Cleaning path planning method and cleaning device

Abstract

The present invention discloses a cleaning path planning method and a cleaning device, which belong to the field of automatic control technology. The method comprises the following steps: collecting point cloud data of an area to be cleaned, constructing an initial composite map, setting a perception marking method, dividing the point cloud data into sub-areas, marking mutation areas after matching with historical map features, and locally updating the composite map using graph optimization; constructing a safety buffer zone at the edge of the area to be cleaned, setting a buffer adjustment method, performing trajectory prediction on the mutation area, and dynamically adjusting the safety buffer distance according to the prediction result; calling a basic cleaning path, generating a hierarchical path, setting a planning adjustment method, updating the hierarchical path according to the mutation area during the cleaning process, calculating the pollution index of the area to be cleaned, constructing a time series prediction model, predicting the pollution diffusion trend, and dynamically adjusting it; after the area to be cleaned is cleaned, performing decision iteration and adapting to user habits.

Description

Technical Field

[0001] The invention relates to a cleaning path planning method and a cleaning device, and belongs to the technical field of automatic control. Background Art

[0002] With the development of environmental perception technologies such as lidar, visual sensors, and SLAM algorithms, cleaning robots have advanced from the early random collision mode to the stage of intelligent autonomous navigation based on environmental maps. By integrating path optimization algorithms with multi-sensor data, they can achieve efficient coverage cleaning in dynamic environments. At the same time, the integration of AI semantic recognition, multi-viewing and other technologies enables them to adapt to the autonomous cleaning needs of complex scenarios such as shopping malls and hotels.

[0003] However, existing technologies are still limited by the lack of real-time adaptation capabilities of mobile objects in dynamic unstructured environments, and face multiple challenges: when the environmental topology changes suddenly, such as the movement of obstacle positions and changes in spatial morphology, existing SLAM algorithms rely on global map reconstruction mechanisms, resulting in path replanning response delays, and traditional control strategies have difficulty maintaining motion continuity during sudden environmental changes. In addition, the lack of dynamic obstacle adaptive control mechanisms makes it impossible for mobile objects to adjust safety buffer distances and obstacle crossing parameters in real time. The imbalance of multi-objective optimization in unstructured environments leads to a sudden increase in motion resistance and energy consumption. The lack of trajectory optimization strategies for special terrains further reduces the passability of complex terrains. Insufficient semantic understanding of environmental features leads to the inability to synchronize control parameters when dynamic attributes such as materials and terrain change, resulting in real-time and robustness defects, making it difficult to meet the autonomous adaptation needs of mobile objects in unstructured scenarios. Summary of the Invention

[0004] In response to the shortcomings of the existing technology, the purpose of the present invention is to provide a cleaning path planning method and a cleaning device, avoid global reconstruction through incremental SLAM, use LSTM to predict obstacle trajectories and adjust the safety buffer distance, optimize the path through RRT and gradient descent, combine material semantics and pollution index to dynamically adjust parameters, and use Q learning iterative optimization strategy to adapt to user habits, so as to solve the real-time and robustness problems.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] Cleaning path planning method, including:

[0007] Collect multi-source data of the area to be cleaned, build an initial composite map, set a perception marking method, divide the point cloud data into sub-areas, mark the mutation areas after matching with historical map features, and use graph optimization to locally update the composite map;

[0008] A safety buffer zone is constructed at the edge of the area to be cleaned, a buffer adjustment method is set, a trajectory prediction is performed on the mutation area, and the safety buffer distance is dynamically adjusted according to the prediction result, and the motion trajectory is adjusted in real time through the crawler differential drive control;

[0009] Call the basic cleaning path, generate a hierarchical path, set the planning adjustment method, and during the cleaning process, update the hierarchical path in real time according to the sudden change area, calculate the pollution index of the area to be cleaned, build a time series prediction model, predict the pollution diffusion trend, and make dynamic adjustments;

[0010] After the area to be cleaned is cleaned, decision iteration is performed to adapt to user habits.

[0011] Preferably, the specific steps of constructing the initial composite map include:

[0012] Before cleaning the area to be cleaned, trigger a function self-check;

[0013] The point cloud data and depth image are acquired through the LiDAR and depth camera, and the surface pressure distribution matrix is ​​generated by combining the pressure sensor, and the material thickness distribution matrix is ​​acquired by the capacitive thickness sensor;

[0014] Perform statistical filtering, voxel downsampling and bilateral filtering on the collected data, and complete spatiotemporal alignment through the external parameter matrix to generate a fused point cloud;

[0015] The RANSAC algorithm is used to fit the plane area, filter out the plane point set in the fused point cloud, and generate the plane area;

[0016] The Canny edge detection algorithm is used to extract the contour edge, calculate the point cloud normal vector, and filter out the edge area point set in combination with the gravity direction to generate the edge area;

[0017] Eliminating plane point sets and edge area point sets from the fused point cloud, and dividing the undulating area using a region growing algorithm;

[0018] Based on octree partitioning, potential hidden areas are screened out, and the boundaries of hidden areas are inferred by combining Bayesian filtering to generate hidden areas;

[0019] The initial composite map is constructed through the incremental SLAM algorithm, which includes semantic labels with planes, edges, undulations, and hidden areas, and keyframe-associated geometric features and pose information.

[0020] Preferably, the perceptual marking method comprises:

[0021] Acquire real-time point cloud data, reduce density through voxel grid filtering, and automatically adjust the octree resolution according to the point cloud density to generate variable resolution sub-region grids;

[0022] For each sub-region, use the pruned and compressed PointNet++ network to extract the composite vector and update the sub-region grid;

[0023] Calculating the overlap rate by using the octree intersection volume ratio, and extracting the historical key frames overlapping with the current sub-region in the initial composite map;

[0024] Using the Transformer attention network, the current sub-region features are mapped into a query vector, and the historical keyframe features are mapped into a key vector and a value vector to calculate the attention weight;

[0025] Combining bidirectional nearest neighbor matching and weighted Euclidean distance, valid point pairs are screened and the displacement of the sub-region is calculated using the least squares method;

[0026] Use a sliding window to store historical mutation displacement data, and dynamically adjust the mutation detection threshold through the incremental variance formula;

[0027] Subregions whose displacement exceeds the mutation detection threshold are screened out as candidate mutation regions, and connected domain analysis is performed to generate a continuous set of candidate mutation regions.

