Control method and device for cleaning robot for underwater steps of swimming pool

Abstract

The embodiment of the invention relates to the technical field of robot control, and provides a control method and device of a swimming pool underwater step cleaning robot. The method comprises the following steps: scanning a step structure in a target step area to construct a three-dimensional step model; an initial control strategy is generated through a depth deterministic strategy gradient algorithm according to the three-dimensional step model before water entry cleaning, and the robot is started to enter a target step area based on the initial control strategy; in the cleaning process, motion state data and posture monitoring data of the robot are collected in real time; based on a self-adaptive control model of an inverted pendulum balance theory, performing collaborative optimization on motion state data and attitude monitoring data which are acquired in real time and step structure parameters in the three-dimensional step model to obtain a real-time control strategy of the robot; and the motion state of the robot is adjusted based on the real-time control strategy, the step cleaning process in the target step area is adapted, and the environment adaptability, the posture stability and the operation efficiency of the robot in the complex underwater step environment are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot control, and in particular to a control method and device of a pool underwater step cleaning robot. BACKGROUND

[0002] At present, cleaning robots can be used to replace traditional manual underwater cleaning operations to achieve targeted cleaning in the pool underwater step area, remove scale, silt, algae and other pollutants in the gap of the step surface, and ensure the cleanliness and use safety of the pool underwater environment.

[0003] In related technologies, due to the differences in the height, width and vertical inclination angle of different pool steps, and the fixed structure of the movement mechanism of the existing cleaning robot, it is difficult to flexibly adapt to steps of different specifications, and situations such as step climbing jam and deviation at corners are prone to occur, resulting in cleaning omission in some areas. In addition, due to the buoyancy effect of the underwater environment and the unevenness and complex structure of the step surface, the attitude stability of the cleaning robot is poor, which is prone to cause the cleaning robot to tilt or overturn during the cleaning process, reduce the cleaning intensity, and cause robot failure. Therefore, it is urgent to propose a technical solution to solve at least one of the above technical problems. SUMMARY

[0004] The main purpose of the embodiments of the present application is to provide a control method and device of a pool underwater step cleaning robot, which aims to solve at least one of the technical problems of poor attitude stability, low cleaning intensity and robot failure caused by the inadaptability between the robot and the step structure in related technologies.

[0005] In a first aspect, the embodiments of the present application provide a control method of a pool underwater step cleaning robot, comprising: scanning the step structure in the target step area to construct a three-dimensional step model; before entering the water for cleaning, generating an initial control strategy according to the three-dimensional step model by a deep deterministic policy gradient algorithm, the initial control strategy at least including one of the following parameters: track opening angle, drive wheel spacing and propulsion power, and starting the robot into the target step area based on the initial control strategy; in the cleaning process, real-time collecting the motion state data and attitude monitoring data of the robot; based on an adaptive control model of inverted pendulum balance theory, cooperatively optimizing the real-time collected motion state data, attitude monitoring data and step structure parameters in the three-dimensional step model to obtain the real-time control strategy of the robot; the adaptive control model is constructed by a sliding mode control, a fuzzy control and an inverted pendulum control model; adjusting the motion state of the robot based on the real-time control strategy to adapt to the step cleaning process in the target step area.

[0006] In a second aspect, the embodiments of the present application provide a control device of a pool underwater step cleaning robot, comprising: an acquisition module configured to scan a step structure in a target step region to construct a three-dimensional step model; a starting module configured to generate an initial control strategy according to the three-dimensional step model by a deep deterministic policy gradient algorithm before water entry cleaning, the initial control strategy comprising at least one of the following parameters: a track opening angle, a driving wheel distance and a propulsion power, and starting the robot to enter the target step region based on the initial control strategy; a cleaning module configured to acquire motion state data and attitude monitoring data of the robot in real time during the cleaning process; an adaptive control model based on an inverted pendulum balance theory is used to cooperatively optimize the motion state data, the attitude monitoring data and step structure parameters in the three-dimensional step model to obtain a real-time control strategy of the robot; the adaptive control model is constructed by a sliding mode control, a fuzzy control and an inverted pendulum control model; and the motion state of the robot is adjusted based on the real-time control strategy to adapt to the step cleaning process in the target step region.

[0007] In a third aspect, the embodiments of the present application further provide an electronic device, which comprises a processor, a memory for storing a computer program, and the processor is configured to execute the computer program and implement the control method of the pool underwater step cleaning robot according to the first aspect or any of the embodiments of the present application.

[0008] In a fourth aspect, the embodiments of the present application further provide a computer readable storage medium, which stores a computer software program, and the computer software program is executed by a processor to implement the control method of the pool underwater step cleaning robot according to the first aspect or any of the embodiments of the present application.

[0009] The embodiment of the present application provides a control method and device of a pool underwater step cleaning robot. In the embodiment of the present application, a three-dimensional step model is constructed by scanning step structures in a target step region. Before water entry cleaning, an initial control strategy is generated by a deep deterministic policy gradient algorithm according to the three-dimensional step model, the initial control strategy at least including one parameter of a track opening angle, a driving wheel distance and a propulsion power, and the robot is started to enter the target step region based on the initial control strategy. In the cleaning process, motion state data and attitude monitoring data of the robot are collected in real time. Then, the motion state data, the attitude monitoring data and step structure parameters in the three-dimensional step model are cooperatively optimized based on an adaptive control model of an inverted pendulum balance theory, so as to obtain a real-time control strategy of the robot. The adaptive control model is constructed by a sliding mode control, a fuzzy control and an inverted pendulum control model. Finally, the motion state of the robot is adjusted based on the real-time control strategy, so as to adapt to the step cleaning process in the target step region. The embodiment of the present application can improve the environmental adaptability, the attitude stability and the operation efficiency of the robot in a complex underwater step environment, and solves the problem of insufficient adaptability of a traditional underwater cleaning device to an unstructured step environment. BRIEF DESCRIPTION OF DRAWINGS

[0010] Figure 1 A flowchart of a control method of a pool underwater step cleaning robot is provided for the embodiment of the present application. Figure 2 A scene diagram of a control method of a pool underwater step cleaning robot is provided for the embodiment of the present application. DETAILED DESCRIPTION

[0011] Some embodiments of the present application will be described in detail below with reference to the accompanying drawings. Figure 1 A flowchart of a control method of a pool underwater step cleaning robot is provided for the embodiment of the present application. Please refer to Figure 1 The method comprises the following steps: Step S101, a three-dimensional step model is constructed by scanning step structures in a target step region.

[0012] In the embodiment of the present application, the pool can be a public pool, a private pool, a landscape pool and various pools with water environment, which is a specific operation object of the whole cleaning process.

[0013] Step S101 relieves the sensing accuracy limitation caused by underwater light scattering and turbid water through multi-source sensor fusion and graph neural network topology modeling, and realizes the targeted extraction of structure parameters such as step height, width, and vertical inclination angle, and surface flatness information. In addition, three-dimensional model construction not only provides the step structure as prior knowledge for subsequent control strategy, but also balances model accuracy and computing efficiency through dynamic grid optimization, improves the technical problems of traditional underwater robots such as fuzzy perception of complex step structure and loss of details, and lays the environmental data foundation for the subsequent targeted generation of control strategy.

[0014] In the embodiment of the application, the target step area refers to a specific area in the swimming pool where steps are set and cleaning operations are needed. The specific range needs to be determined in combination with the type and structure of the swimming pool, such as the ladder area at the entrance of the public swimming pool for convenient personnel access, the step area between the shallow and deep water areas of the private swimming pool, the steps and surrounding connection areas in the landscape swimming pool with landscaping design, etc. These areas not only include the steps themselves, but also cover the transition part connecting the steps and the pool wall and the pool bottom, and the area around the steps where water flow disturbance or pollutant accumulation may exist. It is the specific action range of the cleaning robot when performing underwater step cleaning tasks, and is also the core object of subsequent three-dimensional model construction and control strategy formulation. For example, the target step area can be Figure 2 The underwater step shown, the robot can be Figure 2 The underwater robot shown. Figure 2 In the embodiment, the underwater robot is cleaning the target step area.

