Cleaning robot autonomous path planning control method based on deep learning

Abstract

The invention discloses a cleaning robot autonomous path planning control method based on deep learning, and the method comprises the following steps: obtaining the pose and obstacle information of a cleaning robot, and generating a data packet containing obstacle distribution and blocked coordinates when detecting that a path is blocked; activating a reconnaissance granulocyte antibody family, generating a plurality of micro-paths constrained by curvature, evaluating the accessibility, and outputting safety direction information; encoding the path data and the safety direction to form an antigen vector, and generating a candidate antibody set; executing cloning, variation, concentration inhibition and negative selection according to the matching degree, and screening an optimal antibody path as a repair segment for output; and performing phagocytic recovery on the invalid antibody and updating variation prior, and performing closed-loop splicing on the repair path and the original path to generate an executable track. According to the invention, a reconnaissance granulocyte antibody family and a phagocytosis-recovery mechanism are introduced into an artificial immune algorithm, so that the path continuity and the operation stability in a complex environment are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent robot control, and particularly relates to a cleaning robot autonomous path planning control method based on deep learning. BACKGROUND

[0002] With the rapid development of service robots and intelligent household appliances, cleaning robots are increasingly widely used in home and commercial environments. Cleaning robots usually realize autonomous movement and area coverage through environment perception, map construction and path planning modules, and the stability and environmental adaptability of the path planning algorithm directly affect the cleaning efficiency and operation safety.

[0003] In the prior art, the path planning of the cleaning robot mostly adopts a combination of global planning based on a grid map and real-time correction based on local obstacle avoidance. In terms of global planning, common algorithms include A algorithm, D algorithm, DWA dynamic window algorithm and its improved type. These methods rely on a relatively complete environment map and static obstacle information. When there are dynamic obstacles, soft obstacles or unstructured scenes in the working environment of the robot, the planned path is easy to fail, and global planning needs to be re-executed, resulting in operation interruption or repeated coverage.

[0004] In terms of local path correction, in recent years, various strategies based on deep learning, reinforcement learning and swarm intelligence have been introduced. These methods improve the path adaptive ability to some extent, but there are still two outstanding problems: first, after the local path is blocked, it depends on the whole map re-planning or random sampling expansion, which has large calculation amount and lagging response, and cannot quickly recover the movement; second, there is a lack of continuous learning mechanism, and the algorithm repeatedly occurs the same blocking phenomenon in different scenes, and it is difficult to form long-term memory and self-repairing ability.

[0005] Therefore, how to provide a cleaning robot autonomous path planning control method based on deep learning is a problem that those skilled in the art need to solve. SUMMARY

[0006] One purpose of the present application is to propose a deep learning-based autonomous path planning control method for cleaning robots, which introduces a scout granulocyte antibody family and a phagocytosis-recovery mechanism into an artificial immune algorithm, constructs a path self-repair control system for cleaning robots with rapid detection and self-learning capabilities, and realizes local rapid repair and long-term immune learning when the path is blocked. This method can quickly determine the feasible direction by generating a tentative micro-path through the scout granulocyte antibody family and evaluating the passability without re-planning the global path, and dynamically optimizes the cloning and mutation control parameters and forms new mutation priors by updating the path parameters of the failed antibodies through the phagocytosis-recovery mechanism, so that the algorithm has continuous evolution and self-adaptation capabilities. By performing direction and curvature continuity checking and generating a smooth transition section during path splicing, the present application ensures seamless connection between the repaired path and the original path, effectively improving the path continuity, planning stability and running efficiency of the cleaning robot in a complex dynamic environment.

[0007] According to the deep learning-based autonomous path planning control method for cleaning robots, the method comprises the following steps:

[0008] Obtain the pose and surrounding obstacle information of the cleaning robot through the environment sensor, calculate the path pass rate, and generate a path blocking data packet containing obstacle distribution and blocked area coordinates when detecting that the front path is blocked;

[0009] Activate the scout granulocyte antibody family according to the path blocking data packet, generate a plurality of micro-path sets with preset micro-step length and curvature constraints according to the current position and blocked direction, evaluate the passability of each micro-path, and output a set of safe direction information;

[0010] Combine and encode the path blocking data packet and the set of safe direction information to form an antigen vector, and generate an initial candidate antibody set in combination with the mutation prior parameters stored in the immune memory library, each antibody corresponding to a feasible repair path segment;

[0011] Calculate the matching degree of the candidate antibody and the antigen vector, perform cloning and mutation operations according to the matching degree, obtain an updated antibody set, and form an affinity ranking table;

[0012] Perform concentration inhibition and negative selection according to the affinity ranking table to obtain a screened antibody set, select the antibody with the highest matching degree, and output the corresponding path segment as the repair path;

[0013] Perform phagocytosis-recovery operation on the antibodies excluded in the screening, count the path parameter distribution, dynamically update the control parameters of cloning and mutation based on the statistical results, form new mutation priors, and store the new mutation priors in the immune memory library;

[0014] The repair path output is then joined with the original running path in a closed loop to generate an executable trajectory, which the cleaning robot uses to resume its movement.

[0015] Furthermore, the step of generating the path-blocked data packet includes:

[0016] The system uses a visual sensor to acquire environmental images, a lidar sensor to collect distance point cloud data, and an inertial measurement unit to collect attitude and velocity information, and performs time synchronization and spatial registration on the information.