[0028] Preferably, the perceptual marking method further comprises:

[0029] For each candidate mutation region, query the historical detection results of the previous three frames. If it is a single-frame mutation, calculate the similarity with the historical mutation pattern and eliminate candidate mutation regions with similarity less than the similarity threshold;

[0030] Perform spatial verification of edge areas and undulating areas, calculate the convex hull volume change rate, and eliminate candidate mutation areas whose absolute value of the change rate is lower than the change threshold;

[0031] The grayscale change rate of the depth image is used for screening, and the confidence is enhanced. The Euclidean distance of the spectral features is calculated to screen and generate a set of effective mutation regions.

[0032] For the effective mutation area set, calculate the point cloud mutation ratio, the grayscale change rate of the depth image, and the material spectrum mutation degree, and fuse them according to the weight to obtain the multimodal mutation level;

[0033] Retrieve historical optimization cases whose similarity with the current mutation pattern exceeds a priori threshold from the memory pool as priors;

[0034] Construct a local pose graph containing point cloud reprojection constraints and history memory constraints, and use the g2o optimizer to solve it to update the local topological relationship of the mutation area;

[0035] The semantic confidence in the initial composite map is updated according to the mutation frequency and a lifecycle label is assigned to generate a fused dynamic composite map.

[0036] Preferably, the specific steps of constructing a safety buffer zone include:

[0037] Collect multi-source data, align them in time and space through the sensor extrinsic matrix, and generate edge point sets with confidence by weighted fusion;

[0038] Based on the edge point set, the warning layer, deceleration layer and braking layer are generated by expanding outward along the normal vector direction, which trigger visual warning, speed attenuation function calculation and emergency braking respectively.

[0039] Extracting a terrain slope from the dynamic composite map, and increasing the outward expansion distance of each layer once the terrain slope exceeds a slope threshold;

[0040] An infrared spectrometer collects spectral data of ground materials and calculates humidity values. Once the humidity value exceeds the humidity threshold, the braking distance is increased and the anti-skid driving mode is activated.

[0041] Compare the edge load data collected in real time with the historically stored edge load data. When the deviation exceeds the deviation threshold, increase the safety buffer zone expansion distance. At the same time, if the edge point cloud displacement is detected to exceed the displacement threshold, the safety buffer zone moves in real time with the edge.

[0042] Preferably, the buffer adjustment method includes:

[0043] The QuickHull algorithm is used to generate a three-dimensional convex hull and an axis-aligned bounding box, and the geometric eigenvector of the convex hull is calculated to characterize the regularity of the obstacle.

[0044] The historical position of the obstacle is extracted, and after removing the static frames, the trajectory is smoothed through Kalman filtering to generate a motion feature vector containing velocity, acceleration, and steering angle;

[0045] Build a motion trend model, output the predicted trajectory for the future period, and generate an error ellipse by combining Monte Carlo Dropout;

[0046] Based on the obstacle movement speed and prediction uncertainty, the safety buffer distance is calculated by superimposing speed compensation and uncertainty compensation on the basic buffer distance. , and determine the advance adjustment time;

[0047] Calculate the relative distance between the current position and the predicted obstacle position and azimuth;

[0048] like , maintain the base speed ;like , start steering control;

[0049] Real-time calculation of slip rate , and compensate for the speed to generate a PWM signal to control the motor.

[0050] Preferably, the planning adjustment method includes:

[0051] A three-dimensional search space is generated with the mutation area as the center, and a grid map to be passed is constructed by octree partitioning;

[0052] Obtain the predicted obstacle position and use a dynamic bias sampling strategy to guide the sampling point to offset the predicted obstacle position;

[0053] Using the RRT algorithm, a set of candidate paths is generated, and the smoothness of each candidate path is calculated, and the one with the smallest smoothness is retained. Paths;

[0054] A multi-objective function is constructed based on path distance, energy consumption and terrain resistance, and the path parameters are optimized using the gradient descent method to generate the optimal path.

[0055] Preferably, the planning adjustment method further includes:

[0056] The infrared spectrometer collects data and identifies the material type, and combines it with pressure data to verify consistency. If the consistency verification fails, the material type is re-identified, otherwise the cleaning parameters are dynamically adjusted;

[0057] Calculate pollution indices, predict pollution trends and trigger compensation strategies;

[0058] Convert the hand-drawn key area coordinates to the map coordinate system, extract the geometric features, generate the grid path, and output the path priority index;

[0059] A state space and action space were established, and the DQN algorithm was used to iteratively optimize path parameters based on historical cleaning data and user logs to adjust the visit frequency of high-frequency residual areas; among them, high-frequency residual areas are areas with fixed spatial positions and where contamination remains after completing a preset number of cleaning times.

[0060] The cleaning device includes: a perception module, a planning module and an execution control module;

[0061] The perception module is used to collect multi-dimensional data through multi-sensor collaboration, eliminate noise through pre-processing and complete spatiotemporal alignment;

[0062] The planning module is used to construct an initial composite map based on the initial scan data and dynamically update it, generate a hierarchical path through the RRT algorithm and multi-objective optimization, dynamically adjust path parameters based on material semantic recognition and pollution index, and use a reinforcement learning algorithm to analyze user operation data to iteratively optimize the path strategy;

[0063] The execution control module is used to fuse multi-sensor edge data to generate an adaptive safety buffer zone, dynamically adjust the safety buffer distance, predict the obstacle movement trend through LSTM, and correct the movement trajectory in combination with track differential drive.

[0064] Preferably, the planning module includes a mapping unit, a path optimization unit and a learning unit;

[0065] The mapping unit constructs and updates a composite map containing semantic labels through an incremental SLAM algorithm and dynamic mutation detection, uses a variable-resolution octree and spatiotemporal attention matching to detect environmental topology changes in real time, and updates the local topological relationship of the mutation area through graph optimization;

[0066] The path optimization unit generates a hierarchical path based on the semantic labels of the initial composite map, adopts the RRT algorithm and multi-objective optimization, adjusts the path planning according to dynamic parameters, updates the path parameters in real time in response to environmental changes, and predicts the diffusion trend based on the pollution index;

[0067] The learning unit uses user-defined area planning and reinforcement learning algorithms to dynamically adjust path priorities based on historical cleaning data and user habits, and iteratively optimize the cleaning strategy.