[0015] The step structure refers to various physical form features that constitute the underwater steps of the swimming pool, mainly including the horizontal step surface, i.e. the flat part for personnel to tread or the robot to move. The vertical or inclined step surface, i.e. the side part connecting two adjacent step surfaces. The connection part between the step surface and the step surface, such as the corner, gap and other detail structures. In addition, the step structure also includes the overall arrangement form of the step itself, and the physical state of the step surface, such as whether there are protrusions, depressions and other uneven features. These structural features together determine the spatial form and cleaning difficulty of the step, and are the core object of scanning and obtaining structure parameters (such as height, width, and vertical inclination angle) and flatness information, directly affecting the construction accuracy of the subsequent three-dimensional model and the adaptability of the control strategy.

[0016] The three-dimensional step model is a digital model constructed based on the structural parameters and flatness information of the target step region obtained by robot scanning, through multi-source data fusion and graph neural network topology modeling technology. It can intuitively and accurately present the spatial layout and detailed features of the target step region. The model not only integrates the height, width, and facade inclination angle of the step, but also incorporates the flatness information of the step surface, which can clearly reflect the details of the step surface, facade, corner, gap, and surface protrusions and depressions. At the same time, the model has dynamic grid optimization capability, which can adjust the grid density according to the cleaning needs and structural complexity of different regions of the step, balancing the accuracy of the model and the calculation efficiency. As a key basis for subsequent control strategy formulation, the three-dimensional step model can provide prior knowledge of the step structure for initial control strategy generation, and provide dynamic environmental data support for real-time control strategy optimization, effectively improving the problem of fuzzy perception and detail loss of traditional underwater robots for complex step structures, and ensuring that the robot can accurately adapt to the spatial form and surface state of the step during cleaning.

[0017] As an optional embodiment, in step S101, the target step region is scanned by the multi-source sensors integrated by the robot to obtain the structural parameters of the target step region. In the embodiment of the present application, the structural parameters include the height, width, and facade inclination angle of each step in the target step region. Then, the flatness information of the surface of each step in the target step region is obtained by using a ranging sensor. Finally, the spatial topology of the steps is modeled based on the structural parameters and the flatness information by using a graph neural network to construct a three-dimensional step model of the target step region.

[0018] Specifically, in the specific implementation process of step S101, the robot first scans the target step region with the help of integrated multi-source sensors, which usually include a depth camera and a laser radar. The two work together to cope with the particularity of the underwater environment. The depth camera is responsible for capturing the overall visual profile of the step region, and initially identifying the distribution range and arrangement order of the steps. The laser radar relies on high-frequency detection signals to accurately capture the spatial position data of each part of the step, such as the step region at the entrance of a public swimming pool. The depth camera first determines the sequence of steps extending from the pool edge to the pool, and the laser radar then measures each step in turn to obtain the vertical dimension data from the pool bottom to the step surface to determine the step height, and measures the inclination angle of the step facade and the vertical direction to determine the step facade inclination angle. During the scanning process, the robot moves slowly according to the preset path to ensure that the edge area and junction of the step are covered, and the sensor automatically filters out interference signals caused by turbidity of water or scattering of light to ensure that the collected structural parameters accurately match the actual physical form of the step.

[0019] After the structural parameters are collected, the robot switches to the ranging sensor to acquire the flatness information of the step surface. The ranging sensor emits directional probe signals to the step surface and receives reflected signals to determine the surface height of the probe points based on the signal feedback. For example, in the step area of the shallow water area of a private swimming pool, the robot adjusts its posture to make the ranging sensor vertically aligned with the step surface, and moves along the length and width directions of the step surface to detect the height difference between multiple adjacent probe points. The robot identifies whether there are local protrusions or depressions on the step surface, and slows down the detection speed in the corner area where the step surface meets the vertical surface to ensure that the flatness change of this part is captured. Finally, a set of flatness information covering the entire step surface is formed, which, together with the previously acquired structural parameters, constitutes the basic data for modeling.

[0020] Subsequently, the construction of the three-dimensional step model is entered. First, the structural parameters and flatness information are associated according to the detection position to form associated data of each detection point corresponding to the spatial position, structural characteristics, and surface state. Then, these associated data are mapped to the nodes of the graph neural network, each node containing the step height, width, vertical surface inclination angle, and flatness characteristics of the corresponding position. Next, the edges between nodes are defined according to the actual spatial connection relationship of the step, such as the adjacent detection point nodes of the same step and the corresponding nodes of adjacent steps, which are connected by edges to reflect their spatial topological association. The graph neural network learns the characteristics of these nodes and edges, aggregates the information of adjacent nodes, and mines the overall spatial layout rules of the step. For example, in the step area of a landscape pool with an arc-shaped vertical surface, the graph neural network can accurately identify the curvature and connection mode of the arc-shaped structure by learning the edge characteristics of the nodes of the arc-shaped vertical surface and the adjacent step surface nodes. Finally, all data are integrated into a three-dimensional model based on the learned topological relationship, and the model details are adjusted according to the structural complexity and cleaning needs of different areas of the step, such as strengthening the detail description of areas with frequent flatness changes and appropriately simplifying the presentation of structured and flat areas. Finally, a three-dimensional step model is formed that can accurately reflect the spatial form and surface state of the target step area.

[0021] It is worth noting that the training process of the graph neural network used in step S101 is as follows: first, a training data set is constructed. Collecting step data of different scenes such as public, private and landscape swimming pools, extracting the height, width, facade inclination angle and surface flatness quantitative value of each step as node features, the height difference, horizontal distance and connecting surface slope of adjacent steps as edge features, and labeling the real spatial topology relationship matrix of the corresponding steps. In order to simulate underwater interference, different intensity noises are added to the data to realize data enhancement. The graph neural network takes the node and edge features as input, aggregates the features through the graph convolution layer, focuses on the key nodes through the attention layer, and outputs the topology matrix through the full connection layer. The loss function adopts mean square error, and calculates the error between the predicted topology matrix and the real matrix, and the error between the predicted structure parameters and the actual parameters. Further, a large amount of simulated step data is used to pretrain the graph neural network, so that the graph neural network can master the basic topology rule. Then, the real underwater scanning data is introduced for fine tuning, and the parameters of each layer are optimized to adapt to the actual environment. The Adam optimizer is used for batch iteration, and the validation set error is monitored to prevent overfitting, until the graph neural network stably outputs the accurate topology relationship matrix, meeting the three-dimensional step modeling accuracy requirement.

[0022] Further optionally, in the above step, the step of modeling the spatial topological relationship of the steps based on the structure parameters and the flatness information by the graph neural network to construct a three-dimensional step model of the target step region comprises: encapsulating the structure parameters of each step and the flatness information of the corresponding region as nodes of the graph neural network, the structure parameter features of each node comprising one of: step height, step width, step facade inclination angle, and step surface flatness quantization value. Further, the edge features between the nodes are defined according to the spatial positional relationship between the steps, the edge features comprising one of: height difference, horizontal distance, and connecting surface slope of adjacent steps. Then, the graph neural network composed of a graph convolution layer, an attention mechanism layer, and a fully connected layer is used to aggregate the structure parameter features and the edge features, learn the topological correlation between the steps, and output a topological relationship matrix containing the global spatial layout. Then, the topological relationship matrix and the structure parameter features of each node are fused to generate a three-dimensional step model containing the step geometric parameters, surface state, and spatial connection relationship. The grid density of the three-dimensional step model is dynamically adjusted by a quadtree algorithm, and based on the optimized grid, the step three-dimensional geometric contour is constructed by a triangulation algorithm combined with the structure parameters of each node. The grid elements corresponding to each node are embedded with multi-dimensional textures converted from the flatness data by a luminance remapping algorithm to reflect the surface flatness difference between the nodes, so as to obtain the final output three-dimensional step model. Further optionally, the three-dimensional step model is dynamically corrected by the scanning data fed back by the sensor in real time, the grid density is dynamically adjusted according to the scanning progress, sparse grids are used in the flat step area to reduce the computational load, and the grid is automatically encrypted in the flatness mutation area. The convex features and the concave features of each node corresponding to the steps are stored. For example, the convex features and the concave features are obtained based on the structure parameters and the flatness information. The convex features are feature parameters corresponding to uneven structures in the three-dimensional step model that protrude upward compared to a preset reference surface of the step, and the concave features are feature parameters corresponding to uneven structures in the three-dimensional step model that recess downward compared to the preset reference surface of the step.

