Inspection garbage cleaning method of cleaning robot

Abstract

The invention discloses an inspection garbage cleaning method of a cleaning robot, and particularly relates to the technical field of robot cleaning, the method comprises the following steps: S1, garbage treatment: according to a garbage thermodynamic diagram of an area, through multiple times of scanning and observation of the same area, and through stability, distance and historical feedback filtering, identifying and confirming the existence of preferentially treated garbage and secondary garbage, determining the garbage treatment efficiency; grouping the identified garbage categories and distances; s2, determining a cleaning sequence and a decision: moving along a global path, identifying garbage by using a multi-modal sensor, and confirming an effective cleaning target; s3, the garbage type is judged, and if the garbage is common garbage, the step S3 is executed; and if the stains are stubborn viscous stains, skipping to S4. According to the method, common garbage and stubborn viscous stains are distinguished, it is ensured that cleaning strategies are matched with garbage types, the garbage recognition hit rate can be increased through recognition and scientific grouping, a clear basis is provided for follow-up cleaning sequence planning, and efficiency loss caused by undifferentiated cleaning is avoided.

Description

Technical Field

[0001] This application relates to the field of robot cleaning technology, and more specifically, to a method for inspecting and cleaning garbage using a cleaning robot. Background Technology

[0002] This invention belongs to the field of cleaning robots, specifically relating to an automated method for targeted waste cleaning that integrates sensing technology. With the rapid development of automation technology and artificial intelligence, robots are gradually replacing humans in performing various repetitive and manual labor tasks. Especially in the cleaning scenario, cleaning robots can utilize advanced navigation and sensing technologies to complete daily cleaning tasks without human intervention. These technologies reduce human resource costs and automate cleaning tasks. However, current technologies often cause robots to simply follow pre-set paths and perform general cleaning, resulting in low cleaning efficiency. A search revealed that publication number CN115933685A discloses a method, apparatus, equipment, and storage medium for inspection and cleaning. The method includes path planning for the area to be cleaned, determining the robot's initial cleaning path; collecting image data of the current sensing area as the robot travels along the initial cleaning path, and inputting the image data into a pre-trained dirt and debris detection model to determine the objects to be cleaned within the current sensing area; wherein the dirt and debris detection model is a multi-task detection network model; and determining the target object to be cleaned based on the robot's current cleaning task and the objects to be cleaned, controlling the robot to perform the current cleaning task on the target object. By using a multi-task detection model to determine the objects to be cleaned, the robot only needs to clean the objects within the area to be cleaned, rather than performing full-path cleaning, thus improving cleaning targeting and efficiency. The inventors discovered the following problems with the existing technology during the development of this application: Existing robotic inspection solutions clean areas as cleaning units. This approach has two main drawbacks: 1. Inflexible cleaning routes; 2. Even on clean areas, the robot may still perform unnecessary cleaning tasks, leading to low cleaning efficiency and wasted cleaning resources. Therefore, a method for cleaning robots to inspect and clean up garbage is proposed to address the above problems. Summary of the Invention

[0003] In order to overcome the above-mentioned defects of the prior art, this application provides a method for cleaning robots to inspect and clean up garbage, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, this application provides the following technical solution: a method for cleaning and sweeping garbage using a cleaning robot, comprising the following steps: S1. Waste Management: Based on the waste heat map of the area, the same area is scanned and observed multiple times, and stability filtering, distance adaptive filtering and historical feedback filtering are performed in sequence to identify and confirm the existence of priority waste and secondary waste, and the identified waste is grouped according to category and distance. S2. Determine cleaning sequence and make decisions: Move along the global path, use multimodal sensors to identify debris, and confirm effective cleaning targets; determine the type of debris. If it is ordinary debris, skip to S3; if it is stubborn sticky stains, skip to S4. S3, Routine Cleaning: Cluster and sort ordinary waste, plan the optimal path for a single operation, and perform cleaning. S4. Iterative Cleaning: Initiate a dynamic iterative cleaning cycle, which cleans stubborn stains multiple times through execution, evaluation, and adjustment until they are removed or the maximum number of attempts is reached. S5. Generation Path: For each waste area, the equipment generates multiple candidate cleaning points around it, and determines the final cleaning reference node by evaluating the cleaning cost of these candidate points; S6. Learning and Updating: Based on all the cleaning results, update the environmental cleanliness heatmap to provide an optimization basis for the next cleaning task; Preferably, in step S2, the waste sorting process uses a cost evaluation model to evaluate the waste cleaning cost, and the expression of the cost evaluation model is: cost = w1*D + w2*P + w3*I; Where D represents distance cost, P represents orientation cost, I represents interference cost, w is the respective weight, dis_obs is the distance between the robot and the obstacle, and length_robot is the length of the robot; The evaluation model considers the costs and weights of different factors, such as the absolute distance between the garbage and the cleaning robot, the orientation in front, to the side or behind, the expected interference or obstacles based on sensor data, and the prediction of the time and energy consumption required for cleaning.