[0068] Beneficial effects of the present invention:

[0069] Through incremental SLAM, a dynamic composite map is constructed, and only the mutation area is updated to avoid global reconstruction, shortening the path planning response time, solving the SLAM algorithm delay problem, and improving real-time performance; the LSTM neural network is used to predict the trajectory of dynamic obstacles, dynamically adjust the safety buffer distance, and combine the track differential drive control to achieve adaptive obstacle avoidance and make up for the lack of dynamic obstacle control; the RRT algorithm is used to generate candidate paths, and the multi-objective function balances distance, energy consumption and terrain resistance to solve the imbalance of unstructured environment optimization and improve path efficiency and complex terrain passability; material semantic recognition and pollution index analysis are performed, path parameters and auxiliary functions are dynamically adjusted, and path strategies are iteratively optimized based on historical data and user habits to enhance the adaptability to the dynamic properties of the environment, realize the autonomous adaptation of mobile objects in unstructured scenes, and improve real-time performance and robustness. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 Schematic diagram of the cleaning path planning method;

[0071] Figure 2 is a flow chart of the perceptual marking method of the present invention;

[0072] Figure 3 is a flow chart of the buffer adjustment method of the present invention;

[0073] Figure 4A flow chart of the planning and adjustment method of the present invention;

[0074] Figure 5 This is the structural diagram of the cleaning device. DETAILED DESCRIPTION

[0075] The technical solution of the present invention is described in detail below through the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations on the technical solution of the present invention. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0076] Example 1:

[0077] refer to Figures 1 to 4 As shown, this embodiment introduces a cleaning path planning method, which includes the following steps:

[0078] The robot is started to perform cleaning operations, and the surface point cloud data and depth images of the area to be cleaned are collected. The bearing capacity and material thickness of the surface of the area are perceived in real time, realizing multi-dimensional perception of surface morphology, bearing capacity and material properties in unstructured environments. An incremental SLAM algorithm is used to construct an initial composite map based on the first scan. The initial composite map includes flat areas, undulating areas, edge areas and hidden areas. A perception marking method is set to divide the real-time point cloud data into sub-region grids, and feature matching is performed with the historical map. Sub-regions with displacements exceeding the threshold are screened out and marked as mutation areas. The environmental mutation level is judged by the proportion of mutation areas. The local topological relationship is updated through graph optimization to update the mutation areas in the initial composite map. A dynamic composite map is output to avoid path planning delays caused by global reconstruction and meet the real-time mapping requirements in dynamic environments.

[0079] The system synchronously integrates edge detection sensor data to construct a safety buffer zone at the edge of the area to be cleaned, accurately identifies the surface boundary of the area to be cleaned, sets a buffer adjustment method, and uses an LSTM neural network to predict the trajectory of obstacles in the mutation area. It generates a motion trend model for future time periods and dynamically adjusts the safety buffer distance based on the prediction results. This allows for early adjustment of the safety buffer distance, solves the problem of lack of adaptive control for dynamic obstacles, and simultaneously adjusts the motion trajectory through track differential drive control to avoid the risk of collision with dynamic obstacles.

[0080] The basic cleaning path is called to generate a composite map-driven hierarchical path, and the area to be cleaned is actively cleaned. A planning adjustment method is set. During the cleaning process, the hierarchical path is updated in real time according to the identified mutation area. The RRT algorithm is used to generate a set of candidate paths in the mutation area. The multi-objective function including distance, energy consumption, and terrain resistance is solved by the gradient descent method to achieve real-time update of the path, solving the optimization imbalance problem in unstructured environments, improving path planning efficiency, performing material semantic recognition, and dynamically adjusting path parameters such as speed and driving force. At the same time, the pollution index of the area to be cleaned is calculated based on environmental sensor data, and a time series prediction model is constructed to predict the pollution diffusion trend and dynamically adjust auxiliary functions and path priorities.

[0081] After the cleaning area is completed, decision iteration and user habit adaptation are carried out. First, based on the historical cleaning trajectory and environmental detection data, the path strategy is iteratively optimized through the Q learning algorithm, and the visit frequency of high-frequency residual functional areas is automatically increased to improve the cleaning coverage of key areas. At the same time, user operation data is periodically analyzed. When the priority setting of a specific area is detected multiple times in a row, the path planning weight of the area is automatically increased to achieve dynamic matching of mobile body strategy and user habits, thereby improving scene adaptation efficiency.

[0082] Preferably, the specific steps of constructing the initial composite map include:

[0083] Because the soft surface of the bed requires high sensor accuracy, for example, the pressure sensor must distinguish between the load differences between bedding and mattresses. Strict self-tests ensure subsequent data reliability and avoid map distortion due to sensor errors. The robot is powered on, the main control system is activated, and the sensor self-test program is triggered. The 3D LiDAR, depth camera, pressure sensor, and capacitive thickness sensor are tested for functionality in sequence. The 3D LiDAR requires verification of the stability of the transmitted and received signals, the depth camera requires confirmation of image clarity, the pressure sensor requires calibration of the sampling circuit, and the capacitive thickness sensor requires accuracy calibration.

[0084] After passing the self-test, the robot moves to the preset starting position of the area to be cleaned using the wheel encoder. It then uses a 3D lidar to emit laser beams at a specific frequency to scan the environment horizontally and vertically, acquiring point cloud data containing 3D coordinates and reflection intensity. Simultaneously, the depth camera captures depth images at a specific frame rate and generates point cloud depth information through binocular parallax calculation, thereby obtaining a depth image sequence.

[0085] Pressure sensors are used to sense the surface's bearing capacity in real time, with sampling point spacing controlled within 2 cm. This generates a surface pressure distribution matrix, capturing local pressure changes at the folds of the quilt. Capacitive thickness sensors are used to scan along the movement trajectory to obtain fabric thickness data, ultimately forming a material thickness distribution matrix that distinguishes the thickness differences between sheets and thick duvets, providing multi-dimensional data support for subsequent environmental modeling.

[0086] Preprocess the point cloud data and depth image, and use statistical filtering algorithms to remove outliers in the point cloud data, including calculating the distance from each point in the point cloud data to its 50 nearest neighbors and calculating the mean of the distances. and standard deviation , set the threshold to , remove points whose distance exceeds the threshold to filter out abnormal points, and then use voxel grid filtering for downsampling, including setting the voxel size, dividing the point cloud data into different voxels, calculating the mean of the points in each voxel as the representative point, performing bilateral filtering on the depth image, retaining the wrinkle edges of the bedding, eliminating fabric texture noise, and completing the spatiotemporal alignment of the preprocessed point cloud data and the depth image through the external parameter matrix to generate a fused point cloud;

[0087] Since the plane area is the basic feasible domain for path planning, prioritizing its identification can quickly construct the map skeleton and reduce the amount of subsequent calculations. The RANSAC algorithm is used to fit the plane. The iteration and fitting distance thresholds are set. Based on the constructed fitting plane, the vertical distance from the point cloud to the fitting plane is calculated. Points whose vertical distance does not exceed the fitting distance threshold are selected as the plane point set. The plane point set is divided into main planes and scattered planes according to the proportion of the plane point set. When the plane point set accounts for more than 30%, it is marked as a main plane area, such as a standard mattress body. Otherwise, it is classified as a scattered plane area, such as a local plane formed by stacked bedding. The fitting distance threshold is set based on the average height of the bed folds to ensure the complete extraction of the mattress body plane.