[0017] A local grid map is constructed based on the point cloud data, and a detection window is set in the grid map with the robot's current position as the center;

[0018] Based on environmental images and point cloud data, deep feature extraction networks are used to calculate path feasibility.

[0019] When the path availability rate is lower than a preset threshold and the downward trend within a continuous detection period exceeds the rate of change threshold, a path obstruction data packet containing an obstacle distribution matrix, obstruction area coordinates, obstacle type identifier, and timestamp information is generated.

[0020] Furthermore, the step of activating the reconnaissance granulocyte antibody family includes:

[0021] The direction of obstruction is determined by using the coordinates of the obstructed area and the obstacle distribution matrix in the data packets of the obstructed path;

[0022] Forward sectors are set on both sides of the obstructed direction, and direction sampling is performed in the forward sectors at preset angle intervals to form a candidate orientation set.

[0023] At each candidate orientation, several exploratory micropaths are generated based on the cleaning robot's current position, attitude information, and the turning radius, minimum curvature, and speed range of the driving system.

[0024] For each micropath, a accessibility index is calculated, which includes the proportion of the path segment that is not occupied by obstacles, the minimum distance to the obstacle boundary, and the curvature continuity.

[0025] The number of feasible paths is counted under all candidate orientations, and the repeated or similar directions between adjacent orientations are clustered to obtain several independent safe directions;

[0026] The safety directions are sorted according to the degree to which the passability determination conditions are met, forming a safety direction information set.

[0027] Furthermore, the antigen vector formation and candidate antibody set generation steps include:

[0028] The obstacle distribution matrix, obstructed area coordinates, obstacle type identifiers and timestamps in the blocked data packets are unified and aligned in terms of coordinates and time. The image coordinates, lidar coordinates and map coordinates are converted to the same map coordinate system, and a field integrity verification record is established.

[0029] A local detection window is extracted from the obstacle distribution matrix with the coordinates of the obstructed area as the range. Connectivity segmentation and boundary extraction are performed on the local detection window to obtain the obstacle connectivity number, boundary orientation identifier and obstacle density description.

[0030] Based on the direction entries in the safety direction information set, establish the correspondence between directions and obstructed areas, and generate a direction-area correspondence table containing direction index, incident azimuth, permitted start and end areas, and relative positional relationship with obstacle boundaries;

[0031] The obstacle density description, boundary orientation identifier, connected component number, obstructed area boundary coordinates, direction-area correspondence table, obstacle type identifier, and robot current position and attitude information are uniformly encoded according to the preset field order to form an antigen vector containing environmental geometric information, obstructed orientation information, and passable direction information.

[0032] The system reads the mutation prior parameters stored in the previous round or multiple rounds of immune repair from the immune memory bank, and matches them according to scene category, obstacle type and direction index to determine the path length range, curvature range, turning change range and generation frequency used for this generation.

[0033] Starting with each direction entry in the safety direction information set, and combining the allowed start and end regions recorded in the variation prior parameters and direction-region correspondence table, a preset path primitive type is selected and an initial repair path segment prototype is generated according to the primitive combination rules. The path primitive types include arc segments, edge segments, and spline segments.

[0034] Under the map coordinate system, each prototype of the repaired path segment is checked for segment-by-segment occupancy and boundary compatibility. Path segments with occupancy conflicts or cross boundaries are replaced or truncated at the segment level. Adjacent segments are supplemented to form a continuous path while maintaining the constraints of the start and end areas.

[0035] The repair path segments that pass the verification are structured and encoded, and their starting point coordinates, ending point coordinates, direction index, path primitive sequence used, inter-segment connection order, path length summary, curvature range summary and relative position description of the blocked area are recorded to generate corresponding candidate antibody entries.

[0036] All candidate antibody entries are grouped into an initial candidate antibody set according to their orientation index and generation order.

[0037] Furthermore, the cloning and mutation operations include:

[0038] For each antibody in the initial candidate antibody set, its path parameter record is read, and the start and end coordinates, curvature sequence, turning angle change, and minimum distance to the obstacle boundary of the antibody path are compared with the corresponding environmental geometric features and obstruction orientation features in the antigen vector.

[0039] The matching degree value is calculated based on the comparison results. The matching degree value is determined by the weighted comprehensive result obtained from the determination of path continuity, direction consistency and safety margin.

[0040] Candidate antibodies are sorted from highest to lowest matching degree, and the number of clones is determined based on the sorting results. Antibodies with high matching degree generate more clone instances.

[0041] For each clone instance, local parameters of a path segment are randomly selected as mutation locations. The path length, turning angle, or curvature parameters are perturbed according to the range and step size defined in the mutation prior parameters to generate mutant antibodies.

[0042] The mutated antibodies are tested for passability, and antibody instances that cause obstacle collisions or cross boundaries are removed. The mutated antibodies that pass the test are incorporated into the updated antibody set.

[0043] In the updated antibody set, the matching degree is recalculated and a matching degree ranking list is generated. The antibody index, matching degree value, clone source identifier and mutation parameter record are saved, and an affinity ranking table is generated.

[0044] Furthermore, the step of generating the repair path includes:

[0045] Read all antibody records in the affinity sorting table and calculate the similarity of each antibody with other antibodies in terms of path direction, start and end regions, and morphological parameters.

[0046] Based on the similarity determination results, antibody clusters are determined. Within each cluster, the antibody with the highest affinity is retained, and the remaining antibodies are inhibited or deleted, resulting in a concentration-inhibited antibody set.