[0023] Specifically, in the implementation process of constructing a three-dimensional step model based on the structure parameters and flatness information through a graph neural network in the above steps, first, a node packaging operation is performed, that is, for each step in the target step region, the corresponding structure parameters and the flatness information of the step region are integrated into a node recognizable by the graph neural network. For example, in the multi-step region at the entrance of a public swimming pool, for the first step closest to the pool edge, its own height, width, facade inclination angle, and flatness quantification values at different positions on the step surface (such as the flatness difference between the middle flat area and the edge connecting area of the step) are collectively packaged into an independent node. Each node carries a unique identifier to correspond to the specific position of the actual step, ensuring accurate association between the physical step and the digital node during subsequent modeling. After completing node packaging, the edge features between nodes are defined according to the actual spatial position relationship between steps. Taking the three-step transition from the shallow water area to the deep water area in a private swimming pool as an example, between the first and second steps, the edge features are determined according to the height difference in the vertical direction, the distance in the horizontal direction, and the inclination slope of the connecting facade between the two steps. If the facades of the two steps are gently sloping, the connecting surface slope parameter in the edge feature will reflect this gentle characteristic accordingly, and for areas with obvious vertical connection between steps, the edge feature will record the corresponding steep slope information. In this way, the edges between nodes can accurately reflect the spatial connection relationship of the actual steps.

[0024] Subsequently, feature aggregation is performed using a graph neural network composed of a graph convolution layer, an attention mechanism layer, and a fully connected layer. The graph convolution layer iterates through each node and aggregates the feature information of its adjacent nodes. For example, in the step region with an arc design in a landscape pool, the graph convolution layer fuses the structure parameters and edge features of each node in the arc-shaped step to learn the overall spatial arrangement of the arc-shaped step. The attention mechanism layer focuses on key region nodes of the step, such as nodes at step corners and gaps, and their adjacent nodes, giving these nodes higher attention weights during feature learning to ensure that the spatial topological relationship at the corners is accurately captured. The fully connected layer converts the node feature vectors output by the graph convolution layer into comprehensive feature vectors containing structure adaptability, posture influence degree, and other information, and finally outputs a topological relationship matrix that reflects the global spatial layout of the entire target step region.

[0025] Then, the topological relationship matrix is fused with the structural parameter characteristics of each node to generate a preliminary three-dimensional step model, and the grid density of the model is dynamically adjusted through a quadtree algorithm. For example, in a step area with large differences in surface flatness in a swimming pool, for the middle area of the step with a flat surface and simple structure, the quadtree algorithm will use sparse grids to reduce the computational load. For the area with protrusions and depressions on the edge of the step, the algorithm will automatically increase the grid density to ensure that these detailed features can be clearly presented. Then, based on the optimized grid, combined with the structural parameters of each node, the three-dimensional geometric contour of the step is constructed through a triangulation algorithm, such as using the spatial coordinates of each node as the vertices of triangulation to connect and form a geometric surface that fits the actual step shape.

[0026] It is worth noting that further optionally, when constructing the three-dimensional geometric contour of the step through the triangulation algorithm, the core is to use the spatial coordinates of each node in the optimized grid to convert discrete nodes into continuous geometric surfaces that fit the actual step shape. The algorithm will first traverse all the nodes in the three-dimensional step model, select adjacent nodes according to the spatial position correlation of the nodes, and use the coordinates of these adjacent nodes as the vertices of the triangle to gradually construct a triangular grid that covers the entire step area. For example, when dealing with a flat step surface in a swimming pool, the algorithm will select evenly distributed adjacent nodes to form triangles of appropriate size, so that multiple triangles can be connected to form a flat step surface shape. When facing complex structures such as step corners or curved surfaces, the algorithm will automatically increase the number of triangles to fit the right-angle contour of the corner or the curved trajectory of the arc, avoiding deviations between the geometric contour and the actual step structure caused by large triangles. At the same time, the algorithm will also adjust the arrangement direction of the triangles according to the structural parameters of the nodes (such as the inclination angle of the surface), to ensure that the three-dimensional geometric contour accurately reflects the key features of the step, such as height difference and surface slope, and that the digital model is highly consistent with the physical step shape.

[0027] At the same time, the grid elements corresponding to each node are embedded with multi-dimensional textures converted from flatness data through a luminance remapping algorithm, such as the texture brightness of the protruding area of the step surface being significantly different from that of the depressed area, to intuitively reflect the differences in surface flatness between different nodes, and finally form a complete three-dimensional step model.

[0028] Further optionally, the luminance remapping algorithm is used to convert the flatness data of the step surface into a multi-dimensional texture that can be visually recognized, so as to clearly present the flatness difference in the three-dimensional step model. The luminance remapping algorithm first preprocesses the flatness quantization value obtained by the ranging sensor, and distinguishes the flat area, convex area and concave area of the step surface. Then, the flatness states of different types are mapped to the corresponding luminance range, for example, the flatness value of the flat area is corresponding to the medium luminance, the convex area is corresponding to the higher luminance, and the concave area is corresponding to the lower luminance, and the difference of flatness is embodied by the gradient change of luminance. When embedding the texture into the grid cell, the luminance remapping algorithm will give the grid cell a matching luminance attribute according to the node flatness data corresponding to each grid cell, so that the grid cell presents a visual effect consistent with the actual step surface flatness. For example, in the edge area of the public swimming pool step, if there is a slight convexity formed by long-term use, the grid cell corresponding to this area will be given a higher luminance, while the flat area in the middle of the step presents a medium luminance. Through the luminance difference, the flatness distribution of the step surface can be judged intuitively, and a more intuitive surface state reference is provided for robot cleaning path planning and posture control.

[0029] Further, when embedding the texture into the grid cells, the luminance remapping algorithm introduces a double mapping logic of ambient light adaptation and function attribute association, which helps to solve the limitations of traditional single brightness matching. First, the depth information in the three-dimensional step model and the underwater ambient light attenuation characteristics are linked. According to the depth of the grid cell, the baseline of brightness mapping is dynamically adjusted. For example, when the water depth of the step area is deep, the baseline brightness is moderately increased to offset the visual darkening caused by underwater light attenuation. When the water depth is shallow, the baseline brightness is reduced to avoid overexposure, ensuring that the texture of the grid cell at different depths can be clearly presented in the robot visual sensor. At the same time, the algorithm associates the flatness data with the functional requirements of the robot cleaning operation. Instead of directly mapping the brightness based on the flatness value, the algorithm adjusts the brightness gradient in combination with the impact of flatness on cleaning. For example, the demand for track adhesion in convex areas and the requirement for cleaning pressure in concave areas that may accumulate water. For high convex areas that are prone to track slip, in addition to assigning a higher brightness, a soft warm color transition is also added. Through the coordinated mapping of luminance and color, non-independent color markers are realized, which intuitively indicate the high friction demand area. For concave areas prone to water accumulation, a cold color tone is added on the basis of low brightness, which simultaneously transmits the functional signal of the need to increase cleaning pressure. In addition, the algorithm also links the real-time update data of the flatness of the subsequent robot scanning. When new subtle convexities or concavities are detected on the step surface corresponding to a grid cell, not only will the brightness value of that cell be adjusted, but also the brightness gradient transition of adjacent cells will be used to simulate the trend of the actual surface flatness, avoiding visual discontinuity caused by texture mutation. The generated texture not only has environmental adaptability and functional guidance, but also more accurately reproduces the real form of the step surface, providing more abundant visual function fusion reference for subsequent robot cleaning path planning and actuator parameter adjustment.

[0030] Finally, in the subsequent scanning operation of the robot, the three-dimensional step model constructed is dynamically modified through the new scanning data fed back by the sensor in real time. For example, when the robot scans a previously uncovered corner area of the step and finds that there is a small concave feature that was not captured by the initial modeling, the feature parameters of the corresponding node will be adjusted according to the new scanning data, and the quadtree algorithm will be triggered to encrypt the grid in that area. At the same time, the concave feature is associated with the corresponding node and stored. If other step areas are found to have convex features in subsequent scanning, the model will also be updated in the same way and the feature information will be stored, ensuring that the three-dimensional step model can continuously adapt to the actual situation of the target step area.

