priority, and boundary safety bandwidth are incorporated into secondary corrections to ensure that the final output meets both operational safety and efficiency objectives. After completing the above checks, error compensation parameters are formed, including position correction, attitude correction, rotation transformation parameters, suggested boundary layer thickness, and fusion weight adjustment factors.

[0037] In one specific embodiment, the process of calculating the deviation distance from the error state to the sliding surface and generating the sliding mode control component when the positioning error data deviates from the sliding surface can specifically include the following steps: The positioning error data is substituted into the sliding surface function for calculation to obtain the calculation result. When the calculation result is not equal to zero, it is determined that the error state deviates from the sliding surface and the deviation distance from the error state to the sliding surface is output. Calculate the fractional power of the deviation distance and combine it with the sign function to generate the basic sliding control quantity that drives the error trajectory to converge rapidly to the sliding surface; The sliding gain coefficient is generated based on the real-time rate of change of the deviation distance; The basic sliding control quantity is multiplied by the sliding gain coefficient, and a boundary layer compensation term composed of a hyperbolic tangent function is superimposed to eliminate high-frequency chattering, thus forming the sliding mode control component.

[0038] Specifically, the positioning error data is substituted into the sliding surface function time-by-time for calculation. Time-domain alignment and coordinate system unification ensure that the input and sliding surface are defined within the same reference frame. The calculation result of the sliding surface function is read and compared with the zero value. When the calculation result is not equal to zero, it can be determined that the error state deviates from the sliding surface. Simultaneously, the deviation distance from the output error state to the sliding surface is measured based on the norm or equivalent distance of the sliding surface function. Before output, amplitude limiting and noise reduction filtering are used to suppress instantaneous spikes. Based on the deviation distance, a basic sliding control quantity is constructed. While keeping the deviation distance non-zero, the fractional power is calculated and combined with the sign function to form a nonlinear driving term with finite-time arrival characteristics. The fractional power exponent is selected to fall within the interval between 0 and 1 to enhance the arrival speed and suppress overshoot. The sign function provides direction determination capability, causing the error trajectory to be quickly attracted along the sliding surface normal vector. Simultaneously, a smooth approximation is used near the sliding surface to avoid numerical jitter, ensuring the engineering feasibility of continuous differentiability of the control law. The sliding gain coefficient is generated based on the real-time rate of change of the deviation distance. The rate of change of the deviation is used as an adaptive indicator of the arrival intensity: when the absolute value of the rate of change increases and the duration exceeds the minimum dwell time, the sliding gain increases the arrival intensity according to the preset rising slope, so as to maintain sufficient convergence thrust in the stage of increased disturbance or amplified model uncertainty; when the absolute value of the rate of change decays and stabilizes within the threshold range, the sliding gain slowly falls back to reduce the actuator burden and suppress the waste of control energy. At the same time, the upper and lower limits of the sliding gain are set in combination with the error energy index and the observation confidence to avoid triggering excessive amplification in the noise-dominated range. The nonlinear arrival value of the main control channel is obtained by multiplying the basic sliding control quantity by the sliding gain coefficient, and a boundary layer compensation term is superimposed within the same sampling period to eliminate high-frequency chattering. The boundary layer compensation term adopts a continuous approximation based on the hyperbolic tangent function. By dividing the deviation distance by the boundary layer thickness and inputting it into the hyperbolic tangent function, a vibration suppression channel with a smooth transition in the neighborhood of the sliding surface is constructed. When the deviation distance enters the boundary layer, the compensation term dominates the control output and gradually weakens the discontinuity caused by the sign function, thereby reducing the high-frequency switching behavior of the actuator. In engineering implementation, it is combined with a first-order hysteresis element and rate limit to further reduce mechanical shock and electrical noise, forming the sliding mode control component. To ensure that the control component works in coordination with the overall optimal tracking framework, the sliding mode component is subject to safety supervision and consistency verification: the output and optimal feedback component are decoupled in amplitude and spectrum to prevent the two channels from amplifying measurement noise in the same frequency band; based on dynamic constraints and operational constraints, saturation protection and gradual-in / gradual-out strategies are implemented for the sliding mode component to ensure that both position correction and attitude correction can be safely tracked by the kinematic chain and drive chain.