[0047] Negative selection detection was performed on the inhibited antibody aggregate. The negative selection detection included:

[0048] For each antibody path, perform local accessibility verification to determine whether the path segment overlaps with the obstacle boundary, whether it exceeds the preset curvature or steering angle constraint, and whether there are path segments that do not match the size of the robot chassis.

[0049] Antibodies that fail validation after two or more consecutive tests are marked as invalid antibodies and removed from the current antibody set to form a set of effective antibodies after screening.

[0050] In the selected set of effective antibodies, the mean matching degree of each antibody is recalculated and sorted from high to low. The antibody with the highest ranking is selected as the optimal antibody, and the corresponding path segment is output as the repair path.

[0051] Furthermore, the swallowing-recycling operation includes:

[0052] Collect records of ineffective antibodies that were eliminated during concentration inhibition and negative selection, extract the path length, curvature change, turning angle change and start and end coordinate information of each ineffective antibody, and establish a set of parameters for ineffective antibodies;

[0053] The set of failed antibody parameters is classified and statistically analyzed. The mean path length, curvature distribution range, and turning angle variation range are calculated according to the obstacle type, obstruction location, and safe direction index, and the corresponding occurrence frequency is recorded.

[0054] Based on the statistical results, a parameter distribution table is constructed. The parameter range of similar ineffective antibodies is used as the basis for correction. The path length, curvature step size and turning perturbation range in the existing clonal mutation control parameters are dynamically adjusted.

[0055] The adjusted clonal mutation control parameters are summarized into a new set of mutation prior parameters, and corresponding scene index, obstacle category index and time index are generated.

[0056] The new set of mutation prior parameters is stored in the immune memory bank, and the parameter entries for the corresponding scenarios in the memory bank are updated.

[0057] For old parameter entries already stored in the immune memory bank, perform weight decay and overwrite updates according to the time index.

[0058] Furthermore, the steps for generating and applying the executable trajectory include:

[0059] Receive the repair path segment corresponding to the optimal antibody in the repair path output, and extract the start coordinates, end coordinates, path length and curvature information;

[0060] Locate the path node adjacent to the starting point of the repair path segment in the original running path, and use this node as the splicing starting point to establish the connection relationship between the original path segment and the repair path segment.

[0061] The path direction, curvature and pose at the connection point are continuously checked. When the direction angle or curvature difference exceeds the preset threshold, a transition segment is inserted at the splicing point.

[0062] The repaired path segment and the original path segment are resampled in the same coordinate system to generate a sequence of trajectory points containing continuous coordinate points, direction angles, speed commands and timestamps;

[0063] The kinematic executability of the trajectory point sequence is verified, including the rate of change of steering angle, wheel speed difference and acceleration range. Once the verification is passed, the trajectory point sequence is packaged into an executable trajectory file.

[0064] The executable trajectory file is input into the underlying motion control module of the cleaning robot, and the execution instructions of the original path segment are replaced with the updated trajectory.

[0065] The beneficial effects of this invention are:

[0066] This invention, upon detecting path obstruction, activates a reconnaissance granulocyte antibody family to generate multiple tentative micropaths and evaluates their feasibility. This allows for rapid determination of feasible directions within a local area without re-executing global planning, significantly shortening path recovery time. By incorporating a phagocytosis-recycling mechanism during the negative selection phase, the parameters of rejected antibody paths are statistically analyzed and summarized, and clonal mutation control parameters are dynamically updated to form new mutation priors. This enables the algorithm to progressively optimize path generation strategies in different environments, demonstrating long-term self-learning capabilities.

[0067] This invention performs continuity checks on direction, curvature, and pose during the path repair splicing stage, and generates transition segments when necessary to achieve a smooth connection between the repaired path and the original path, ensuring the robot's motion continuity in narrow or dynamic environments. By encoding the obstacle distribution matrix, obstacle type identifier, and safety direction information in the path obstruction data packet into an antigen vector, the algorithm can perform unified processing in various scenarios and has strong environmental adaptability.

[0068] This invention stores prior parameters of mutations under different scenarios in an immune memory bank, enabling the accumulation and retrieval of experiential knowledge. This allows the system to quickly generate better path solutions in subsequent operations, improving overall planning efficiency and stability. Attached Figure Description

[0069] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0070] Figure 1 This is a flowchart of a deep learning-based autonomous path planning and control method for cleaning robots proposed in this invention.

[0071] Figure 2 This is a flowchart illustrating the activation of the reconnaissance granulocyte antibody family in a deep learning-based autonomous path planning and control method for cleaning robots proposed in this invention.

[0072] Figure 3 This is an immune cycle flowchart of an autonomous path planning and control method for cleaning robots based on deep learning proposed in this invention. Detailed Implementation

[0073] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0074] refer to Figure 1 - Figure 3 A deep learning-based autonomous path planning and control method for cleaning robots includes the following steps:

[0075] The robot obtains its pose and information about surrounding obstacles through environmental sensors, calculates the path availability, and generates a path obstruction data packet containing the distribution of obstacles and the coordinates of the obstructed area when it detects that the path ahead is blocked.

[0076] The reconnaissance granulocyte antibody family is activated based on the blocked data packet. The reconnaissance granulocyte antibody family generates a set of micropaths with preset microstep size and curvature constraints based on the current position and the blocked direction. The accessibility of each micropath is evaluated and a safe direction information set is output.