[0031] Further optionally, when the robot acquires new scan data through the sensor and discovers a feature (such as a small indentation in the corner of a step) that is not captured by the initial model, it will first locate the corresponding model region of the feature through spatial coordinate matching. Specifically, the new scan data will carry three-dimensional coordinate information containing the indentation region, which will be compared with the coordinates of each node in the existing three-dimensional step model to find one or more nodes closest to it. If the coverage range of the existing node cannot completely contain the complete region of the indentation, a new node will be added at this position to accurately correspond to the spatial range of the indentation. Subsequently, according to the specific measurement value in the new scan data, the feature parameters of the corresponding node are adjusted. For example, if the indentation depth is a specific value, the flatness quantization value of the node will be updated to reflect this indentation feature, and the associated structure parameters (such as local height value) of the node will be corrected to ensure that the node features are consistent with the physical properties of the actual indentation.

[0032] After completing the adjustment of the node feature parameters, the quadtree algorithm will automatically trigger the encryption processing of the grid in this region. The quadtree algorithm will first identify the location of the indentation region in the existing grid and analyze the feature complexity of the region. Since the indentation is a flatness mutation feature, the algorithm will determine that the region needs higher grid precision to present the details. At this time, the quadtree will split the existing grid cell containing the indentation into four sub-cells, and if the split sub-cells are still insufficient to cover the details of the indentation (such as smaller undulations on the edge of the indentation), the recursive splitting will continue until the size of the grid cell matches the feature size of the indentation. For example, for a small indentation in depth and range, it may only need to be split 1-2 times to obtain a grid with sufficient precision. If the edge of the indentation is irregular and there are secondary protrusions, further splitting will be performed to capture these subtle features. The encrypted grid cell will be associated with the updated and adjusted node feature parameters, so that the grid in this region of the model can not only accurately wrap the outline of the indentation, but also carry more abundant flatness information through the number of dense grid cells, thereby improving the presentation accuracy of the three-dimensional step model in this region and making it more consistent with the actual step surface morphology.

[0033] Step S102, before entering the water for cleaning, an initial control strategy is generated according to the three-dimensional step model by a deep deterministic policy gradient algorithm, and the robot is started into the target step area based on the initial control strategy. In the embodiment of the present application, the initial control strategy at least includes one of the following parameters: track opening angle, driving wheel spacing and propulsion power. In the above step, the initial control strategy is generated based on the three-dimensional step model by the deep deterministic policy gradient algorithm, which can reduce the risk of robot operation caused by blind entry into water. The deep deterministic policy gradient algorithm combines virtual scene pre-training and fault case fine-tuning, so that the initial control strategy can be adapted to the step topological features (such as height difference and inclination angle) and water flow disturbance in advance, greatly reducing the risk of attitude deviation when the robot enters the target area. In particular, the pre-adaptation for high-risk convex edges, such as increasing the track adhesion of the robot, solves the problems of traditional robots, such as jamming and overturning caused by unknown environment at the initial stage of entering water, and improves the safety and efficiency of the water entry stage.

[0034] It can be understood that the initial control strategy is a set of preset motion control scheme optimized and generated by the deep deterministic policy gradient algorithm relying on the structure characteristics of the target step area presented by the three-dimensional step model before the robot enters the water to perform cleaning work. The core purpose is to provide adaptive parameter guidance for the robot to enter the target step area, avoid problems such as jamming, attitude deviation or insufficient power of the robot after entering the water due to unknown underwater step environment, and ensure that the robot can smoothly and safely transition from the initial parking position to the target working area, laying a foundation for the subsequent cleaning process. The track opening angle parameter in the strategy is mainly used to adapt to the height and vertical surface inclination characteristics of the steps in the three-dimensional step model. For example, when the model shows that the vertical surface inclination angle of the step is large, adjusting the track opening angle can increase the contact area of the track and the vertical surface of the step, and improve the adhesion of the robot when climbing the step, avoiding slipping. If the step height is low and the vertical surface is flat, the track opening angle can be adjusted to a more compact state to reduce the moving resistance. The driving wheel spacing parameter is set according to the step width and surface flatness in the model. When the step width is wide, the adaptive driving wheel spacing can enhance the stability of the robot moving on the step surface and prevent left and right deviation. If there is local unevenness on the step surface, reasonable wheel spacing can also reduce the collision between the driving wheel and the uneven part and reduce the actuator loss. The propulsion power parameter is determined in combination with the step distribution density and water flow disturbance in the model. If the steps are densely distributed and need to be frequently turned, the propulsion power will be set to a moderate level to balance the power and attitude stability. If the water flow disturbance is small, the power can be appropriately increased to improve the entry efficiency, and vice versa. The three parameters together constitute the core of the initial control strategy, ensuring that the robot can accurately adapt to the basic environment of the target step area after entering the water.

[0035] As an optional embodiment, in step S102, the initial control strategy is generated according to the three-dimensional step model through the deep deterministic policy gradient algorithm, including: extracting the step height difference, facade inclination angle, surface flatness gradient and spatial connection relationship from the three-dimensional step model, and combining the water flow disturbance coefficient obtained by fluid dynamics simulation to jointly constitute the state space input. The water flow disturbance coefficient is obtained by fluid dynamics simulation based on the pre-collected historical water flow velocity data and historical step space structure data. Further, the track opening angle adjustment range, drive wheel spacing variation interval and propulsion power threshold are set as the continuous action space, the physical constraint model of the actuator is embedded in the Actor-Critic double network architecture, so that the control parameters output by the Actor network meet the mechanical movement limit. Then, the dynamic reward function is set, the high-risk areas identified in the three-dimensional step model are set with a penalty coefficient, and an excessive reward is given to the pre-avoidance action, and the reward weight is adjusted adaptively according to the step complexity. The high-risk areas include the edges of the steps with a convex height greater than the convex obstacle threshold. Then, a virtual reality data fusion training mechanism is used. Specifically, a plurality of virtual scenes are generated by using the three-dimensional step model for pre-training, and the Actor network output precision is optimized by using the deep deterministic policy gradient algorithm. The Critic network dynamically evaluates the action value and corrects the policy deviation in the reverse direction. Subsequently, the trained Actor-Critic double network architecture is applied to the three-dimensional step model to generate an initial control strategy including a track opening angle dynamic adjustment curve, a drive wheel spacing matching scheme and a propulsion power gradient distribution table, and the initial control strategy is pre-checked and corrected in combination with the convex features and concave features in the three-dimensional step model. After generating the initial control strategy, the posture stability of the initial control strategy in the step transition area is verified by model simulation, and if the predicted inclination angle exceeds the safety threshold, the parameter adaptive correction of the initial control strategy is triggered.

[0036] Specifically, in the step of generating the initial control strategy, first, the state space input is constructed. The staff will extract key information reflecting the step structure characteristics from the generated three-dimensional step model, such as the height difference between adjacent steps, the inclination angle of each step facade, the flatness gradient change of the step surface from flat to convex area, and the spatial connection relationship of the step and pool wall, pool bottom connection, and the overall arrangement logic of the step is clear. At the same time, the water flow situation of the step area is simulated by fluid dynamics simulation, such as the local water flow disturbance caused by personnel entering and exiting at the entrance, and the corresponding water flow disturbance coefficient is obtained. These structure information and water flow disturbance coefficient are integrated to constitute the state space input required by the deep deterministic policy gradient algorithm, providing comprehensive environmental data support for subsequent strategy generation.

[0037] Then, the adjustable parameter range of the robot actuator is converted into a continuous action space, i.e. the adjustable interval of the track opening angle, the change range of the drive wheel spacing, and the upper and lower thresholds of the propulsion power. In this process, the physical constraint model of the actuator is embedded in the Actor-Critic double network architecture. For example, according to the mechanical structure of the robot track, the maximum and minimum adjustment limits of the track opening angle are set to avoid the Actor network outputting control parameters beyond the mechanical carrying capacity. For example, if the track opening angle is too large, it may cause the track to break. The constraint model will intercept such unreasonable output in real time to ensure that all control parameters meet the actual movement limits of the robot.