[0005] Preferably, in step S5, as the device moves along the area, the constructed environmental cleanliness heatmap is dynamically updated along the movement path. This dynamically updated cleanliness heatmap is essentially a Bayesian probability model, the expression of which is: P t(i,j) represents the probability that there is trash in grid (i,j) that needs to be cleaned at time t; t (i,j) is the amount of garbage actually found in grid (i,j) during this cleaning at time t. Garbage amount can be quantified by garbage detection confidence, pixel percentage, or garbage quantity. I max : Normalized upper bound for waste amount; α: Learning rate (0 < α < 1), representing the influence of the current observation on the probability update; γ: History decay factor (0 < γ < 1), representing the self-decay of historical probabilities. The self-decay mechanism is used to simulate situations where waste is naturally cleared or no longer appears. N i,j : The weights of the neighboring grid (k,l) (e.g., the closer the distance, the higher the weight); w k,l The sum of all weights, used for normalization; During the inspection, the robot prioritizes going to P. t The region with the highest (i,j) value.

[0006] Preferably, in step S6, after the device completes a cleaning task, its reinforcement learning framework uses the Q-Learning algorithm to replan the optimal solution that appears during the cleaning process. The expression of the Q-Learning algorithm is: Define the state, S: the robot's state, which can be represented as S = waste type and ground material; Define Action, A: The robot's cleaning action, which can be represented as A = brush speed level, suction level, and path angle; Define Reward, R: The immediate feedback after the action is executed; R = β1 (reduction in cleaning area) - β2 (time consumed) - β3 (power consumption) - β4 (wear and tear on cleaning mechanism); β is the weighting coefficient of each indicator; During the learning process, each area cleaning by the device is a learning session. The reinforcement learning framework it uses employs the Q-Learning algorithm to record problems that occur during the cleaning process. A Q-table is built based on the problems encountered, and the cleaning content is updated based on the Q-table. The expression for the Q table is: Q(S,A)←Q(S,A)+η·[R+ξ·max A′ Q(S′,A′)-Q(S,A)]; Q(S, A): The long-term expected value of taking action A in state S; η: Learning rate 0 < η < 1, controlling the impact of new information; ξ: Discount factor 0 < ξ < 1, measuring the importance of future rewards; S: The new state reached after performing action A (it may still be the same type of waste, but the cleanliness level has changed); max A′ Q(S′,A′): The maximum expected value that can be obtained under the new state S′; The device abstracts the cleaning scene as a state and the parameter configuration of the cleaning mechanism as an action. Through the Q-Learning algorithm, the device obtains reward signals by interacting with the environment. The reward signals are calculated by comprehensively considering the cleaning effect, efficiency, and energy consumption, and its Q value is continuously updated accordingly.