[0077] The blocked data packet and the security direction information set are combined and encoded to form an antigen vector, and combined with the mutation prior parameters stored in the immune memory bank to generate an initial set of candidate antibodies, with each antibody corresponding to a feasible repair path segment;

[0078] Calculate the matching degree between candidate antibodies and antigen vectors, perform cloning and mutation operations based on the matching degree, obtain an updated antibody set, and form an affinity ranking table;

[0079] Based on the affinity ranking table, concentration inhibition and negative selection are performed to obtain the set of screened antibodies. The antibody with the highest matching degree is selected, and the corresponding path segment is output as the repair path.

[0080] The antibodies that were rejected during the screening were subjected to phagocytosis-recovery operation, their path parameter distribution was statistically analyzed, and the control parameters of clonal mutations were dynamically updated based on the statistical results to form new mutation priors. The new mutation priors were then stored in the immune memory bank.

[0081] The repair path output is then joined with the original running path in a closed loop to generate an executable trajectory, which the cleaning robot uses to resume its movement.

[0082] In this embodiment, the step of generating the path-blocked data packet includes:

[0083] The system uses a visual sensor to acquire environmental images, a lidar sensor to collect distance point cloud data, and an inertial measurement unit to collect attitude and velocity information, and performs time synchronization and spatial registration on the information.

[0084] A local grid map is constructed based on the point cloud data, and a detection window is set in the grid map with the robot's current position as the center;

[0085] Occupancy probability is estimated for each grid cell within the detection window using ray projection-based occupancy probability update, resulting in an occupancy probability matrix;

[0086] The cells with probabilities not lower than the occupancy threshold in the occupancy probability matrix are mapped to a binary occupancy map, and the row and column indices of the binary occupancy map in the detection window coordinate system are used to form an obstacle distribution matrix;

[0087] In the binary occupancy map, select the occupancy connected region that intersects with the forward sector of the current planned path, and calculate the map coordinates of the vertices of its circumscribed rectangle or polygon as the coordinates of the obstructed area.

[0088] Perform target detection on the environmental image and output obstacle type identifiers corresponding to the obstacle distribution matrix;

[0089] Based on environmental images and point cloud data, the network uses deep feature extraction to calculate path passability. The network adopts a fixed dual-branch encoding and fusion structure. The image branch uses a convolutional neural network to extract color, texture and boundary semantic features, while the point cloud branch uses a sparse 3D convolutional network to extract spatial shape and obstacle geometric distribution features.

[0090] The two features are fused through cross-modal attention and then upsampled and reconstructed in a multi-scale decoder to output a accessibility probability map with the same size as the local grid map. Each grid corresponds to a confidence value of environmental accessibility.

[0091] The average passability probability of all grid cells within the detection window is used as the path passability rate of the current path segment.

[0092] When the path availability rate is lower than a preset threshold and the downward trend within a continuous detection period exceeds the rate of change threshold, a path obstruction data packet containing an obstacle distribution matrix, obstruction area coordinates, obstacle type identifier, and timestamp information is generated.

[0093] In this embodiment, the step of activating the reconnaissance granulocyte antibody family includes:

[0094] The direction of obstruction is determined by using the coordinates of the obstructed area and the obstacle distribution matrix in the data packets of the obstructed path;

[0095] Forward sectors are set on both sides of the obstructed direction, and direction sampling is performed in the forward sectors at preset angle intervals to form a candidate orientation set.

[0096] At each candidate orientation, several exploratory micropaths are generated based on the cleaning robot's current position, attitude information, and the turning radius, minimum curvature, and speed range of the driving system.

[0097] The micropath is a path segment with defined length and curvature conditions within a local raster map, which is divided into multiple discrete points according to the sampling step size and stored in sequence form;

[0098] For the micropath, occupancy detection is performed on the local grid map to determine whether each sampling point in the path segment overlaps with an obstacle unit or is within a safe distance range;

[0099] For each micropath, a accessibility index is calculated, which includes the proportion of the path segment that is not occupied by obstacles, the minimum distance to the obstacle boundary, and the curvature continuity.

[0100] Micropaths that meet the feasibility criteria are marked as feasible paths, and their corresponding direction information is recorded.

[0101] The number of feasible paths is counted under all candidate orientations, and the repeated or similar directions between adjacent orientations are clustered to obtain several independent safe directions;

[0102] The safety directions are sorted according to the degree of satisfaction of the passability determination conditions to form a safety direction information set, and this safety direction information set is provided as input to the antigen encoding and candidate antibody generation steps.

[0103] In this embodiment, the steps of antigen vector formation and candidate antibody set generation include:

[0104] The obstacle distribution matrix, obstructed area coordinates, obstacle type identifiers and timestamps in the blocked data packets are unified and aligned in terms of coordinates and time. The image coordinates, lidar coordinates and map coordinates are converted to the same map coordinate system, and a field integrity verification record is established.

[0105] A local detection window is extracted from the obstacle distribution matrix with the coordinates of the obstructed area as the range. Connectivity segmentation and boundary extraction are performed on the local detection window to obtain the obstacle connectivity number, boundary orientation identifier and obstacle density description.