[0038] After that, the dynamic reward function guides the algorithm optimization direction, which considers risk prediction, step passability, and posture stability in three dimensions. Taking the step area with multiple convex edges in the landscape pool as an example, the three-dimensional step model identifies the step edges with excessive convex height as high-risk areas and sets a penalty coefficient for these areas. If the algorithm-generated strategy may cause the robot to approach the area, the corresponding reward will be deducted. Conversely, if the strategy includes adjusting the track grip in advance to avoid high-risk areas, an excess reward will be given. At the same time, the reward weight will be adjusted adaptively according to the step complexity. For example, for step areas with complex structure and large flatness difference, the reward weight of posture stability dimension will be increased to guide the algorithm to generate control strategies that can ensure the stability of the robot.

[0039] Subsequently, a virtual reality data fusion training mechanism is used to train the algorithm. The three-dimensional step model generates diverse virtual training scenes, such as simulating the straight steps of public pools, the arc-shaped steps of landscape pools, and the staggered steps of private pools. The algorithm is pre-trained in virtual scenes to familiarize with the control requirements of various step structures. After pre-training, historical cleaning data fault cases are imported, such as the case where the robot was stuck in a certain step concave area due to improper drive wheel spacing. The algorithm parameters are fine-tuned, and the output accuracy of the Actor network is continuously optimized through the deep deterministic policy gradient algorithm to make the control parameters generated by the Actor network more suitable for actual work requirements. At the same time, the Critic network evaluates the value of the Actor network output action in real time. If it finds that a certain action causes the robot posture to deviate, it will correct the strategy bias to ensure that the algorithm output strategy is efficient and safe.

[0040] After the training, the Actor-Critic dual network architecture is applied to the three-dimensional model of the current target step area to generate an initial control strategy. This strategy is presented in a structured form, including a dynamic adjustment curve of the track angle with respect to the step height, a drive wheel spacing matching scheme for different width steps, and a propulsion power gradient distribution table based on water flow disturbance and step slope. After generating the strategy, pre-checking and correction are performed in combination with the convex and concave features of the steps in the three-dimensional model. For example, if the model shows that a certain section of the step has continuous concave, it is checked whether the drive wheel spacing matching scheme can allow the robot to pass smoothly. If the spacing is too large, it may cause the robot to sink into the concave, so the drive wheel spacing is appropriately reduced to optimize the passability.

[0041] Finally, the effectiveness of the initial control strategy is verified through model simulation, focusing on simulating the motion state of the robot in the step transition area, such as the transition process from a one-step step to a two-step step. The simulation predicts the inclination angle of the robot during this process. If the predicted inclination angle exceeds the safety threshold, parameter adaptive correction is triggered immediately, such as adjusting the power value in the propulsion power distribution table during the climbing phase or optimizing the slope of the track angle adjustment curve, until the simulation results show that the robot's posture in the step transition area is stable and the inclination angle is controlled within the safety range. Finally, the initial control strategy that meets the actual work requirements is determined.

[0042] Exemplarily, in the above steps, the information extracted from the three-dimensional step model covers more detailed step structure features when constructing the state space input, combined with the water flow data from the fluid dynamics simulation to form a multi-dimensional input. For example, in the standard step area of a public swimming pool, the uniform height difference between adjacent steps (such as a sequence of steps with consistent height) and the irregularly distributed step height difference in a private swimming pool (such as non-uniform arrangement of some high steps and some low steps) are extracted. For the inclination angle of the facade, both the vertical facade perpendicular to the pool bottom and the gentle slope facade (smaller inclination angle) and the arc-shaped transition facade (continuous change in angle) designed to create a natural form in a landscape pool are included. The extraction of the surface flatness gradient is refined to the characteristics of steps of different materials, such as the regular protruding gradient of the step surface formed by the joints of the ceramic tile paving, the local wear and tear gradient of the concrete step after long-term use, and the irregular damage gradient of the step edge due to collision. The extraction of the spatial connection relationship covers more spatial correlation forms, such as the right-angle hard connection of the step and the pool wall, the smooth transition connection of the step and the shallow water area platform, and the insertion connection of some steps in the step group and the hidden drain. The water flow disturbance coefficient obtained by fluid dynamics simulation also includes diversified scenarios, such as the intermittent strong disturbance of the entrance step area of the swimming pool due to frequent entry and exit of personnel, the persistent vortex disturbance of the deep water area step due to water flow circulation, and the weak disturbance of the step corner due to water flow stagnation. These more detailed structure features and water flow data together constitute the state space input, allowing the deep deterministic policy gradient algorithm to better understand the environmental characteristics of the target step area and provide more abundant basic information for generating an accurate and adaptive initial control strategy.

[0043] As an optional embodiment, in step S102, starting the robot into the target step area based on the initial control strategy comprises: based on the surface flatness gradient of the entrance area in the three-dimensional step model and the water flow disturbance coefficient, screening a safe entrance point position meeting the preset safe entrance condition to generate a path planning from the initial parking position of the robot to the entrance point position. The preset safe entrance condition comprises: the water flow disturbance coefficient is lower than a set low disturbance threshold, and the surface flatness gradient meets a set convex obstacle threshold. Then, the robot actuator pre-checking is started to drive the track to adjust to the initial value of the opening angle in the initial control strategy, and drive the wheel to match to the interval reference value. Whether the actuator load meets the physical constraint is detected by the torque sensor. If the load is over the threshold, the corresponding parameter is fine-tuned. Then, the robot is controlled to enter the water and enter the path execution stage. The water flow velocity data is collected in real time. The water flow disturbance deviation value is calculated from the real-time collected water flow velocity data. The propulsion power is dynamically adjusted according to the propulsion power gradient distribution table of the initial control strategy combined with the water flow disturbance deviation value to avoid the attitude deviation caused by water flow impact. When the distance from the entrance point position is less than a set distance, the pre-adaptation of the track opening angle dynamic adjustment curve is triggered in advance based on the convex feature and the concave feature in the three-dimensional step model. If the robot is in the step edge with the convex height greater than the convex obstacle threshold in the entrance area, the track gripping force is increased. After the robot enters the initial section of the target step area, the real-time motion state data and the attitude data are collected and compared with the expected structure parameters of the three-dimensional step model. If the attitude deviation is less than or equal to the preset threshold, it is confirmed to be stably entered. If the attitude deviation is greater than the preset threshold, the preliminary attitude correction is performed based on the inverted pendulum control model.

[0044] Specifically, in step S102, in the process of starting the robot into the target step area based on the initial control strategy, the safe entrance path needs to be determined first. It is necessary to analyze whether the surface flatness gradient of the entrance area in the three-dimensional step model is lower than the set convex obstacle threshold to determine whether there is a convex obstacle in the entrance area. In addition, the spatial connection relationship between the steps and the pool edge at the entrance of the public swimming pool, the arrangement density of the step group can be combined to determine whether there is a convex obstacle in the entrance area. At the same time, the water flow disturbance coefficient obtained by fluid dynamics simulation is combined to screen the area with smooth water flow and no obvious convex obstacle as the safe entrance point position. For example, in a private swimming pool, if the model shows that the water flow disturbance of a certain step entrance near the shallow water area is small, and there is no convexity on the step edge, this position will be set as the entrance point. Then, the path planning from the pool edge platform where the robot initially stops to the entrance point is generated. The path will avoid the concave area and the high water flow disturbance zone marked in the model to ensure that the robot can approach the target area smoothly during movement.

[0045] Next, the robot's actuator is pre-verified to ensure that all components meet the initial control strategy settings before entering the water. The track is driven to the initial angle specified in the initial control strategy, such as a medium angle for a flat step, and the drive wheels are adjusted to the reference value of the step width. Then, the torque sensor detects the load of the track and drive wheels. If the track angle is too large, causing the drive motor to exceed the safety threshold, the angle is slightly reduced until the load meets the physical constraints. If the drive wheel spacing is too wide, causing the axle to be abnormally stressed, the spacing is appropriately narrowed to ensure that the actuator is in a stable working state before entering the water. After completing the pre-verification, the robot slowly enters the water and enters the path execution phase. During the water entry process, the water flow sensor carried by the robot collects real-time water flow data. Then, based on the initial control strategy's propulsion power gradient distribution table and real-time water flow disturbance, the propulsion power is dynamically adjusted. For example, in a landscape pool, if a sudden disturbance is detected near the entrance due to landscaping water flow, the propulsion power is immediately reduced to reduce the impact of water flow on the robot's attitude. If the water flow is smooth, the medium power is maintained according to the distribution table to ensure moving efficiency while avoiding attitude deviation.