[0007] Preferably, in step S2, as the device moves within the area and passes a stubborn stain, the device establishes a comprehensive confidence score C. stain To determine whether it is a stubborn stain, the overall confidence score C is used. stain Specifically, it is by: C stain =w color ·ΔC+w texture ·ΔT x +w torque ·ΔM; ΔC: Color difference between the stained area and the surrounding floor (ΔE value can be used in the CIELAB color space); ΔT x The difference in texture features between the stained area and the surrounding floor (features such as contrast and entropy can be calculated using the Gray-Level Co-occurrence Matrix (GLCM)); ΔM: The deviation between the average motor torque of the cleaning mechanism when it passes through this area and the standard value; w: These are the weighting coefficients for color, texture, and torque, respectively; color +w texture +w torque =1; C threshold : Boundaries of model decision-making; Therefore, when C stain ≥C threshold If the stain is found to be stubborn, the stubborn stain cleaning mode will be activated. After each iteration of cleaning, the effectiveness score E of the cleaning is evaluated; the effectiveness score E is specifically determined by: The conclusion is A before A after The visually visible area of ​​the stain before and after this cleaning; M before M after Before and after this cleaning, the cleaning motor maintained the same brush speed and average torque value; the decrease in torque indicates a reduction in sticky deposits. α, β: weighting coefficients, and α+β=1, which can be adjusted according to the type of stain. For example, for very sticky stains, a higher weight β is given to the torque change. After cleaning is complete, when E≥E threshold If the cleaning is successful, the cycle is exited. In step S1, the equipment performs stable filtration, eliminating temporary obstacles and mobile debris, focusing on deterministic waste; multiple scans of the same area confirm the continued presence of waste. If debris does not appear in several consecutive observations, it is considered temporary debris that is not cleaned, implementing a time window mechanism. For example, if an object has been present for a certain period of time, it is considered stable waste. At the same time, image recognition technology is used to distinguish the inherent parts of the waste and the environment, eliminating misjudged objects such as those that may just be ground patterns or stains that are difficult to clean for a long time.

[0008] Preferably, in step S1, the device completes stability filtration, performs distance filtration, defines the working range by combining the robot's position and cleaning ability, and selects waste that meets the conditions.

[0009] Preferably, in step S1, after completing stable filtration, the device records and analyzes the cleaning logs of the area's waste, marks the waste that has been repeatedly cleaned but failed, and avoids it.

[0010] Preferably, in step S1, the device uses the DBSCAN algorithm to group the identified waste according to its category and distance, and uses the DBSCAN algorithm to group similar waste that are close to each other in space into the same group.

[0011] The technical effects and advantages of this application are as follows: Compared with existing technologies, this method for cleaning robots to inspect and clean garbage offers high accuracy in garbage identification and classification, reducing ineffective cleaning actions at the source. In the S1 garbage disposal step, a triple filtering mechanism of "stability + distance + historical feedback" effectively eliminates temporary obstacles, mobile debris, and misjudgments of ground patterns—garbage is only considered a cleaning target after multiple scans confirm its continuous presence; if debris is not observed continuously, it is judged as temporary debris, preventing the robot from wasting resources on unnecessary targets. Simultaneously, the DBSCAN algorithm is used to group garbage by category and distance, achieving centralized management of similar garbage. In the S2 step, multimodal sensors combined with a comprehensive confidence score C... stain By distinguishing between ordinary waste and stubborn, sticky stains, and ensuring that cleaning strategies match waste types, this identification and scientific grouping improves the accuracy of waste identification and provides a clear basis for subsequent cleaning sequence planning, avoiding efficiency losses caused by "indiscriminate cleaning".

[0012] Compared with existing technologies, this method for cleaning robots to inspect and clean up garbage involves several improvements. In step S2, when determining the cleaning sequence, a cost evaluation model is used to comprehensively consider distance cost, orientation cost, and interference cost, and weighted allocation is used to quantify the cleaning cost. In step S5, multiple candidate cleaning points are generated around the garbage, and the final reference node is determined by evaluating the cost, further optimizing the single cleaning path. Furthermore, in steps S5 and S6, a Bayesian probability model is used to construct a dynamic environmental cleanliness heatmap, updating the garbage presence probability P of grid (i,j) in real time. t+1 (i,j) represents the area with the highest probability of being inspected by the robot, avoiding blind traversal and enabling the robot to cover high-priority areas with the lowest energy consumption and shortest time, reducing ineffective movement and resource waste.