[0039] In one specific embodiment, the process of performing step 104 may specifically include the following steps: Error compensation parameters are applied to the positioning system of each robot to correct the current position coordinates and attitude angles of each robot, thereby obtaining the second pose data; The positioning accuracy of each robot and the spatial consistency index of the second local map data are recalculated based on the second pose data, and a reliability score is generated based on the positioning accuracy and spatial consistency index. The target weights of each robot in map fusion are dynamically assigned based on the reliability score. The target weights are applied to the second local map data to perform fusion, while the coordinate alignment and overlapping area merging between different robot maps are handled to generate fused global map data.

[0040] Specifically, error compensation parameters are applied to the localization systems of each robot to dynamically correct their current position coordinates and attitude angles, resulting in second pose data. The localization accuracy of each robot is recalculated based on this second pose data, and a spatial consistency index is obtained by combining it with the corrected second local map data. The calculation of localization accuracy relies on the deviation between the expected trajectory and the actual trajectory, and is comprehensively evaluated by combining indicators such as position offset, attitude angle error, and convergence rate. The spatial consistency index is reflected by comparing the correspondence of landmarks in overlapping areas of the local map, the degree of boundary contour fitting, and the matching rate of occupied grids. Higher localization accuracy indicates better alignment between the robot's motion control within its local area and the real environment; stronger spatial consistency indicates a lower probability of misalignment and ghosting when the local map is overlaid with other robot maps. By weighting and combining localization accuracy and spatial consistency index, a quantitative reliability score is generated, reflecting the robot's contribution and credibility in global map construction. Based on the reliability score, target weights are dynamically assigned to each robot in map fusion. The dynamic weight allocation mechanism is based on real-time localization accuracy and... Figure 1Consistency-based flexible adjustment: When a robot has extremely high localization accuracy and its local map exhibits excellent spatial consistency, its target weight will increase, enhancing its contribution to the global map. Conversely, if a robot's error increases due to temporary sensor failure or environmental interference, its weight will decrease to prevent unreliable data from negatively impacting the fusion result. The target weight is applied to the fusion process of the second local map data. During the fusion operation, coordinate alignment between different robot maps is achieved based on the calibrated second pose data. Rotation matrices and translation vectors unify each local map to the same global coordinate system. In spatially overlapping areas, a weighted average and consistency optimization method is used for merging. High-weighted robot data will play a dominant role in conflict areas, while low-weighted data will be weakened or marginalized as redundant information. For non-overlapping areas, the unique environmental information of different robots is preserved to ensure the integrity of the global map's coverage. After the above processing, the final fused global map data is generated.

[0041] The map and localization fusion method for multi-cooperative robots in embodiments of the present invention has been described above. The map and localization fusion system for multi-cooperative robots in embodiments of the present invention will be described below. Please refer to [link / reference]. Figure 2 One embodiment of the map and localization fusion system for multiple collaborative robots in this invention includes: The acquisition module 201 is used to acquire positioning error data and first local map data of multiple robots in the target cleaning area; The compensation module 202 is used to compensate the first local map data based on the positioning error data to obtain the second local map data; Error analysis module 203 is used to perform error analysis based on the second local map data and positioning error data to obtain error compensation parameters; The fusion module 204 is used to fuse the second local map data according to the error compensation parameters to obtain fused global map data.