[0106] Based on the direction entries in the safety direction information set, establish the correspondence between directions and obstructed areas, and generate a direction-area correspondence table containing direction index, incident azimuth, permitted start and end areas, and relative positional relationship with obstacle boundaries;

[0107] The obstacle density description, boundary orientation identifier, connected component number, obstructed area boundary coordinates, direction-area correspondence table, obstacle type identifier, and robot current position and attitude information are uniformly encoded according to the preset field order to form an antigen vector containing environmental geometric information, obstructed orientation information, and passable direction information.

[0108] The system reads the mutation prior parameters stored in the previous round or multiple rounds of immune repair from the immune memory bank, and matches them according to scene category, obstacle type and direction index to determine the path length range, curvature range, turning change range and generation frequency used for this generation.

[0109] Starting with each direction entry in the safety direction information set, and combining the allowed start and end regions recorded in the variation prior parameters and direction-region correspondence table, a preset path primitive type is selected and an initial repair path segment prototype is generated according to the primitive combination rules. The path primitive types include arc segments, edge segments, and spline segments.

[0110] Under the map coordinate system, each prototype of the repaired path segment is checked for segment-by-segment occupancy and boundary compatibility. Path segments with occupancy conflicts or cross boundaries are replaced or truncated at the segment level. Adjacent segments are supplemented to form a continuous path while maintaining the constraints of the start and end areas.

[0111] The repair path segments that pass the verification are structured and encoded, and their starting point coordinates, ending point coordinates, direction index, path primitive sequence used, inter-segment connection order, path length summary, curvature range summary and relative position description of the blocked area are recorded to generate corresponding candidate antibody entries.

[0112] All candidate antibody entries are aggregated into an initial candidate antibody set according to their orientation index and generation order, and this set is submitted to the cloning and mutation steps as input data.

[0113] In this embodiment, the cloning and mutation operations include:

[0114] For each antibody in the initial candidate antibody set, its path parameter record is read, and the start and end coordinates, curvature sequence, turning angle change, and minimum distance to the obstacle boundary of the antibody path are compared with the corresponding environmental geometric features and obstruction orientation features in the antigen vector.

[0115] The matching degree value is calculated based on the comparison results. The matching degree value is determined by the weighted comprehensive result obtained from the determination of path continuity, direction consistency and safety margin.

[0116] ;

[0117] in, Let be the matching degree value of the i-th candidate antibody. This represents the directional difference between the antibody pathway and the direction of antigen blockage. This is the difference between the average curvature of the antibody pathway and the curvature of the antigen-blocked boundary. This represents the minimum safe distance from all sampling points within the antibody pathway to the barrier boundary. This is a path continuity index with a value range of [0,1]. w1, w2, w3, and w4 are weighting coefficients that satisfy w1 + w2 + w3 + w4 = 1. The maximum curvature allowed under the current environment. This represents the maximum passable distance within the current detection window.

[0118] ;

[0119] in This represents the number of sampling points along the path. and The first The orientation angle and curvature of each sampling point.

[0120] Candidate antibodies are sorted from highest to lowest matching degree, and the number of clones is determined based on the sorting results. Antibodies with high matching degree generate more clone instances.

[0121] For each clone instance, local parameters of a path segment are randomly selected as mutation locations. The path length, turning angle, or curvature parameters are perturbed according to the range and step size defined in the mutation prior parameters to generate mutant antibodies.

[0122] The mutated antibodies are tested for passability, and antibody instances that cause obstacle collisions or cross boundaries are removed. The mutated antibodies that pass the test are incorporated into the updated antibody set.

[0123] The matching degree is recalculated in the updated antibody set and a matching degree ranking list is generated. The antibody index, matching degree value, clone source identifier and mutation parameter record are saved. An affinity ranking table is generated and provided to the concentration inhibition and negative selection steps.

[0124] In this embodiment, the step of generating the repair path includes:

[0125] Read all antibody records in the affinity sorting table and calculate the similarity of each antibody with other antibodies in terms of path direction, start and end regions, and morphological parameters.

[0126] ;

[0127] in, The overall similarity between the two antibody pathways is given. For path direction similarity, Similarity between the start and end regions. Morphological similarity describes the similarity of changes in path curvature. For the weighting coefficients, satisfying .

[0128] ;

[0129] in, and The average direction angle of the two paths.

[0130] ;

[0131] in, and Paths The coordinates of the start and end points, This represents the maximum passable distance within the current detection window.

[0132] ;

[0133] in, and The first The curvature of each sampling point This is the maximum number of sampling points set.

[0134] Based on the similarity determination results, antibody clusters are determined. Within each cluster, the antibody with the highest affinity is retained, and the remaining antibodies are inhibited or deleted, resulting in a concentration-inhibited antibody set.

[0135] Negative selection detection was performed on the inhibited antibody aggregate. The negative selection detection included:

[0136] For each antibody path, perform local accessibility verification to determine whether the path segment overlaps with the obstacle boundary, whether it exceeds the preset curvature or steering angle constraint, and whether there are path segments that do not match the size of the robot chassis.

[0137] Antibodies that fail validation after two or more consecutive tests are marked as invalid antibodies and removed from the current antibody set to form a set of effective antibodies after screening.

[0138] In the selected set of effective antibodies, the mean matching degree of each antibody is recalculated and sorted from high to low according to the mean. The antibody with the highest ranking is selected as the optimal antibody, and its corresponding path segment is output as the repair path and passed to the path splicing step.