[0013] Compared with existing technologies, this method for cleaning robot inspection and garbage collection introduces the Q-Learning reinforcement learning framework, allowing the robot to accumulate experience through each cleaning task and continuously optimize its cleaning strategy. In step S6, while learning using Q-Learning, the device records the long-term expected value of actions in different states through Q(S,A) in the Q-table. After each cleaning, the robot updates the Q-table based on the actual results, adjusting the learning rate η and the discount factor ξ to ensure a balance between new information and future rewards. This allows the device to learn continuously without human intervention, adapting to different cleaning scenarios (such as homes, shopping malls, and office buildings). The cleaning effect gradually improves with long-term use, while reducing maintenance costs. Attached Figure Description

[0014] Figure 1 This is a flowchart of a cleaning robot's inspection and garbage cleaning method according to the present invention; Figure 2 This is a flowchart illustrating the waste disposal process of a cleaning robot's inspection and waste cleaning method according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0016] Example A method for a cleaning robot to inspect and clean up garbage includes the following steps: S1. Waste Management: Based on the waste heat map of the area, the same area is scanned and observed multiple times, and stability filtering, distance adaptive filtering and historical feedback filtering are performed in sequence to identify and confirm the existence of priority waste and secondary waste, and the identified waste is grouped according to category and distance. S2. Determine cleaning sequence and make decisions: Move along the global path, use multimodal sensors to identify debris, and confirm effective cleaning targets; determine the type of debris. If it is ordinary debris, skip to S3; if it is stubborn sticky stains, skip to S4. S3, Routine Cleaning: Cluster and sort ordinary waste, plan the optimal path for a single operation, and perform cleaning. S4. Iterative Cleaning: Initiate a dynamic iterative cleaning cycle, which cleans stubborn stains multiple times through execution, evaluation, and adjustment until they are removed or the maximum number of attempts is reached. S5. Generation Path: For each waste area, the equipment generates multiple candidate cleaning points around it, and determines the final cleaning reference node by evaluating the cleaning cost of these candidate points; S6. Learning and Updating: Based on all the cleaning results, update the environmental cleanliness heatmap to provide an optimization basis for the next cleaning task; In this embodiment, the operation of the cleaning robot relies on the collaborative work of a series of core algorithms, including but not limited to: the DBSCAN clustering algorithm for garbage grouping, the cost model for evaluating cleaning priority, the Bayesian probability model for updating environmental cognition, the Q-learning algorithm for reinforcement learning, and the Hybrid A* path planning algorithm for final path generation. The specific application of each algorithm will be described in detail below in conjunction with each step as a preferred implementation. In step S2, the garbage sorting process uses a cost evaluation model to evaluate the cleaning cost of garbage. The expression of the cost evaluation model is: cost = w1*D + w2*P + w3*I; Where D represents distance cost, P represents orientation cost, I represents interference cost, w is the respective weight, dis_obs is the distance between the robot and the obstacle, and length_robot is the length of the robot; The evaluation model considers the costs and weights of different factors, such as the absolute distance between the garbage and the cleaning robot, the orientation in front, to the side or behind, the expected interference or obstacles based on sensor data, and the prediction of the time and energy consumption required for cleaning.

[0017] In a preferred embodiment, in step S5, as the device moves along the area, the constructed environmental cleanliness heatmap is dynamically updated along the movement path. This dynamically updated cleanliness heatmap is essentially a Bayesian probability model, the expression of which is: P t+1 (i,j) represents the probability that there is trash in grid (i,j) that needs to be cleaned at time t; t (i,j) is the amount of garbage actually found in grid (i,j) during this cleaning at time t. Garbage amount can be quantified by garbage detection confidence, pixel percentage, or garbage quantity. I max : Normalized upper bound for waste amount; α: Learning rate (0 < α < 1), representing the influence of the current observation on the probability update; γ: History decay factor (0 < γ < 1), representing the self-decay of historical probabilities. The self-decay mechanism is used to simulate situations where waste is naturally cleared or no longer appears. N i,j : The weights of the neighboring grid (k,l) (e.g., the closer the distance, the higher the weight); w k,l The sum of all weights, used for normalization; During the inspection, the robot prioritizes going to P. t The region with the highest (i,j) value.