[0046] When the robot is within a set distance from the entry point, the convex feature and concave feature data in the three-dimensional step model for that area are pre-adapted to trigger the dynamic adjustment curve of the track angle. For example, if the model shows that the edge of the entrance area has a high-risk part with a convex height greater than the convex obstacle threshold, the track angle is appropriately expanded to increase the contact area between the track and the step surface, thereby increasing the gripping force and preventing the robot from slipping when entering due to the edge convexity. For slight concave areas in the entrance area, the track angle is slightly adjusted to make the track more conform to the concave surface, ensuring smooth transition.

[0047] After the robot enters the initial section of the target step area, it immediately starts collecting motion state and attitude data, including moving speed, acceleration, and inclination angle. These real-time data are compared with the expected structure parameters (such as step height and surface flatness) of the three-dimensional step model for that section. If the comparison result shows that the robot's attitude deviation is within the preset range, such as a slight deviation from the horizontal baseline, it is confirmed that the robot has entered a stable state. If the deviation exceeds the threshold, such as a significant inclination caused by a sudden convexity on the step surface, an algorithm based on the inverted pendulum balance theory is immediately enabled for preliminary attitude correction. By adjusting the drive wheel spacing and propulsion power distribution, the robot quickly recovers to a stable attitude, laying the foundation for subsequent cleaning operations.

[0048] Step S103, during the cleaning process, the motion state data and the attitude monitoring data of the robot are collected in real time. In the embodiment of the present application, the motion state data of the robot during the cleaning process refers to a parameter set reflecting the dynamic behavior of the robot during the movement in the underwater step area, including the moving speed of the robot (such as the speed of advancing or retreating along the step surface horizontally, the speed of turning around the axis of the robot), the acceleration (such as the speed of increasing when starting, the amplitude of deceleration when encountering a convex feature), the moving direction (such as whether to travel along the preset cleaning path, whether to deviate from the direction due to water flow or step obstacle), and the operating parameters of the actuator (such as the rotation rate of the track, the rotation speed of the driving wheel, the working speed of the cleaning brush, etc.). These data can directly reflect the moving efficiency of the robot and the working state of the actuator, for example, a sudden decrease in the rotation speed of the track may mean that the robot encounters a convex obstacle on the step surface, and a slowdown in the advancing speed may be affected by water flow disturbance.

[0049] The attitude monitoring data is a parameter for capturing the attitude state of the robot in three-dimensional space, mainly including the inclination angle of the robot (such as the up-down inclination angle of the front end relative to the step surface when climbing the step, the inclination angle of the left and right sides when approaching the edge of the step), the angular velocity (such as the speed of changing the inclination angle, such as the rate of rapidly increasing the inclination angle when suddenly driving into the concave area of the step), the attitude stability parameter (such as whether the robot maintains a horizontal attitude, whether there is a shaking frequency caused by water flow impact or step structure), and the contact state data between the robot and the step surface (such as the contact pressure of the track on the step surface, the adhesion of the driving wheel to the step, etc.). These data are directly related to the attitude safety of the robot, for example, an excessively large inclination angle may cause the robot to roll over, and insufficient contact pressure may mean that the track has poor adhesion to the step surface and there is a risk of slipping.

[0050] During real-time data collection, these motion state data and attitude monitoring data are continuously transmitted to the robot control end, establishing a dynamic feedback channel between the state of the robot itself and the underwater step environment. The collected data are processed by spatio-temporal alignment to correspondingly associate the motion data and the attitude data at the same time and the same spatial position, such as matching the advancing speed at a certain time with the inclination angle at that time, to ensure that the data accurately reflect the corresponding relationship between the motion behavior and the attitude change. At the same time, through an abnormal data elimination algorithm, invalid data (such as instantaneous speed abnormal peak value caused by water flow impact, inclination angle error value caused by sensor false triggering) caused by water flow disturbance to underwater sensors and signal fluctuation are filtered out, ensuring the timeliness and accuracy of the data. This accurate real-time data feedback can enable the subsequent adaptive control model to timely capture the attitude change (such as sudden inclination) of the robot and the environmental dynamics (such as sudden water flow and unexpected step obstacle), providing a reliable basis for the dynamic optimization of the control strategy, and effectively alleviating the problem of large sensor data noise and multiple disturbances in the underwater environment.

[0051] In the above steps, real-time data collection during the robot cleaning process establishes a dynamic feedback channel for the robot's own posture and the correlation with the environment, improving the real-time performance of the robot's running state monitoring. Through time-space alignment and abnormal data rejection, the timeliness and accuracy of the data are ensured, which helps to solve the problem of underwater sensors being easily disturbed by water flow and large data noise, so that the subsequent control strategy optimization can accurately respond to changes in the robot's posture and the environment.

[0052] In step S104, an adaptive control model based on the inverted pendulum balance theory is used to cooperatively optimize the real-time collected motion state data, posture monitoring data, and step structure parameters in the three-dimensional step model, to obtain the real-time control strategy of the robot.

[0053] The step structure parameters are a set of key information extracted from the three-dimensional step model that reflect the physical characteristics of the target step area, including the height of each step and the height difference between adjacent steps, the width of the step surface, the inclination angle of the vertical surface, the flatness gradient change of the step surface from flat to convex or concave, and the spatial connection relationship between steps (such as whether there is a smooth transition or a right-angle connection), the specific location and size of the surface convexity and concavity, etc. These parameters collectively outline the spatial form and surface state of the steps and are the core basis for the robot to understand the structure of the environment, such as a larger inclination angle of the vertical surface of a step indicating that the robot needs stronger adhesion when climbing, and a large flatness gradient indicating the presence of more convexities or concavities, which requires frequent adjustment of the posture to avoid slipping.

[0054] The principle of the synergistic optimization in the above steps can be understood as establishing a correlation mapping between the motion state data, the posture monitoring data, and the step structure parameters, so that the three complement each other and dynamically adapt in the control strategy generation process. Specifically, the motion state data (such as speed and acceleration) reflects the current movement efficiency and actuator working state of the robot, the posture monitoring data (such as inclination angle and angular velocity) embodies the balance stability of the robot, and the step structure parameters provide prior information of the environment (such as the height difference of the upcoming step and the surface flatness), and the combination of the three achieves spatiotemporal alignment. Further, the motion data, the posture data at the same time and the same position, and the step structure parameters corresponding to the position are associated, input into the adaptive control model, the model will predict the environmental change (such as entering a high protrusion area) according to the step structure parameters, judge whether the current movement is adapted to the environment (such as whether the speed is too fast) in combination with the motion state, and evaluate the balance state (such as whether the inclination is caused by the speed being too fast) through the posture data, and finally adjust the control parameters by comprehensively considering the three information to ensure that the robot adapts to the environment while maintaining stability. For example, when the step structure parameters show that there is a step with a large height difference in front, the motion state data shows that the current speed is fast, and the posture data shows that there is a tendency to lean forward, the model will synergistically optimize and simultaneously reduce the propulsion power, adjust the track angle to increase the gripping force, to achieve smooth transition.