[0139] In this embodiment, the swallowing-recycling operation includes:

[0140] Collect records of ineffective antibodies that were eliminated during concentration inhibition and negative selection, extract the path length, curvature change, turning angle change and start and end coordinate information of each ineffective antibody, and establish a set of parameters for ineffective antibodies;

[0141] The set of failed antibody parameters is classified and statistically analyzed. The mean path length, curvature distribution range, and turning angle variation range are calculated according to the obstacle type, obstruction location, and safe direction index, and the corresponding occurrence frequency is recorded.

[0142] Based on the statistical results, a parameter distribution table is constructed. The parameter range of similar ineffective antibodies is used as the basis for correction. The path length, curvature step size and turning perturbation range in the existing clonal mutation control parameters are dynamically adjusted.

[0143] The adjusted clonal mutation control parameters are summarized into a new set of mutation prior parameters, and corresponding scene index, obstacle category index and time index are generated.

[0144] The new set of mutation prior parameters is stored in the immune memory bank, and the parameter entries for the corresponding scenarios in the memory bank are updated to provide parameter references for the antibody generation steps when the subsequent path is blocked.

[0145] For old parameter entries already stored in the immune memory bank, weight decay and overwrite updates are performed according to the time index to maintain the timeliness and continuity of the memory bank parameters.

[0146] In this embodiment, the steps for generating and applying the executable trajectory include:

[0147] Receive the repair path segment corresponding to the optimal antibody in the repair path output, and extract its start coordinates, end coordinates, path length and curvature information;

[0148] Locate the path node adjacent to the starting point of the repair path segment in the original running path, and use this node as the splicing starting point to establish the connection relationship between the original path segment and the repair path segment.

[0149] The path direction, curvature, and pose at the connection point are continuously checked. When the direction angle or curvature difference exceeds a preset threshold, a transition segment is inserted at the splicing point. The transition segment generates a smooth connection curve based on the curvature and direction difference of the paths on both sides.

[0150] The repaired path segment and the original path segment are resampled in the same coordinate system to generate a sequence of trajectory points containing continuous coordinate points, direction angles, speed commands and timestamps;

[0151] The kinematic executability of the trajectory point sequence is verified, including the rate of change of steering angle, wheel speed difference and acceleration range. Once the verification is passed, the trajectory point sequence is packaged into an executable trajectory file.

[0152] The executable trajectory file is input into the underlying motion control module of the cleaning robot, and the execution instructions of the original path segment are replaced with the updated trajectory.

[0153] Example 1:

[0154] To verify the feasibility of this invention in practice, it was applied to a prototype wheeled cleaning robot, and an online self-repair experiment was conducted in a typical home setting under obstructed path conditions. The experimental setting was a two-bedroom home environment of approximately 65 square meters, including a living room, dining room, bedroom, hallway, and kitchen. The floor material was a mix of wood flooring and carpet, and the furniture was densely packed with both soft and dynamic obstacles. The robot was equipped with a LiDAR, an RGB-D camera, an inertial measurement unit, and touch sensors, and its edge computing unit was a 6TOPS embedded system.

[0155] To verify the actual performance of this invention, two path repair strategies were selected for comparison:

[0156] Comparison with Method A (Global Replanning): When the path is blocked, the A* algorithm for global path planning is re-executed;

[0157] Comparison Method B (Reinforcement Learning Local Obstacle Avoidance): Based on a deep reinforcement learning model, predict locally passable directions and generate path corrections.

[0158] Table 1 Experimental Configuration Table

[0159] Item Setting content Grid resolution 5cm Blocking threshold Passable rate below 30% Detection window 2.0m x 2.0m Change rate threshold Decline by more than 15% for 5 consecutive periods Direction sampling resolution 10° Micro-path step range 0.10–0.20m Minimum turning radius 0.45m Safety distance threshold 8cm Inhibition similarity threshold 0.7 Memory bank update period Every 20 repairs or every 10 minutes to take the mean once

[0160] During the experiment, the robot performed a complete coverage task, recording a total of 10 obstruction events (6 hard obstacles and 4 soft obstacles). For each event, the path repair time, whether global planning was triggered, the number of path splicing checks, and the task continuous completion rate were statistically analyzed.

[0161] Table 2 Experimental Results

[0162] Blocked event number Obstacle type Global replanning method repair delay (ms) Strengthening learning local obstacle avoidance method repair delay (ms) Invention repair delay (ms) 1 Table leg (hard) 682 395 185 2 Rise of floor mat (soft) 710 428 212 3 Displacement of dining chair (hard) 643 386 179 4 Slippers (soft) 709 410 225 5 Power cord (soft) 697 421 207 6 Sofa edge (hard) 675 392 190 7 Threshold (hard) 652 404 204 8 Carpet edge (soft) 724 417 219 9 Corridor corner (hard) 648 381 181 10 Dining table leg (hard) 672 389 188

[0163] As shown in Table 2, the average repair latency of the method of this invention is approximately 200 milliseconds across all obstructed events, which is about 70% lower than the traditional global replanning method and about 50% lower than the reinforcement learning local obstacle avoidance method. This indicates that the present invention can complete path repair within a single local planning cycle.

[0164] In the experiment, the global replanning method requires re-searching the entire map, and the computation time increases significantly with the complexity of the environment. Although reinforcement learning methods can achieve local obstacle avoidance, they are prone to multiple corrections or path oscillations in soft obstacle scenarios. In contrast, this invention rapidly generates micropaths in a local area by detecting granulocyte antibody families and selects the optimal repair path by combining antigen-antibody matching mechanism, thereby completing path recovery without changing the global topology.