[0088] Because edge areas present a risk of falling, safety buffer zones can be constructed after marking to prevent the robot from crossing the boundary. Based on the preprocessed depth image, the Canny edge detection algorithm is used to extract contour edge pixels, such as the edges of bedding and mattress boundaries. The contour edge pixels are mapped to the point cloud coordinate system, and the point cloud normal vector is calculated. The direction of gravity is combined to determine whether it is a physical edge. Once the angle between the point cloud normal vector and the direction of gravity exceeds 45°, it is determined to be a physical edge. The physical edges are integrated to generate an edge area point set to generate the edge area.

[0089] Because bedding wrinkles and the undulating and hidden areas under pillows are blind spots for cleaning, semantic segmentation can improve the coverage accuracy of special areas. Planar point sets and edge area point sets are removed from the fused point cloud. The remaining point cloud is divided into continuous undulating areas using the region growing algorithm. The average slope of each area is calculated to evaluate the terrain resistance.

[0090] Leveraging the sparsity of LiDAR point clouds, the system uses octree spatial partitioning to identify shadowed areas where the laser cannot directly illuminate. Combining this with the parallax information from the binocular camera, the system marks areas in the depth image with a confidence level less than a preset confidence threshold as potentially hidden areas. Bayesian filtering is then used to fuse laser and visual data to infer the boundaries of hidden areas and complete missing parts of the map.

[0091] Incremental SLAM is used to construct an initial composite map, including a fused point cloud based on the first scan. The ORB-SLAM3 algorithm is used to construct a keyframe sequence. Each keyframe contains a pose. The geometric features of planes, undulations, edges, and hidden areas are associated with the keyframe poses, including point sets, equations, and boundaries, and stored as an initial composite map with semantic labels.

[0092] Preferably, the perceptual marking method comprises:

[0093] Since the global point cloud matching calculation is too large, dividing the sub-regions can realize distributed feature matching and improve real-time performance; obtain real-time point cloud data collected by multiple sensors, reduce the point cloud density through voxel grid filtering, and automatically adjust the resolution of the octree division according to the point cloud density to generate a sub-region grid with variable resolution. Each leaf node stores the sub-region point cloud and spatial coordinate range; Among them, automatically adjusting the resolution of the octree division includes calculating the point cloud density and setting a secondary density threshold to divide the high-density area, the regular density area and the low-density area. Different density areas are configured with different resolutions, and the resolution can be set by those skilled in the art. For example, the high-density area corresponds to pillows and bedding wrinkles, and the octree resolution is set to 0.5cm. The regular density area corresponds to the mattress plane, and the resolution is set to 1cm. The low-density area corresponds to the gap under the bed, and the resolution is set to 2cm.

[0094] For each sub-region, a lightweight PointNet++ network is used to extract a composite vector containing semantic and geometric features, thereby outputting a sub-region mesh. The semantic features are processed through a softmax layer to output the material semantic probability, such as the probability of a pillow being 0.8. The geometric features include curvature and normal vectors.

[0095] Extract historical keyframes that overlap with the current sub-region in the initial composite map. Calculate the octree intersection volume ratio as the overlap rate. When the overlap rate exceeds the preset overlap threshold, it is considered an overlap. Each keyframe contains a pose and feature vector. Perform spatiotemporal attention matching on the current sub-region features and the historical keyframe features in the initial composite map. Construct a Transformer attention network, map the current sub-region features to a query vector, and map the historical keyframe features to a key vector and a value vector, thereby calculating the attention weight. Since this embodiment uses an 8-head Transformer attention network, when calculating the attention weight, first calculate the single-head attention weight, then splice the 8-head outputs and linearly map them to obtain the final attention weight, thereby focusing on salient features such as the pillow edge and bedding wrinkles.

[0096] For the current sub-region point cloud and the historical point cloud, bidirectional nearest neighbor matching is used. For each point in the current sub-region point cloud, the point with the closest Euclidean distance is found in the historical point cloud to form a candidate point pair. The feature descriptor distance of the candidate point pair is calculated, and only point pairs whose feature descriptor distance does not exceed the preset distance threshold are retained to reduce mismatching;

[0097] For the candidate point pairs after screening, the weighted Euclidean distance is calculated in combination with the attention weight. A weighted distance threshold is set, and the point pairs whose weighted Euclidean distance is less than the weighted distance threshold are screened as valid point pairs. The sub-region displacement is calculated for the valid point pairs using the least squares method, thereby obtaining a sub-region displacement list, which provides a quantitative basis for the screening of mutation regions.

[0098] Set the size of the sliding window to store historical mutation displacement data, use the incremental variance formula to update the historical variance in real time, and dynamically adjust the mutation detection threshold based on the set basic threshold. For example, in highly dynamic scenes, the variance increases, which increases the mutation detection threshold and reduces false detections caused by temporary movement. In static scenes, the threshold remains unchanged to maintain sensitivity. The expression is as follows:

[0099] ;

[0100] Where, For time, is the basic threshold; is the adjustment coefficient to balance sensitivity and robustness, is the historical variance;

[0101] Based on the sub-region displacement list, sub-regions whose displacement exceeds the mutation detection threshold are selected as candidate mutation regions. Connected domain analysis is then performed to merge spatially adjacent candidate mutation regions with consistent displacement directions to generate a continuous set of candidate mutation regions. Each candidate mutation region records the point cloud, displacement vector, and spatial volume.

[0102] Each candidate mutation region is verified in the time dimension, and the historical detection results of the previous three frames are queried to analyze the mutation consistency. If three consecutive frames are marked as mutations and the displacement trajectory is consistent, such as pillow translation, the candidate mutation region is retained. If only a single frame mutation occurs, the similarity with the historical mutation pattern is calculated, and candidate mutation regions with similarity less than the similarity threshold are eliminated to eliminate single-frame noise mutations.