[0055] In the embodiment of the present application, the adaptive control model is constructed by a sliding mode control, a fuzzy control and an inverted pendulum control model. It is worth noting that the adaptive control model based on the inverted pendulum balance theory is to equivalent the motion posture of the robot on the step to the balance problem of the inverted pendulum. For example, the main body of the robot is regarded as a pendulum rod, and the step surface is regarded as the support surface of the pendulum rod. By maintaining the vertical balance of the pendulum rod (robot) relative to the support surface (step), the stable posture of the robot is ensured. The model structure integrates the inverted pendulum control model, the sliding mode control and the fuzzy control: the inverted pendulum control model as the core, the dynamic equation is established through the inverted pendulum balance theory, the attitude data such as inclination angle and angular velocity is converted into the control quantity (such as track angle adjustment direction) required for maintaining balance. The sliding mode control module is responsible for processing the uncertain disturbances such as water flow disturbance and uneven step, and setting the sliding mode to quickly respond to disturbances and reduce attitude fluctuations. The fuzzy control module is based on expert experience and historical data to perform fuzzy reasoning on the step complexity (such as height difference and protrusion density) and the degree of attitude deviation, dynamically optimizes the balance parameters of the inverted pendulum control model and the gain coefficient of the sliding mode control, and improves the adaptability of the model to complex environments. The function of the model is to receive three types of data of motion state, attitude monitoring and step structure in real time, output control parameters that can accurately adapt to the current environment through multi-module collaborative calculation, ensure the cleaning efficiency of the robot in the flat step area, and maintain the attitude stability in complex areas such as high-risk protrusions and large-angle vertical surfaces, and solve the problem of insufficient adaptability of traditional control methods in underwater multi-disturbance environment.

[0056] In the embodiment of the present application, the real-time control strategy of the robot mainly contains specific control parameters for the current environment and the state of the robot. For example, the real-time correction value of the track angle (based on the current step vertical surface inclination angle and the attitude inclination angle), the dynamic adjustment amount of the driving wheel spacing (adapted to the current step width and surface flatness), the adaptive value of the propulsion power (balance cleaning efficiency and attitude stability), and the adjustment frequency of each parameter (set according to the step complexity). It can be understood that the types of real-time control strategy are divided according to different functional emphases, including but not limited to: attitude stability strategy (when the inclination angle approaches the safety threshold, the track angle and the driving wheel spacing are adjusted to reduce the inclination risk), obstacle avoidance strategy (when a high protrusion area is detected, the propulsion power is reduced and the track adhesion is increased), efficiency optimization strategy (in a flat and undisturbed area, the propulsion power is increased and the attitude control threshold is relaxed to improve the cleaning speed), transition adaptation strategy (in the transition area with large step height difference, the power and track parameters are adjusted simultaneously to realize smooth climbing or descending), and the like. These strategies will be dynamically switched according to real-time data to ensure that the robot can efficiently and safely complete the cleaning work in different scenarios.

[0057] In the above steps, the multi-disturbance problem of the underwater step environment is improved by the fusion of inverted pendulum balancing theory, sliding mode control and fuzzy control. Specifically, the inverted pendulum control model can equivalently convert the robot posture into a pendulum balancing problem, achieving accurate quantification of the tilt state. Sliding mode control can enhance the robustness to water flow disturbance and uneven steps. Fuzzy control helps to dynamically optimize control parameters through fuzzy reasoning of step complexity and posture deviation. The synergistic optimization of inverted pendulum balancing theory, sliding mode control and fuzzy control enables the real-time control strategy to adapt to different step structures, such as focusing on efficiency in flat areas and stability in convex areas, helping to solve the poor adaptability and posture instability problems of traditional control methods in complex environments, control the robot tilt angle within a safe threshold, and ensure the stability of the cleaning process.

[0058] It should be noted that sliding mode control is a robust control method for nonlinearity, its core idea is to design a specific control strategy to force the state trajectory to reach and remain on the preset sliding mode surface in a limited time, thereby achieving insensitivity to parameter changes and external disturbances. Specifically, first, a sliding surface is set according to the control objective, which is usually a linear combination of state variables representing the ideal motion trajectory. Then, through the switching control law, the state converges to the sliding surface from any initial position, and once it reaches the sliding surface, it will move stably along the sliding surface, even if there are external disturbances or parameter fluctuations, the stable state can be maintained. For example, in robotic arm control, sliding mode control can offset the effects of load changes by rapidly switching control signals, ensuring that the robotic arm moves along the predetermined trajectory. This strong robustness makes it particularly suitable for complex environments with uncertain disturbances.

[0059] Fuzzy control is an intelligent control method based on fuzzy logic, its principle is to simulate the fuzzy thinking in human decision-making process, convert precise input quantities into fuzzy sets (such as high temperature, slow speed, etc. fuzzy concepts) through fuzzification, then infer according to the preset fuzzy rule base (if expert experience or experimental data summary rules such as that), finally convert the inference results into precise control quantities through defuzzification, realize the control. It does not need to establish an accurate mathematical model, especially suitable for nonlinear, time-varying systems that are difficult to model accurately. For example, in room temperature control, fuzzy control can output specific power adjustment values according to the fuzzy information "current temperature and target temperature deviation is large and changes fast", and the rule "if the deviation is large and changes fast, then adjust the heating power significantly", to flexibly cope with the complex changes of environmental temperature.

[0060] The inverted pendulum control model takes the inverted pendulum as the control object. The core principle is to monitor the tilt angle, angular velocity of the pendulum rod, and the position, velocity and other state information of the base in real time, design a feedback control strategy, and dynamically adjust the movement of the base (such as horizontal force or displacement) to offset the tilting tendency of the pendulum rod due to gravity, and maintain the upright balance state of the pendulum rod. The inverted pendulum itself is a typical unstable, strongly coupled nonlinear system. The key to control is to establish the dynamic relationship between the pendulum rod and the base, and to correct the control quantity in real time through negative feedback mechanism. When the pendulum rod tilts forward, control the base to move forward to generate inertial force, so that the pendulum rod swings backward. When the pendulum rod tilts backward, control the base to move backward, so as to achieve dynamic balance. This control idea is applied to scenes that need to maintain dynamic balance, such as robot gait control, and its essence is to stabilize the unstable state in the target state through fast response to posture changes.

[0061] As an optional embodiment, in step S104, the speed and acceleration in the real-time motion state data, the tilt angle and angular velocity in the posture monitoring data, the step height difference in the step structure parameter extracted from the three-dimensional step model, the surface flatness gradient, the convex feature, and the concave feature are spatiotemporally aligned, the abnormal data is removed by fusing the water flow disturbance coefficient, and an input data set is generated. Further, a robot dynamics model is established based on the inverted pendulum balance theory as an inverted pendulum control model. Taking the tilt angle in the input data set as a balance deviation reference, the balance torque required for the robot to maintain the posture balance is calculated by the robot dynamics model in combination with the step height difference, and the adjustment direction of the track opening angle and the drive wheel spacing is generated based on the balance torque. Then, the water flow disturbance coefficient and the uncertainty disturbance coefficient caused by the unevenness of the steps are obtained by the sliding mode control module, and the sliding mode switching gain is dynamically adjusted according to the water flow disturbance coefficient, the uncertainty disturbance coefficient and the flatness gradient. The fuzzy inference of the step complexity and the posture deviation degree is performed by using the fuzzy control module, and the balance parameters of the inverted pendulum control model and the gain coefficient of the sliding mode control are adaptively optimized, and the balance weight of the high-risk area is increased. The step complexity is determined by the height difference and the convex density. Finally, the adjustment direction of the track opening angle and the drive wheel spacing, the sliding mode switching gain, the balance parameter, the gain coefficient, and the balance weight are cooperatively integrated, and the mapping relationship between the propulsion power and the posture stability is combined to generate a real-time control strategy including the real-time correction value of the track opening angle, the dynamic adjustment amount of the drive wheel spacing, and the adaptive value of the propulsion power.

[0062] Exemplarily, first, the real-time collected motion state data (such as the speed of the robot moving along the step surface, the deceleration acceleration when encountering a protrusion), the attitude monitoring data (such as the inclination angle of the front end when climbing the steps, the angular velocity when turning), and the corresponding position step structure parameters (such as the height difference between the current step and the next step, the flatness gradient of the step surface from flat to concave, the position characteristics of the local protrusion, and the size characteristics of the corner concave) extracted from the three-dimensional step model are precisely matched according to the collection time and spatial coordinates, ensuring that the same time and the same position of the three types of data form a corresponding association. At the same time, the water flow disturbance coefficient obtained by fluid dynamics simulation is used to determine whether the real-time collected water flow velocity data is affected by abnormal water flow. For example, in the entrance step area of a public swimming pool, if the water flow sensor detects a sudden strong disturbance, the water flow velocity data collected at this time will have a sudden peak value, which will be determined as an abnormal value and excluded, providing a reliable data basis for subsequent model calculation.