[0165] Furthermore, in all 10 events, the present invention was able to complete the path splicing verification in one go without secondary transition adjustments, indicating that the continuity processing of direction and curvature constraints during the closed-loop splicing process is effective; while the reinforcement learning method showed discontinuity in direction in 3 soft obstacle scenarios, requiring secondary adjustments.

[0166] After multiple experiments, the immune memory bank of this invention gradually formed prior parameters for different types and locations of obstacles, making the repair process for subsequent similar obstacles more stable. Experimental records show that in the first 5 soft obstacle events, the average system repair latency was 216 milliseconds, while it decreased to 196 milliseconds in the subsequent 5 events, indicating that the phagocytosis-recycling and memory update mechanism can achieve online self-learning and parameter adaptation.

[0167] All three methods can achieve global coverage, but the global replanning method resulted in multiple repeated cleanings; the reinforcement learning method experienced two brief pauses in complex furniture areas; the method of this invention did not experience any interruption throughout the entire task, maintained complete path continuity, and the control commands did not exceed the chassis steering constraints, verifying the executability and stability of the method in complex environments.

[0168] In summary, the experiments show that the artificial immune pathway self-repair control method proposed in this invention can complete path recovery with low computational cost in obstructed scenarios, and continuously optimize parameter configuration through immune memory mechanism, thus possessing practical application feasibility and long-term operational stability.