[0103] Simultaneously, spatial dimension verification is performed to estimate the neighborhood normal vectors of the point cloud in the candidate mutation region set. The percentage of points whose normal vectors have an angle greater than 45° with the gravity direction is counted. Once the percentage of points exceeds the threshold, it is determined to be an edge region; otherwise, it is an undulating region. The convex hull volume change rate is calculated, and candidate mutation regions with an absolute value of the change rate below the change threshold are eliminated.

[0104] The candidate mutation region is projected onto the depth image, the grayscale change rate is calculated, and the regions where the grayscale change rate exceeds the grayscale threshold are screened out and the confidence is enhanced. At the same time, the spectral features of the candidate mutation region are extracted, the Euclidean distance with the historical spectrum is calculated, and the regions where the distance is less than the distance threshold are eliminated, thereby generating a set of valid mutation regions;

[0105] For the effective mutation area set, the point cloud mutation ratio, the grayscale change rate of the depth image, and the material spectrum mutation degree are calculated, and the multimodal mutation level is obtained by weighted fusion. Once the calculated mutation level exceeds the preset value, it is determined to be a severe mutation, such as a quilt being completely opened;

[0106] Retrieve historical optimization cases similar to the current mutation pattern from the memory pool, calculate the cosine similarity between the current mutation region features and the memory pool features, and once the similarity exceeds the prior threshold, extract the corresponding historical keyframe poses and constraint relationships as priors. The memory pool is used to store historical graph optimization results.

[0107] Construct a local pose graph, set the objective function including point cloud reprojection constraints and historical memory constraints, use the g2o optimizer to solve, update the local topological relationship of the mutation area, and output the optimized pose and topology structure; the local pose graph contains the key frames related to the mutation area and their adjacent key frames, the nodes are camera poses, and the edges are point cloud reprojection constraints and historical memory constraints;

[0108] For each element in the initial composite map, the semantic confidence is updated based on the mutation frequency of the last 10 scans, and a life cycle label is assigned, including nascent (such as a new pillow), mature (stable area), and declining (long-term unchanged area). The life cycle label is divided by setting a secondary confidence threshold;

[0109] The corresponding part of the initial composite map is replaced with the optimized geometric features of the mutation area, the non-mutation area is retained, and semantic confidence, lifecycle labels, real-time materials, and carrying data are added to generate a dynamic composite map.

[0110] Preferably, the specific steps of constructing a safety buffer zone include:

[0111] After the robot starts, it simultaneously collects data from the lidar, binocular camera, and infrared sensor. After voxel filtering and outlier removal of the lidar point cloud, edge points whose normal vectors are at an angle exceeding 45° to the direction of gravity are extracted. The depth image is denoised using bilateral filtering, and the Canny algorithm is used to extract contours and map them into three-dimensional space. Infrared sensor data is used to complete the edges of low-reflective objects. The three types of edge point sets are spatially and temporally aligned using the sensor extrinsic parameter matrix and fused according to weights to generate a confidence-based edge point set.

[0112] Since traditional fixed buffer zones cannot adapt to complex terrain, layered adaptive buffer zones can dynamically adjust the safety margin according to the environment to reduce the risk of collision. Based on the edge point set, a three-layer buffer zone is constructed, including the edge point set as the basis, and the buffer zone is expanded along the normal vector direction. At this time, the visual warning is triggered, the top LED light flashes at a frequency of 2Hz, and the voice prompts "approaching the edge" and expands outward again. A deceleration layer is generated based on the speed attenuation function defined at this time to drive the speed attenuation, and finally expand outward. A braking layer is generated after a certain distance, and emergency braking is initiated;

[0113] The terrain slope at the current location is extracted from the dynamic composite map. Once the terrain slope exceeds the slope threshold, the outward expansion distance of each layer is increased by 20% to compensate for the risk of slope slippage. At the same time, an infrared spectrometer is used to collect spectral data of the ground material and calculate the humidity value. Once the humidity value exceeds the humidity threshold, the braking layer distance is increased by 50% and the anti-skid driving mode is activated;

[0114] The track pressure sensor is used to collect load-bearing data in real time and compare it with the edge load-bearing data stored in the dynamic composite map. When the deviation exceeds the deviation threshold, such as the transition from a mattress to bedding, the outward expansion distance of the safety buffer zone increases by 10%-20%. If the edge point cloud displacement exceeds the displacement threshold, such as the bedding being pulled, the safety buffer zone moves in real time with the edge to maintain a safe distance.

[0115] Preferably, the buffer adjustment method includes:

[0116] Based on the mutation areas of the dynamic composite map, areas with the semantic label "dynamic obstacles" are screened, such as pillows and moving bedding. The QuickHull algorithm is used to generate a three-dimensional convex hull. The convex hull vertices are recursively found to generate the minimum convex polyhedron containing all the convex hull vertices. The volume of the convex hull is also calculated. The principal axis direction vector is obtained through the eigendecomposition of the covariance matrix to represent the longest axis direction of the obstacle. An axis-aligned bounding box is generated, recording the length, width, and height dimensions. The standard deviation of the angle between the convex hull surface normal vector and the gravity direction is calculated to characterize the regularity of the obstacle. If the standard deviation is less than the standard threshold, it is judged as a regular obstacle, such as a pillow. Otherwise, it is an irregular obstacle, such as a wrinkled bedding, and a geometric feature vector is generated.

[0117] Extract the obstacle's most recent 20-frame historical poses from the positioning cache, remove static frames with displacements less than a threshold, and treat them as noise. Keep the valid trajectory point set, smooth the trajectory using a Kalman filter, and calculate the velocity, acceleration, and steering angle to generate a motion feature vector.

[0118] Based on a bidirectional LSTM network, a motion trend model is constructed. By inputting historical trajectories, a predicted trajectory for the future period is output. Monte Carlo Dropout is used to generate a 95% confidence error ellipse to capture the nonlinear motion trends of obstacles.