[0018] First, as the device moves along the area, the environmental cleanliness heat map it constructs is dynamically updated along the path of movement. A sampling area is generated with the garbage to be cleaned as the center. The shape of the sampling area is determined according to the type of garbage and the environmental layout. The environmental layout is circular, elliptical, or ring-shaped, and the area is large enough to ensure that the cleaning robot has enough space to adjust and approach the garbage from multiple angles. Within this sampling area, a reference node of grid (i,j) is generated according to a specific interval.

[0019] For each reference node, its properties are evaluated, including the longitudinal distance of the cleaning position relative to the debris; the lateral distance of the cleaning position relative to the debris; and the dynamic range and shape of the side brush of the specific cleaning mechanism of the cleaning robot being cleaned. Based on the attributes evaluated above, a cleaning cost is calculated for each reference node. This cost takes into account the difficulty of reaching the point, the efficiency of cleaning, and the need for obstacle avoidance. These factors are quantified using a Bayesian probability model, including the distance from the pre-cleaning location to the robot's current location and the estimated cost of the feasible travel path; the angular deviation between the cleaning direction and the robot's current orientation; and the impact of potential moving or fixed obstacles in the environment on the cleaning operation. Then, the cleaning costs of each reference node are compared, and the cleaning reference node with the best cleaning effect and the lowest cost is selected as the optimal cleaning reference node. This cleaning reference node provides the best cleaning angle, distance and path to ensure the efficiency and safety of the cleaning operation. Based on the cleaning reference node information of the first garbage finally selected, the navigation path of the robot from the current position to the pre-cleaning position is generated by the Hybrid A* path planning algorithm. Based on the given cleaning sequence and cleaning reference node information, a polynomial curve is used to connect the pre-cleaning and precise cleaning locations of multiple wastes in sequence to determine an initial cleaning reference route.

[0020] Using dynamic programming algorithms, based on an initial cleaning reference route, path planning is completed by constructing a state space, defining state transition equations, determining objective functions and constraints, handling initial conditions and edge cases, making decisions in stages, and performing recursive and iterative calculations. This generates a feasible, safe, and efficient cleaning path that must conform to the robot's motion characteristics and adapt to environmental constraints and cleaning task requirements.

[0021] Based on the cleaning path, the specific action sequence for performing the cleaning task is specified, including the instructions of the cleaning robot's side brushes, roller brushes, and other cleaning mechanisms. The cleaning result, the removal of garbage at position (i,j), will be used as key data and fed back to the Bayesian model to execute the update formula P_t+1(i,j), thereby reducing the probability value of that point.

[0022] In a preferred embodiment, in step S6, after the device completes a cleaning task, its reinforcement learning framework uses the Q-Learning algorithm to replan the optimal solution that appears during the cleaning process. The expression of the Q-Learning algorithm is: Define the state, S: the robot's state, which can be represented as S = waste type and ground material; Define Action, A: The robot's cleaning action, which can be represented as A = brush speed level, suction level, and path angle; Define Reward, R: The immediate feedback after the action is executed; R = β1 (reduction in cleaning area) - β2 (time consumed) - β3 (power consumption) - β4 (wear and tear on cleaning mechanism); β is the weighting coefficient of each indicator; During the learning process, each area cleaning by the device is a learning session. The reinforcement learning framework it uses employs the Q-Learning algorithm to record problems that occur during the cleaning process. A Q-table is built based on the problems encountered, and the cleaning content is updated based on the Q-table. The expression for the Q table is: Q(S, A) ← Q(S, A) + η·[R + ξ·max A′ Q(S′,A′)-Q(S,A)]; Q(S,A): The long-term expected value of taking action A in state S; η: Learning rate 0 < η < 1, controlling the impact of new information; ξ: Discount factor 0 < ξ < 1, measuring the importance of future rewards; S: The new state reached after executing action A. It still deals with the same type of garbage, but the cleanliness level has changed here. max A′ Q(S′,A′): The maximum expected value that can be obtained under the new state S′; The device abstracts the cleaning scene as a state and the parameter configuration of the cleaning mechanism as an action. Through the Q-Learning algorithm, the device obtains reward signals by interacting with the environment. The reward signals are calculated by comprehensively considering the cleaning effect, efficiency, and energy consumption, and its Q value is continuously updated accordingly.