[0169] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1.A deep learning-based autonomous path planning control method for a cleaning robot, characterized by, The method comprises the following steps: Obtaining the pose and surrounding obstacle information of the cleaning robot through the environmental sensor, calculating the path passability, and generating a path blocked data packet containing the obstacle distribution and blocked area coordinates when detecting that the front path is blocked; According to the path blocked data packet, activating the scout granulocyte antibody family, which generates a plurality of micro-path sets with preset micro-step length and curvature constraints according to the current position and blocked direction, evaluates the passability of each micro-path, and outputs a set of safe direction information; Combining the path blocked data packet and the set of safe direction information to form an antigen vector, and combining the stored variation prior parameters in the immune memory library to generate an initial candidate antibody set, each antibody corresponding to a feasible repair path segment; Calculating the matching degree of the candidate antibody and the antigen vector, performing cloning and variation operations according to the matching degree, obtaining an updated antibody set, and forming an affinity ranking table; According to the affinity ranking table, performing concentration inhibition and negative selection to obtain a screened antibody set, selecting the antibody with the highest matching degree, and outputting the corresponding path segment as the repair path; Performing phagocytosis-recovery operation on the antibodies excluded in the screening, counting the path parameter distribution, and dynamically updating the control parameters of cloning and variation based on the statistical results to form new variation priors, and storing the new variation priors in the immune memory library; Splicing the repair path output and the original running path to form an executable trajectory, and the cleaning robot resumes motion according to the executable trajectory. 2.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The generation step of the path blocked data packet comprises: Using a visual sensor to obtain an environmental image, using a laser radar sensor to collect distance point cloud data, using an inertial measurement unit to collect attitude and speed information, and time synchronizing and spatially registering the information; Based on the point cloud data, a local grid map is constructed, and a detection window is set in the grid map with the current position of the robot as the center; Based on the environmental image and the point cloud data, a depth feature extraction network is used to calculate the path passability; When the path passability is lower than a preset threshold and the downward trend in the continuous detection period exceeds a change rate threshold, a path blocked data packet containing the obstacle distribution matrix, the blocked area coordinates, the obstacle type identifier and the timestamp information is generated. 3.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The step of activating the scout granulocyte antibody family comprises: Using the blocked area coordinates and the obstacle distribution matrix in the path blocked data packet to determine the blocked direction; Setting a forward sector on both sides of the blocked direction, and sampling the direction in the forward sector according to a preset angle interval to form a set of candidate directions; At each candidate direction, a plurality of exploratory micro-paths are generated according to the current position and attitude information of the cleaning robot, as well as the turning radius, minimum curvature and speed range of the travel system; Calculating the passability index of each micro-path, which includes the proportion of the path segment not occupied by obstacles, the minimum distance from the obstacle boundary, and the curvature continuity; Statistically counting the number of feasible paths under all candidate directions, and clustering the repeated or similar directions between adjacent directions to obtain a plurality of independent safe directions; Sorting the safe directions according to the satisfaction degree of the passability judgment condition to form a set of safe direction information. 4.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The antigen vector formation and candidate antibody set generation step comprises: The obstacle distribution matrix, the obstacle region coordinates, the obstacle type identifier and the timestamp in the path blocked data packet are subjected to coordinate unification and time alignment, the image coordinates, the laser radar coordinates and the map coordinates are converted to the same map coordinate system, and a field integrity check record is established; A local detection window is intercepted in the obstacle distribution matrix with the obstacle region coordinates as the range, connected domain segmentation and boundary extraction are performed on the local detection window, and the obstacle connected domain number, the boundary trend identifier and the obstacle density description are obtained; According to the direction entries in the safe direction information set, a corresponding relationship between the directions and the obstacle regions is established, and a direction-region correspondence table containing the direction index, the incident position, the allowed start and end region and the relative position relationship with the obstacle boundary is generated; The obstacle density description, the boundary trend identifier, the connected domain number, the obstacle region boundary coordinates, the direction-region correspondence table, the obstacle type identifier and the current position and attitude information of the robot are uniformly encoded in a preset field order to form an antigen vector containing environmental geometric information, blocked direction information and passable direction information; The variation prior parameters stored in the last one or more rounds of immune repair processes are read from the immune memory library, and are matched according to the scene category, the obstacle type and the direction index to determine the path length range, the curvature range, the steering change range and the generation frequency used for this generation; Each direction entry in the safe direction information set is taken as a starting point, the allowed start and end region recorded in the direction-region correspondence table is combined with the variation prior parameters, a preset path primitive type is selected, and an initial repair path segment prototype is generated according to the primitive combination rule, the path primitive type including an arc segment, an edge following segment and a spline segment; The segment-level replacement or truncation is performed on the path segment that occurs occupancy conflict or boundary violation, and the adjacent segments are completed under the premise of maintaining the start and end region constraints to form a continuous path under the map coordinate system; The repair path segment that passes the check is structured and encoded to record the start point coordinates, the end point coordinates, the belonging direction index, the adopted path primitive sequence, the segment connection order, the path length abstract, the curvature range abstract and the relative position description of the blocked region, and a corresponding candidate antibody entry is generated; All the candidate antibody entries are collected as an initial candidate antibody set according to the direction index and the generation order. 5.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The cloning and variation operation comprises: The path parameter record of each antibody in the initial candidate antibody set is read, and the start and end coordinates of the antibody path, the curvature sequence, the steering angle change and the minimum distance from the obstacle boundary are compared with the corresponding environmental geometric features and blocked direction features in the antigen vector item by item; The matching degree value is calculated according to the comparison result, and the matching degree value is determined by the weighted comprehensive result of the path continuity, the direction consistency and the safety margin determination; The candidate antibodies are sorted from high to low according to the matching degree, and the number of clones is determined according to the sorting result, and the antibodies with high matching degree generate more clone instances; Randomly select local parameters of path segments of each clone instance as mutation positions, perform perturbation on path length, turning angle or curvature parameters according to the range and step length defined in the mutation prior parameters, and generate mutated antibodies; Perform passability test on the mutated antibodies, eliminate antibody instances that appear obstacle collision or out-of-bound, and integrate the tested mutated antibodies into the updated antibody set; Recalculate the matching degree in the updated antibody set and form a matching degree ranking list, save the antibody index, matching degree value, clone source identification and mutation parameter record, and generate an affinity ranking table. 6.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The generating step of the repair path comprises: Read all antibody records in the affinity ranking table, and count the similarity of each antibody with other antibodies in path direction, start and end area and shape parameters; Determine the antibody clustering set according to the similarity determination result, retain the antibody with the highest affinity in each clustering set, and inhibit or delete the remaining antibodies to obtain the antibody set after concentration inhibition; The negative selection detection on the inhibited antibody set comprises: Perform local passability verification on each antibody path to determine whether the path segment overlaps with the obstacle boundary, whether it exceeds the preset curvature or turning angle constraint, and whether there is a path segment that does not match the robot chassis size; Mark the antibody that fails to pass the verification for two or more times as a failed antibody, and eliminate it from the current antibody set to form a screened effective antibody set; Recalculate the matching degree average of each antibody in the screened effective antibody set, and sort them from high to low according to the average value, select the antibody with the highest ranking as the optimal antibody, and output the corresponding path segment as the repair path. 7.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The phagocytosis-recovery operation comprises: Collect the records of failed antibodies eliminated in the concentration inhibition and negative selection process, extract the path length, curvature change, turning angle change and start and end coordinate information of each failed antibody, and establish a failed antibody parameter set; Classify and count the failed antibody parameter set, calculate the path length average, curvature distribution range and turning angle change range according to the obstacle type, blocked direction and safe direction index, and record the corresponding frequency; Based on the statistical results, construct a parameter distribution table, use the parameter range of the same type of failed antibody as the correction basis, and dynamically adjust the path length, curvature step length and turning perturbation range in the existing clone mutation control parameters; Induce the adjusted clone mutation control parameters into a new set of mutation prior parameters, and generate corresponding scene index, obstacle category index and time index; Store the new set of mutation prior parameters in the immune memory library, and update the parameter entries corresponding to the scene in the memory library; Perform weight decay and coverage update on the old parameter entries stored in the immune memory library according to the time index. 8.The deep learning-based autonomous path planning control method of a cleaning robot according to claim 1, characterized in that, The generating and applying step of the executable trajectory comprises: Receive the repair path segment corresponding to the optimal antibody in the repair path output, and extract the start point coordinate, end point coordinate, path length and curvature information; Locate the path node adjacent to the start point of the repair path segment in the original running path, and use the adjacent path node as the splicing start point to establish the connection relationship between the original path segment and the repair path segment; The continuity of the path direction, curvature and pose at the joint is checked, and when the direction angle or the curvature difference exceeds a preset threshold, a transition section is inserted at the joint; The repaired path section and the original path section are uniformly resampled in a coordinate system to generate a trajectory point sequence containing continuous coordinate points, direction angles, speed instructions and time stamps; The kinematic executability of the trajectory point sequence is verified, and the verification content includes the steering angle change rate, the wheel speed difference and the acceleration range, and when the verification is passed, the trajectory point sequence is packaged as an executable trajectory file; The executable trajectory file is input into the underlying motion control module of the cleaning robot to replace the execution instructions of the original path section with the updated trajectory.

Images