[0119] Since the movement speed and prediction uncertainty of dynamic obstacles directly affect the safety buffer distance, dynamic adjustment can avoid the conservatism of fixed safety buffer zone. The basic buffer distance is set to , obtain the safety buffer distance through speed compensation and uncertainty compensation , and calculate the advance adjustment time , ensuring that the obstacle avoidance action is executed in advance; the expression is as follows:

[0120] ;

[0121] ;

[0122] ;

[0123] Where, is the robot's cleaning radius, is the safety margin, is the speed coefficient, To predict the velocity of the trajectory, is the standard deviation coefficient, is the radius of the major axis of the error ellipse, is the current cleaning speed, Forecast lead time;

[0124] Calculate the relative distance between the robot's current position and the predicted obstacle position , and the azimuth, which is the angle between the robot and the obstacle; if , maintain the base speed , that is, ;like , start the steering control, if you turn left, the left track slows down and the right track accelerates, that is, 、 ;in, 、 are the left and right track speeds respectively, is the steering coefficient, The angle between the robot's forward direction and the direction of the obstacle, reflecting the degree of lateral deviation of the obstacle relative to the robot;

[0125] Calculate the slip rate in real time , and compensate for the speed to generate a PWM signal to control the motor and correct the trajectory to avoid collision; the expression is as follows:

[0126] ;

[0127] ;

[0128] ;

[0129] Where, is the speed calculated by the track encoder, is the speed calculated by the visual odometry, 、 are the left and right track speeds after compensation respectively.

[0130] Preferably, the specific steps of generating the hierarchical path include:

[0131] A hierarchical path is generated based on the semantic labels in the dynamic composite map. In the plane area, if it is a single bed plane, a Z-shaped path is called and the cleaning parameters, including row spacing, speed and corner speed, are set. If it is a double bed plane, a segmented A-shaped path is called and the segment length is set to avoid missing wrinkles during long-distance straight cleaning. In the undulating area, a spiral path is used and the initial radius, expansion distance and maximum radius are set. In the edge area, a parallel edge tracking path is used to maintain a safe distance from the edge. In the hidden area, a combination of grid coverage and spiral in-depth is adopted. For example, the area under the pillow is first filled with a spiral with a radius of 0.15m, and then cleaned back and forth along the grid line. The cleaning frequency is automatically increased to 2 times.

[0132] Preferably, the planning adjustment method includes:

[0133] Obtain the detected mutation area and expand it outward a certain distance from the mutation area to generate a three-dimensional search space. Use the octree structure to divide the three-dimensional search space, mark the leaf nodes occupied by obstacles, and construct a grid map to be passed.

[0134] To enable path planning to avoid the future position of obstacles in advance, the predicted obstacle position is obtained. A dynamic bias sampling strategy is used to guide the sampling points to offset to the predicted obstacle position, thereby improving the efficiency of dynamic obstacle avoidance. The expression is as follows:

[0135] ;

[0136] Where, is the sampling point coordinate after bias, are the randomly generated original sampling point coordinates, is the current position of the obstacle, is the predicted position of the obstacle at the future moment, is the bias coefficient, which is used to control the degree of deviation from the predicted obstacle position. ;

[0137] Using the RRT algorithm, a set of candidate paths is generated, the smoothness of each candidate path is calculated and sorted, and the one with the smallest smoothness is retained. paths to update the candidate path set;

[0138] For each candidate path in the candidate path set, a distance target is first constructed based on the ratio of the candidate path length to the reference path length to measure path efficiency. An energy consumption target is then constructed by comprehensively considering cleaning factors. A terrain resistance target is then constructed by combining material resistance weights and terrain slope. An objective function is generated through weight configuration to calculate the functional components of the candidate path. Multi-objective optimization is then used to balance the shortest distance, lowest energy consumption, and terrain adaptability, ensuring that path connectivity is prioritized in the event of severe mutations and avoiding cleaning blind spots. Once a severe mutation is identified, the weights in the objective function are adjusted using an S-shaped function, and the path parameters are optimized using the gradient descent method to generate the optimal path.

[0139] An infrared spectrometer is used to collect band data. After dimensionality reduction through principal component analysis, the Euclidean distance between the real-time collected data and the characteristic values ​​in the dynamic composite map is calculated to determine whether the material has changed. Once the Euclidean distance exceeds the Euclidean distance threshold, it is determined that the material has changed. At this time, various material parameters are retrieved from the material library for matching, the specific material type is identified, and the cleaning parameters are dynamically adjusted for different materials. For example, for carpet materials, the speed is reduced and the driving force is increased, while for silk materials, the speed is increased and the driving force is reduced.

[0140] At the same time, through the consistency verification of pressure and material, when the consistency is lower than the consistency threshold, the material re-identification is immediately triggered to ensure the accuracy of parameter adjustment;

[0141] The pollution index is calculated by integrating dust concentration, stain area ratio, and formaldehyde concentration. A pollution trend prediction model is then used to output pollution trends for future periods. Once the pollution index or trend slope exceeds the compensation threshold, a compensation strategy is activated to increase cleaning frequency, UV intensity, and fragrance volume to proactively address pollution spread.

[0142] To meet the personalized needs of scenarios such as hotels, user-defined paths can be used to improve manual planning efficiency. The coordinates of key areas hand-drawn by users through the APP are converted to the map coordinate system. At the same time, regional geometric features, including area, perimeter, and convex hull, are extracted to determine the regularity of the area. The RRT algorithm is used to increase the sampling density and generate a grid path with fixed spacing to improve the area coverage accuracy. The path priority index is also output to meet personalized cleaning needs.

[0143] A historical cleaning database is established to record the optimal path parameters for different bed types, fabrics, and pollution scenarios. A state space including bed type, material, pollution index, and historical coverage rate, as well as an action space of speed, driving force, row spacing, and frequency are established. The coverage rate, energy consumption, and time are balanced through the reward function, and the DQN algorithm is used for iterative optimization. According to historical cleaning data and user operation logs, the access frequency of high-frequency residual areas is automatically adjusted to improve the scene adaptation efficiency. Among them, the high-frequency residual area is an area with a fixed spatial position and where pollution residues still appear after completing the preset cleaning times.

[0144] Example 2:

[0145] See also Figure 5 , another embodiment provided by the present invention: a cleaning device, comprising: a perception module, a planning module and an execution control module;

[0146] The perception module is used to collect multi-dimensional data of unstructured environments in real time, such as surface morphology, load-bearing capacity, and material properties. The collaborative work of multiple sensors addresses the incomplete data from a single sensor in traditional solutions, enabling accurate capture of complex features. By integrating various devices, such as lidar, depth cameras, pressure sensors, and spectrometers, it simultaneously acquires 3D point clouds, depth images, pressure distribution, and spectral features. Pre-processing eliminates noise and completes spatiotemporal alignment, providing reliable data support for subsequent mapping and planning, avoiding map distortion caused by sensor errors and improving the accuracy and completeness of environmental modeling.