[0023] In a preferred embodiment, in step S2, the device moves within the area and passes by a stubborn stain, and the device establishes a comprehensive confidence score C. stain To determine whether it is a stubborn stain, the overall confidence score C is used. stain Specifically, it is by: C stain =w color ·ΔC+w texture ·ΔT x +w torque ·ΔM; ΔC: Color difference between the stained area and the surrounding floor (ΔE value can be used in the CIELAB color space); ΔT x The difference in texture features between the stained area and the surrounding floor (features such as contrast and entropy can be calculated using the Gray-Level Co-occurrence Matrix (GLCM)); ΔM: The deviation between the average motor torque of the cleaning mechanism when it passes through this area and the standard value; w: These are the weighting coefficients for color, texture, and torque, respectively; color +w texture +w torque =1; C threshold : Boundaries of model decision-making; Therefore, when C stain ≥C threshold If the stain is found to be stubborn, the stubborn stain cleaning mode will be activated. After each iteration of cleaning, the effectiveness score E of the cleaning is evaluated; the effectiveness score E is specifically determined by: The conclusion is A before A after The visually visible area of ​​the stain before and after this cleaning; M before M after Before and after this cleaning, the cleaning motor maintained the same brush speed and average torque value; the decrease in torque indicates a reduction in sticky deposits. α, β: weighting coefficients, and α+β=1, which can be adjusted according to the type of stain. When a very sticky stain is found, a higher weight β is given to the torque change after the second cleaning. After cleaning is complete, when E≥E threshold If the cleaning is successful, the cycle is exited. In the preferred implementation, in step S1, the device performs stable filtration, eliminating temporary obstacles and mobile debris, focusing on deterministic waste. Multiple scans of the same area confirm the continued presence of waste. If debris does not appear in several consecutive observations, it is considered temporary debris and not cleaned, implementing a time window mechanism. For example, if an object has consistently existed for a certain period, it is considered stable waste. Simultaneously, image recognition technology is used to distinguish the waste from the inherent parts of the environment, eliminating misjudged objects such as patterns on the ground or stains that are difficult to clean over a long period.

[0024] In a preferred embodiment, in step S1, the device performs stability filtration, distance filtration, defines the working range by combining the robot's position and cleaning ability, and selects waste that meets the conditions.

[0025] In a preferred embodiment, in step S1, after completing stable filtration, the device records and analyzes the cleaning logs of the area's waste, marks the waste that has been repeatedly attempted to be cleaned but failed, and avoids it.

[0026] In a preferred embodiment, in step S1, the device uses the DBSCAN algorithm to group the identified waste according to category and distance. The DBSCAN algorithm groups similar waste that are close neighbors in space into the same group. DBSCAN is a density-based clustering algorithm that can identify clusters of arbitrary shapes and does not require pre-specifying the number of clusters. Therefore, it is very suitable for clustering spatially distributed data such as waste where the number of clusters cannot be determined. The specific clustering steps are as follows: 1. Parameter Setting: Set the radius (ε) to one car width and the minimum number of points (MinPts) to 1, allowing a single piece of trash to form a cluster. 2. Mark All Objects as Unprocessed: Initially, all trash items are considered unprocessed. 3. Select an Unprocessed Object: Randomly select an unprocessed trash item to start. 4. Find the Core Object Based on Radius ε and MinPts: Check the number of trash items within the ε-neighborhood of the selected trash item. If the number of trash items reaches MinPts, mark the current trash item as the core object and begin creating a new trash cluster. 5. Neighborhood Expansion: If a core trash item is found, check all trash items within its neighborhood and add them to the current trash cluster. For each trash item within its neighborhood, continue searching for samples within its ε-neighborhood. If a core object is found, expand the current cluster. 6. Handling Boundary Objects: Even if a trash cluster is not a core trash cluster, it can become part of a trash cluster as long as it is within the neighborhood of other core trash items. In this invention, MinPts is fixed at 1, so there are no boundary objects. 7. Continue selecting new unprocessed objects: Repeat steps 4-6 until all junk items have been processed.