[0042] Through the collaborative efforts of the aforementioned components, and by establishing an error coupling coefficient matrix that considers the relative distance between robots and the visual overlap area, a precise mathematical description of the error propagation mechanism in a multi-robot system was achieved for the first time. This overcomes the technical deficiency of traditional methods that ignore the mutual influence of errors between robots. A nonlinear map error compensation function was constructed using a radial basis function neural network, which can handle complex spatial error distributions more accurately than traditional linear compensation methods. Through a three-tiered architecture of geometric transformation compensation, topology optimization compensation, and semantic compensation, various errors in the map data are comprehensively eliminated. An integral finite-time sliding surface was designed, and an improved Riccati differential equation was solved, achieving precise convergence of multi-robot positioning errors within a preset finite time. This breaks through the technical bottleneck of traditional control methods that can only converge asymptotically, significantly improving the speed and accuracy of error compensation. A three-layer control architecture was constructed, comprising an optimal feedback control layer, a sliding mode control layer, and an integral compensation control layer. Different control layers handle deterministic errors, modeling uncertainties, and coupling propagation errors respectively, achieving optimized combination of control performance and robustness guarantees. A dynamic weight allocation mechanism based on real-time error compensation is established, adaptively adjusting the fusion weights according to the error suppression effect of each robot and the map quality evaluation results, thus solving the technical problem that traditional fixed-weight fusion cannot adapt to dynamic error changes. Geometric error compensation is achieved through rotation matrix transformation and translation vector correction, landmark topological relationship errors are handled through graph optimization, and different regions are identified and differentiated compensation strategies are executed through semantic segmentation networks, forming a complete multi-level error correction technology system. A fractional power sliding law algorithm is adopted to ensure finite-time convergence characteristics. Combined with adaptive gain adjustment and hyperbolic tangent function boundary layer design, control chattering is effectively suppressed while ensuring fast convergence, overcoming the inherent technical defects of traditional sliding mode control.

[0043] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0044] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0045] The above-described embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A map and localization fusion method for multi-cooperative robots, characterized in that, include: Acquire positioning error data and first local map data of multiple robots in the target cleaning area; The first local map data is compensated based on the positioning error data to obtain the second local map data. Specifically, the process includes: calculating the relative distance and visual overlap area between robots based on the real-time position coordinates of each robot; calculating the error coupling coefficient matrix by performing an exponential decay function on the influence intensity of error propagation among multiple robots based on the relative distance and the visual overlap area; inputting the positioning error data and the error coupling coefficient matrix into a recursive error propagation equation to perform dynamic modeling of cumulative error, obtaining an initial error propagation model; constructing a cumulative error propagation model containing a system state matrix, a control input matrix, and a process noise matrix based on the initial error propagation model; inputting the positioning error data into the cumulative error propagation model for recursive solution to obtain the cumulative positioning error vector for each robot, and performing time difference operation on the cumulative positioning error vector to obtain an error feature dataset; inputting the error feature dataset into a radial basis function neural network for nonlinear compensation to obtain map error compensation data for each robot; correcting the grid coordinates in the first local map data based on the map error compensation data to obtain corrected local map data; and updating the spatial distribution status of occupied grids, idle grids, and unknown grids based on the corrected local map data to obtain second local map data. Error analysis is performed based on the second local map data and the positioning error data to obtain error compensation parameters. Specifically, this includes: parsing clean path constraints and environmental boundary constraints from the second local map data to generate a zero-error expected trajectory; calculating the deviation between the positioning error data and the zero-error expected trajectory, and designing a sliding surface based on the deviation to generate a sliding surface parameter combination; performing pole placement based on the sliding surface parameter combination to obtain a convergence parameter configuration; combining the convergence parameter configuration with the error coupling coefficient matrix to form state equation coefficients, and substituting the state equation coefficients into the Riccati differential equation for recursive solution to obtain a positive definite symmetric matrix solution; generating an optimal feedback gain matrix based on the positive definite symmetric matrix solution, and constructing an optimal tracking controller using the optimal feedback gain matrix and the sliding surface parameter combination; and performing hierarchical control on the positioning error data based on the optimal tracking controller to obtain error compensation parameters. The second local map data is fused according to the error compensation parameters to obtain fused global map data. Specifically, this includes: applying the error compensation parameters to the positioning system of each robot to correct the current position coordinates and attitude angles of each robot to obtain second pose data; recalculating the positioning accuracy of each robot and the spatial consistency index of the second local map data based on the second pose data, and generating a reliability score based on the positioning accuracy and the spatial consistency index; dynamically allocating the target weight of each robot in map fusion according to the reliability score; applying the target weight to the second local map data to perform fusion, while handling the coordinate alignment and overlapping area merging between different robot maps to generate fused global map data.