[0147] The planning module is used to construct an initial composite map based on the initial scan data, dynamically update the composite map, avoid path planning delays caused by global reconstruction, and meet the real-time mapping needs in dynamic environments. At the same time, it generates a hierarchical path based on the semantic labels of the composite map, and realizes real-time path updates through the RRT algorithm and multi-objective optimization. It combines material semantic recognition and pollution index calculation to dynamically adjust path parameters and auxiliary functions, solve the optimization imbalance problem in unstructured environments, and improve path efficiency and cleaning effects. In addition, through user-defined area processing and reinforcement learning algorithm iterative optimization of path strategies, and analysis of user operation data to dynamically match mobile body strategies with user habits, it solves the problems of insufficient personalized services and low scene adaptation efficiency of traditional solutions, and improves the cleaning coverage rate of key areas and user operation convenience.

[0148] The execution module is used to build a dynamic safety buffer zone and achieve obstacle avoidance. By fusing multi-sensor edge data to generate an adaptive buffer area, it solves the problem of poor adaptability of traditional fixed buffer zones in complex terrain. The safety buffer distance is dynamically adjusted according to the terrain slope and material moisture. Combined with LSTM to predict the movement trend of obstacles, the safety distance is adjusted in advance and the motion trajectory is corrected through track differential drive to avoid the risk of collision with dynamic obstacles, ensure the stability of the robot's operation on soft surfaces, and achieve zero-fall safety control.

[0149] Preferably, the planning module includes a mapping unit, a path optimization unit and a learning unit;

[0150] The mapping unit is used to build and update composite maps containing semantic labels. It uses incremental SLAM algorithms and dynamic mutation detection to solve the problem of global mapping delay in dynamic environments. It builds an initial composite map based on the first scan, divides the area, and detects environmental mutations in real time through variable resolution octrees and spatiotemporal attention matching. It uses graph optimization to update the local topological relationship of the mutation area and outputs a dynamic composite map. This ensures that the map is updated in real time as the environment changes, meets the mapping requirements in dynamic scenarios, and improves the timeliness and accuracy of the map.

[0151] The path optimization unit is used to generate and optimize cleaning paths. It generates hierarchical paths based on the semantic labels of the initial composite map. Through the RRT algorithm and multi-objective optimization, it solves the problem of path planning imbalance in unstructured environments. It dynamically adjusts path parameters according to bed type and material, and updates path parameters in real time in response to environmental changes. It uses the pollution index to predict diffusion trends and adjust cleaning strategies, improving path efficiency and coverage accuracy in complex areas, achieving efficient cleaning in different scenarios.

[0152] The learning unit is used to achieve human-machine collaboration and strategy self-optimization. Through user-defined area processing and reinforcement learning algorithms, it solves the problems of personalized cleaning needs and strategy iteration. It supports APP hand-drawn area planning and voice interaction, dynamically adjusts path priority based on historical cleaning data and user habits, and iteratively optimizes cleaning strategies through the DQN algorithm, reducing manual intervention and improving scene adaptation efficiency. The cleaning strategy can continuously evolve over time to meet diverse cleaning needs.

[0153] In summary, the present invention activates the robot's self-detection sensors, collects data such as point clouds and depth images, and uses incremental SLAM to construct an initial composite map containing multiple regions. It then divides subregions in real time to match historical maps, marks sudden changes in regions, and updates the composite map. Sensor data is integrated to construct a safety buffer zone, and obstacle trajectories are predicted using LSTM to dynamically adjust the buffer distance. Obstacle avoidance is achieved through differential drive of the tracks. The base path is called and updated in real time based on sudden changes using the RRT algorithm and gradient descent. Parameters are adjusted based on material semantics and contamination index. After cleaning is complete, the path strategy is iteratively optimized through Q learning based on historical data and user operations to adapt to user habits.

[0154] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiment. All technical solutions based on the concept of the present invention are within the scope of protection of the present invention. It should be noted that for those skilled in the art, various improvements and modifications that do not depart from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A cleaning path planning method, characterized in that: include: Collect multi-source data of the area to be cleaned, build an initial composite map, set a perception marking method, divide the point cloud data into sub-areas, mark the mutation areas after matching with historical map features, and use graph optimization to locally update the composite map; A safety buffer zone is constructed at the edge of the area to be cleaned, a buffer adjustment method is set, a trajectory prediction is performed on the mutation area, and the safety buffer distance is dynamically adjusted according to the prediction result, and the motion trajectory is adjusted in real time through the crawler differential drive control; Call the basic cleaning path, generate a hierarchical path, set the planning adjustment method, and during the cleaning process, update the hierarchical path in real time according to the sudden change area, calculate the pollution index of the area to be cleaned, build a time series prediction model, predict the pollution diffusion trend, and make dynamic adjustments; After the area to be cleaned is cleaned, decision iteration is performed to adapt to user habits.

2. The cleaning path planning method according to claim 1, wherein: The specific steps to construct the initial composite map include: Before cleaning the area to be cleaned, trigger a function self-check; The point cloud data and depth image are acquired through the LiDAR and depth camera, and the surface pressure distribution matrix is ​​generated by combining the pressure sensor, and the material thickness distribution matrix is ​​acquired by the capacitive thickness sensor; Perform statistical filtering, voxel downsampling and bilateral filtering on the collected data, and complete spatiotemporal alignment through the external parameter matrix to generate a fused point cloud; The RANSAC algorithm is used to fit the plane area, filter out the plane point set in the fused point cloud, and generate the plane area; The Canny edge detection algorithm is used to extract the contour edge, calculate the point cloud normal vector, and filter out the edge area point set in combination with the gravity direction to generate the edge area; Eliminating plane point sets and edge area point sets from the fused point cloud, and dividing the undulating area using a region growing algorithm; Based on octree partitioning, potential hidden areas are screened out, and the boundaries of hidden areas are inferred by combining Bayesian filtering to generate hidden areas; The initial composite map is constructed through the incremental SLAM algorithm, which includes semantic labels with planes, edges, undulations, and hidden areas, and keyframe-associated geometric features and pose information.

3. The cleaning path planning method according to claim 2, wherein: The perceptual marking method comprises: Acquire real-time point cloud data, reduce density through voxel grid filtering, and automatically adjust the octree resolution according to the point cloud density to generate variable resolution sub-region grids; For each sub-region, use the pruned and compressed PointNet++ network to extract the composite vector and update the sub-region grid; Calculating the overlap rate by using the octree intersection volume ratio, and extracting the historical key frames overlapping with the current sub-region in the initial composite map; Using the Transformer attention network, the current sub-region features are mapped into a query vector, and the historical keyframe features are mapped into a key vector and a value vector to calculate the attention weight; Combining bidirectional nearest neighbor matching and weighted Euclidean distance, valid point pairs are screened and the displacement of the sub-region is calculated using the least squares method; Use a sliding window to store historical mutation displacement data, and dynamically adjust the mutation detection threshold through the incremental variance formula; Subregions whose displacement exceeds the mutation detection threshold are screened out as candidate mutation regions, and connected domain analysis is performed to generate a continuous set of candidate mutation regions.