[0027] The working process of this application is as follows: First, based on the waste heat map of the area, the same area is scanned and observed multiple times, and the existence of priority waste and secondary waste is identified and confirmed through stability, distance and historical feedback filtering, and the identified waste categories and distances are grouped. Moving along a global path, it uses multimodal sensors to identify debris and confirm effective cleaning targets; it determines the type of debris: if it is ordinary debris, it performs regular cleaning; if it is stubborn, sticky stains, it performs iterative cleaning. Routine cleaning: Cluster and sort ordinary waste, plan the optimal path for a single cleaning, and perform cleaning; Iterative cleaning: Initiate a dynamic iterative cleaning cycle, and clean stubborn stains multiple times through execution, evaluation, and adjustment until they are removed or the maximum number of attempts is reached; For each waste area, the equipment generates multiple candidate cleaning points around it, and determines the final cleaning reference node by evaluating the cleaning cost of these candidate points; Based on all the cleaning results, the environmental cleanliness heatmap is updated to provide an optimization basis for the next cleaning task. The above is the working principle of this cleaning robot's inspection and garbage cleaning method.

Claims

1. A method of garbage sweeping by a cleaning robot, characterized in that, The method comprises the following steps: S1, garbage disposal: based on the garbage heat map of the area, the same area is scanned and observed multiple times, and stability filtering, distance adaptability filtering and historical feedback filtering are sequentially performed to identify and confirm the existence of priority garbage and secondary garbage, and the identified garbage is grouped according to categories and distances; S2, determine the cleaning sequence and decision: move along the global path, identify the garbage using multi-modal sensors, and confirm the effective cleaning target; judge the garbage type, if it is ordinary garbage, jump to S3; if it is stubborn sticky stain, jump to S4; S3, regular cleaning: clustering, sorting and planning the single optimal path for ordinary garbage, and executing cleaning; S4, iterative cleaning: starting a dynamic iterative cleaning cycle, cleaning stubborn stains multiple times through execution, evaluation and adjustment until they are cleaned or the maximum number of attempts is reached; S5, generate path: for each garbage area, the device generates multiple candidate cleaning points around it, and determines the final cleaning reference node by evaluating the cleaning cost of these candidate points; S6, learning and updating: update the environmental cleanliness heat map according to all cleaning results this time to provide optimization basis for the next cleaning task.

2. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S2, the garbage sorting process uses a cost evaluation model to evaluate the cleaning cost of garbage, and the expression of the cost evaluation model is: cost = w1*D + w2*P + w3*I where D represents a distance cost, P represents an orientation cost, I represents an interference cost, w is a respective weight, dis obs is a distance of the robot from the obstacle, length robot is a length of the robot; The evaluation model considers the cost and weight of different factors, such as the absolute distance between the garbage and the cleaning robot, the direction of the front, side or back, the expected interference or obstacle based on sensor data, and the prediction of the time and energy consumption required for cleaning.

3. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S5, the constructed environmental cleanliness heat map is dynamically updated along the moving path when the device moves along the area, and this dynamically updated cleanliness heat map is essentially a Bayesian probability model, and the expression of the Bayesian probability model is: P t (i,j) represents the probability that at time t, the grid (i,j) has garbage that needs to be cleaned; I t (i,j) the amount of litter actually found in the grid (i,j) at time t in the current cleaning; I max : upper limit of normalization of the amount of garbage α: learning rate 0<α<1, indicating the influence degree of this observation value on probability update; γ: historical decay factor 0<γ<1, indicating the self-decay of historical probability, and the self-decay mechanism is used to simulate the situation that garbage is naturally cleaned or no longer appears N i,j : weight of the neighborhood grid (k,l) (e.g. the closer the distance the higher the weight); w k,l : sum of all weights for normalization; Upon inspection, the robot goes to P t the region with the highest (i,j) value.

4. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S6, after completing a cleaning task, the reinforcement learning framework used by the device uses the Q-Learning algorithm to re-plan the optimal solution that appears during cleaning, and the expression of the Q-Learning algorithm is: Define state State, S: the state of the robot, which can be represented as S = garbage type and ground material; Define action Action, A: the cleaning action of the robot, which can be represented as A = brush speed gear, suction force gear and path angle; Define reward Reward, R: the immediate feedback after executing the action; R = β1·(cleaning area reduction amount) - β2·(time consumption) - β3·(electricity consumption) - β4·(cleaning mechanism wear degree)·; β is the weight coefficient of each index; The device learns in the learning process, and each time the area is cleaned, it is a learning. The reinforcement learning framework used by the device itself adopts the Q-Learning algorithm to record the problems that occur during cleaning. The Q table is established based on the problems that occur, and the cleaning content is updated based on the Q table. The expression of Q-table is: Q(S,A)←Q(S,A)+η·[R+ξ·max A′ Q(S',A')-Q(S,A)] Q(S, A): Long-term expected value of performing action A in state S; η: Learning rate 0 < η < 1, controlling the impact of new information; ξ: Discount factor 0 < ξ < 1, measuring the importance of future rewards; S: New state reached after performing action A; max A′ Q(S', A') : the maximum expected value that can be obtained in the new state S' The device abstracts the cleaning scene as a state and abstracts the parameter configuration of the cleaning mechanism as an action. Through the Q-Learning algorithm, the device obtains a reward signal through interaction with the environment. The reward signal is calculated based on cleaning effect, efficiency, and energy consumption, and the Q value is updated accordingly.

5. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S2, the device moves within the area and passes over a stubborn stain, the device sets up a comprehensive confidence score C stain to determine whether it is a stubborn stain, and the comprehensive confidence score C stain is specifically: C stain = w color · ΔC + w texture · ΔT + w x · ΔM torque · ΔM ΔC: Color difference between the stain area and the surrounding floor (ΔE value in CIELAB color space can be used); ΔT x : difference in texture features between the stain area and the surrounding floor (contrast, entropy, etc. features can be computed by a Gray Level Co-occurrence Matrix GLCM); ΔM: Deviation of the average motor torque of the cleaning mechanism from the standard value when passing through the area; w: weight coefficient of color, texture, torque, respectively; w color + w texture + w torque = 1; C threshold : boundaries of model decisions; Thus when C stain ≥ C threshold then the stain is determined to be a tough stain and the tough stain cleaning mode is initiated. The device performs each iteration of cleaning, and evaluates the effectiveness score E of this cleaning; the effectiveness score E is specifically derived from: yielding A before , A after : visual area of the stain before and after the cleaning; M before , M after : the average torque value of the cleaning motor is maintained at the same brush speed before and after cleaning; α, β: Weight coefficients, and α + β = 1; After cleaning is completed, when E ≥ E threshold then cleaning is considered successful and the loop is exited.

6. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S1, the device stabilizes the filter, excludes temporary obstacles and flowable debris, and focuses on determining the certainty of garbage. Multiple scans of the same area confirm the persistence of garbage. If the debris does not appear in several observations, it is considered temporary debris and is not cleaned, implementing a time window mechanism.

7. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S1, the device completes the stability filtering, performs distance filtering, and defines the work range based on the robot position and cleaning capacity, and selects the garbage that meets the conditions.

8. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S1, after completing the stability filtering and stability filtering, the device records and analyzes the cleaning log of the area garbage, marks the garbage that has been unsuccessfully cleaned multiple times, and avoids it.

9. The garbage sweeping method of the cleaning robot according to claim 1, wherein: In step S1, the device uses the DBSCAN algorithm to group the identified garbage according to categories and distances, and uses the DBSCAN algorithm to group similar garbage in the space into the same group.

Images