2. The map and localization fusion method for multi-cooperative robots according to claim 1, characterized in that, The acquisition of positioning error data and first local map data of multiple robots in the target cleaning area includes: Data acquisition is performed by calling multiple LiDAR sensors, wheel encoder sensors, and inertial measurement unit sensors configured on the robot to obtain a multi-sensor dataset for each robot; Based on the multi-sensor dataset, Kalman filter fusion localization is performed in the target cleaning area to obtain the first pose data of each robot; Error calculation is performed based on the first pose data and the preset reference trajectory data to obtain the positioning error data of each robot; Based on the first pose data, the environmental point cloud data collected by the lidar sensor is processed into a rasterized SLAM mapping system to obtain the first local map data for each robot.

3. The map and localization fusion method for multi-cooperative robots according to claim 1, characterized in that, The step of correcting the grid coordinates in the first local map data based on the map error compensation data to obtain corrected local map data includes: The position correction, attitude correction, and rotation transformation parameters of each robot are analyzed from the map error compensation data to obtain the grid coordinate transformation parameters. Based on the grid coordinate transformation parameters, a rotation matrix transformation and translation vector correction are performed on each grid coordinate in the first local map data to obtain grid coordinate data. Based on the spatial relationship of landmarks in the raster coordinate data, graph optimization is performed to obtain landmark coordinate data. The landmark coordinate data is input into a semantic segmentation network to identify clean areas, obstacle areas, and navigation path areas, and corresponding error compensation strategies are executed to obtain corrected local map data. The semantic segmentation network uses a deep learning model to classify the input map data pixel by pixel or grid by grid to identify clean areas, obstacle areas, and navigation path areas in the map.

4. The map and localization fusion method for multi-cooperative robots according to claim 1, characterized in that, The step of performing hierarchical control on the positioning error data based on the optimal tracking controller to obtain error compensation parameters includes: The positioning error data is multiplied by the optimal feedback gain matrix in the optimal tracking controller to perform a linear quadratic optimal control operation, thereby generating the optimal feedback control component; When the positioning error data deviates from the sliding surface, the deviation distance from the error state to the sliding surface is calculated and a sliding mode control component is generated; When an error coupling propagation effect is detected among multiple robots, the historical error influence of each robot and the mutual propagation error are calculated, and an integral compensation control component is generated. The optimal feedback control component, the sliding mode control component, and the integral compensation control component are weighted and superimposed to form the error compensation parameter.

5. The map and localization fusion method for multi-cooperative robots according to claim 4, characterized in that, When the positioning error data deviates from the sliding surface, the step of calculating the deviation distance from the error state to the sliding surface and generating a sliding mode control component includes: The positioning error data is substituted into the sliding surface function for calculation to obtain the calculation result. When the calculation result is not equal to zero, it is determined that the error state deviates from the sliding surface and the deviation distance from the error state to the sliding surface is output. Calculate the fractional power of the deviation distance and combine it with the sign function to generate the basic sliding control quantity that drives the error trajectory to converge rapidly to the sliding surface; The sliding gain coefficient is generated based on the real-time rate of change of the deviation distance; The basic sliding control quantity is multiplied by the sliding gain coefficient, and a boundary layer compensation term composed of a hyperbolic tangent function is superimposed to eliminate high-frequency chattering, thus forming a sliding mode control component.

6. A map and localization fusion system for multiple collaborative robots, characterized in that, A map and localization fusion method for performing a multi-cooperative robot as described in any one of claims 1-5, comprising: The acquisition module is used to acquire positioning error data and first local map data of multiple robots in the target cleaning area; The compensation module is used to compensate the first local map data based on the positioning error data to obtain the second local map data; The error analysis module is used to perform error analysis based on the second local map data and the positioning error data to obtain error compensation parameters; The fusion module is used to fuse the second local map data according to the error compensation parameters to obtain fused global map data.

Images