4. The cleaning path planning method according to claim 3, wherein: The perceptual marking method further includes: For each candidate mutation region, query the historical detection results of the previous three frames. If it is a single-frame mutation, calculate the similarity with the historical mutation pattern and eliminate candidate mutation regions with similarity less than the similarity threshold; Perform spatial verification of edge areas and undulating areas, calculate the convex hull volume change rate, and eliminate candidate mutation areas whose absolute value of the change rate is lower than the change threshold; The grayscale change rate of the depth image is used for screening, and the confidence is enhanced. The Euclidean distance of the spectral features is calculated to screen and generate a set of effective mutation regions. For the effective mutation area set, calculate the point cloud mutation ratio, the grayscale change rate of the depth image, and the material spectrum mutation degree, and fuse them according to the weight to obtain the multimodal mutation level; Retrieve historical optimization cases whose similarity with the current mutation pattern exceeds a priori threshold from the memory pool as priors; Construct a local pose graph containing point cloud reprojection constraints and history memory constraints, and use the g2o optimizer to solve it to update the local topological relationship of the mutation area; The semantic confidence in the initial composite map is updated according to the mutation frequency and a lifecycle label is assigned to generate a fused dynamic composite map.

5. The cleaning path planning method according to claim 4, characterized in that: The specific steps to build a safety buffer zone include: Collect multi-source data, align them in time and space through the sensor extrinsic matrix, and generate edge point sets with confidence by weighted fusion; Based on the edge point set, the warning layer, deceleration layer and braking layer are generated by expanding outward along the normal vector direction, which trigger visual warning, speed attenuation function calculation and emergency braking respectively. Extracting a terrain slope from the dynamic composite map, and increasing the outward expansion distance of each layer once the terrain slope exceeds a slope threshold; An infrared spectrometer collects spectral data of ground materials and calculates humidity values. Once the humidity value exceeds the humidity threshold, the braking distance is increased and the anti-skid driving mode is activated. Compare the edge load data collected in real time with the historically stored edge load data. When the deviation exceeds the deviation threshold, increase the safety buffer zone expansion distance. At the same time, if the edge point cloud displacement is detected to exceed the displacement threshold, the safety buffer zone moves in real time with the edge.

6. The cleaning path planning method according to claim 5, characterized in that: The buffer adjustment method comprises: The QuickHull algorithm is used to generate a three-dimensional convex hull and an axis-aligned bounding box, and the geometric eigenvector of the convex hull is calculated to characterize the regularity of the obstacle. The historical position of the obstacle is extracted, and after removing the static frames, the trajectory is smoothed through Kalman filtering to generate a motion feature vector containing velocity, acceleration, and steering angle; Build a motion trend model, output the predicted trajectory for the future period, and generate an error ellipse by combining Monte Carlo Dropout; Based on the obstacle movement speed and prediction uncertainty, the safety buffer distance is calculated by superimposing speed compensation and uncertainty compensation on the basic buffer distance. , and determine the advance adjustment time; Calculate the relative distance between the current position and the predicted obstacle position and azimuth; like , maintain the base speed ;like , start steering control; Real-time calculation of slip rate , and compensate for the speed to generate a PWM signal to control the motor.

7. The cleaning path planning method according to claim 6, characterized in that: The planning adjustment method includes: A three-dimensional search space is generated with the mutation area as the center, and a grid map to be passed is constructed by octree partitioning; Obtain the predicted obstacle position and use a dynamic bias sampling strategy to guide the sampling point to offset the predicted obstacle position; Using the RRT algorithm, a set of candidate paths is generated, and the smoothness of each candidate path is calculated, and the one with the smallest smoothness is retained. Paths; A multi-objective function is constructed based on path distance, energy consumption and terrain resistance, and the path parameters are optimized using the gradient descent method to generate the optimal path.

8. The cleaning path planning method according to claim 7, characterized in that: The planning adjustment method further includes: The infrared spectrometer collects data and identifies the material type, and combines it with pressure data to verify consistency. If the consistency verification fails, the material type is re-identified, otherwise the cleaning parameters are dynamically adjusted; Calculate pollution indices, predict pollution trends and trigger compensation strategies; Convert the hand-drawn key area coordinates to the map coordinate system, extract the geometric features, generate the grid path, and output the path priority index; A state space and action space were established, and the DQN algorithm was used to iteratively optimize path parameters based on historical cleaning data and user logs to adjust the visit frequency of high-frequency residual areas; among them, high-frequency residual areas are areas with fixed spatial positions and where contamination remains after completing a preset number of cleaning times.

9. A cleaning device for implementing the cleaning path planning method according to any one of claims 1 to 8, characterized in that: include: Perception module, planning module and executive control module; The perception module is used to collect multi-dimensional data through multi-sensor collaboration, eliminate noise through pre-processing and complete spatiotemporal alignment; The planning module is used to construct an initial composite map based on the initial scan data and dynamically update it, generate a hierarchical path through the RRT algorithm and multi-objective optimization, dynamically adjust path parameters based on material semantic recognition and pollution index, and use a reinforcement learning algorithm to analyze user operation data to iteratively optimize the path strategy; The execution control module is used to fuse multi-sensor edge data to generate an adaptive safety buffer zone, dynamically adjust the safety buffer distance, predict the obstacle movement trend through LSTM, and correct the movement trajectory in combination with track differential drive.

10. The cleaning device according to claim 9, characterized in that: The planning module includes a mapping unit, a path optimization unit and a learning unit; The mapping unit constructs and updates a composite map containing semantic labels through an incremental SLAM algorithm and dynamic mutation detection, uses a variable-resolution octree and spatiotemporal attention matching to detect environmental topology changes in real time, and updates the local topological relationship of the mutation area through graph optimization; The path optimization unit generates a hierarchical path based on the semantic labels of the initial composite map, adopts the RRT algorithm and multi-objective optimization, adjusts the path planning according to dynamic parameters, updates the path parameters in real time in response to environmental changes, and predicts the diffusion trend based on the pollution index; The learning unit uses user-defined area planning and reinforcement learning algorithms to dynamically adjust path priorities based on historical cleaning data and user habits, and iteratively optimize the cleaning strategy.

Images