[0063] Then, based on the inverted pendulum balance theory, a robot dynamics model is constructed, and the attitude of the robot on the steps is equivalent to the balance state of the inverted pendulum. The model takes the inclination angle in the input data set as the balance deviation reference. For example, when the robot appears to be tilted forward on the curved steps of a landscape swimming pool, the height difference between the position step and the next step is combined to calculate the balance torque required to maintain the attitude balance through the dynamics equation. The greater the height difference, the greater the torque required to maintain balance. Subsequently, the balance torque is converted into the adjustment direction of the specific actuator: if the forward torque is large, an adjustment direction to increase the track angle will be generated (to increase the contact area of the track with the step surface and enhance the gripping force to offset the forward trend), and at the same time, an adjustment direction to reduce the distance between the drive wheels will be generated (to reduce the risk of lateral tilting of the robot and improve overall stability), ensuring that the adjustment direction directly matches the balance requirement.

[0064] After that, a sliding mode control module is started to handle uncertain disturbances, and the uncertain disturbance coefficient is set based on historical step spatial structure data. The sliding mode control module will dynamically adjust the sliding mode switching gain according to the surface flatness gradient in the input data set. For example, in the concrete step area of a private swimming pool, if the flatness gradient is large (there are many small protrusions and concavities on the surface due to wear), it means that the robot is easily disturbed by the unevenness of the steps during movement. At this time, the sliding mode switching gain will be appropriately increased to enable a faster response to attitude changes. When the robot appears to be momentarily tilted due to a small protrusion, the sliding mode control can quickly output an adjustment signal to prevent the inclination angle from further expanding. If the step surface is flat (the flatness gradient is small), the switching gain is appropriately reduced to reduce the frequent fluctuations of the control parameters, balancing stability and control efficiency.

[0065] Meanwhile, the fuzzy control module will perform fuzzy reasoning on the step complexity and the degree of attitude deviation. The step complexity is determined by the height difference of the current area and the density of the protrusions, such as the continuous multi-step height difference in a public swimming pool and the dense protrusions on the edge of the step area, which will be determined as a high complexity area. The degree of attitude deviation is divided according to the difference between the inclination angle and the safety threshold, such as slight inclination, moderate inclination, and severe inclination. The module dynamically adjusts the balance parameters of the inverted pendulum control model (such as increasing the recovery coefficient to enhance the anti-inclination ability) and the gain coefficient of the sliding mode control according to the preset fuzzy rules (such as high complexity area and moderate inclination to greatly optimize the balance parameters). For high-risk areas identified in the three-dimensional model (such as parts with protrusion height exceeding the standard on the edge of the step), additional balance weight will be added. When calculating the control parameters, more control resources are allocated to the attitude adjustment of the area corresponding to the area, for example, the correction amplitude of the track angle is further increased to ensure the stability of the robot when passing through the high-risk area.

[0066] Finally, the track angle adjustment direction, the drive wheel spacing adjustment direction, the sliding mode switching gain, the optimized balance parameters and gain coefficients, and the balance weight of the high-risk area are calculated, and the mapping relationship between the propulsion power and the attitude stability is combined. If the current inclination angle is small (attitude stable), the propulsion power adaptive value will be set according to the cleaning efficiency priority principle. If the inclination angle is close to the safety threshold (attitude risk increases), the propulsion power adaptive value will be reduced to reduce the moving speed and reduce the attitude fluctuation. For example, when the robot climbs the high step of the landscape swimming pool, the input data set shows that there is a moderate forward inclination angle, a large step height difference (high complexity), and a small protrusion on the surface (flatness gradient medium), which will integrate to generate a real-time control strategy of "increasing the real-time correction value of the track angle, reducing the dynamic adjustment amount of the drive wheel spacing, and reducing the propulsion power adaptive value". The real-time control strategy not only offsets the forward trend through the adjustment of the track and the wheel spacing, but also reduces the power to avoid excessive speed and exacerbate instability, ensuring stable cleaning of the robot in the complex step area.

[0067] Further optionally, in the above embodiment, the robot dynamics model is established based on the inverted pendulum balance theory as an inverted pendulum control model, comprising: first, determining the input parameters of the robot dynamics model, including the physical parameters of the robot, real-time attitude monitoring data, three-dimensional step model extracted structural parameters, and water flow disturbance coefficient. The physical parameters of the robot include rotational inertia, damping coefficient and restoring coefficient, the real-time attitude monitoring data include inclination angle and angular velocity, the three-dimensional step model structural parameters include step height difference and surface flatness gradient, and the water flow disturbance coefficient is obtained by fluid dynamics simulation. Further, the cleaning state of the robot is equivalent to an inverted pendulum device, the inclination angle is mapped to the angle of the inverted pendulum deviating from the balance position, the angular velocity is mapped to the angular rate of the inverted pendulum, the step height difference is converted to the external tilting disturbance torque of the inverted pendulum, the surface flatness gradient is converted to the contact damping disturbance of the inverted pendulum, and the water flow disturbance coefficient is converted to the lateral additional torque of the inverted pendulum, to construct an inverted pendulum dynamics equation containing multiple disturbance terms. Next, a cleaning component contact reaction force correction term is generated based on the inverted pendulum dynamics equation, which is calculated by the friction coefficient between the robot brush and the step surface and the real-time pressure sensor data, and is used to compensate for the additional force of the brush on the robot attitude during cleaning. Then, the damping coefficient in the inverted pendulum dynamics equation is dynamically adjusted according to the flatness gradient of the step surface in the three-dimensional step model. When the flatness gradient increases, the damping coefficient is increased to enhance the attitude disturbance resistance, and when the flatness gradient decreases, the damping coefficient is reduced to improve the attitude adjustment flexibility. Finally, the parameters of the inverted pendulum dynamics equation are iteratively corrected by the attitude monitoring data and the motion state data of the robot collected in the historical cleaning process, the deviation between the balance torque output by the robot dynamics model and the actual required balance torque is calculated, the gradient descent algorithm is used to optimize the rotational inertia and the restoring coefficient, so that the model prediction deviation of the robot dynamics model is controlled within a preset range, and a robot dynamics model adapted to the pool step scene is constructed as an inverted pendulum control model.

[0068] Specifically, in the above embodiment, the input parameters of the model need to be first determined and the collection and integration of the parameters are completed. Among the physical parameters of the robot, the moment of inertia is related to the overall mass distribution of the robot. For example, a robot with a heavy body and a battery at the bottom has a relatively large moment of inertia, which affects the sensitivity of attitude adjustment. The damping coefficient reflects the resistance of the robot when the attitude changes, which is related to the friction characteristics of the track material and the step surface. The recovery coefficient is related to the attitude correction ability of the robot itself, which is determined by the performance of the built-in balance adjustment mechanism. Real-time attitude monitoring data is obtained through the gyroscope and tilt sensor carried by the robot. For example, when climbing a step, the tilt angle represents the angle between the body and the step surface, and the angular velocity reflects the speed of change of the angle. The structural parameters of the three-dimensional step model are extracted from the constructed model, such as the height difference between the current step and the next step, and the flatness gradient change from flat to convex surface. The water flow disturbance coefficient is obtained through fluid dynamics simulation, such as the periodic water flow at the entrance of a public swimming pool due to the flow of personnel, which has a regular fluctuation in the disturbance coefficient. These parameters together constitute the input basis of the model, ensuring that the model can fully reflect the physical characteristics, real-time state, and environmental characteristics of the robot.

[0069] Then, the cleaning operation state of the robot is equivalent to an inverted pendulum device, and parameter mapping and dynamics equation construction are completed. Specifically, the body of the robot is regarded as the swing rod of the inverted pendulum, the step surface is regarded as the support surface of the swing rod, the tilt angle of the robot directly corresponds to the angle of the swing rod deviating from the vertical balance position, and the angular velocity corresponds to the angular rate of the swing rod swinging. For the disturbance factors in the environment, the step height difference is converted into an external tilt disturbance torque acting on the swing rod. For example, a step with a large height difference in a landscape swimming pool will generate a large disturbance torque, simulating the tendency of the swing rod being pushed to the tilt. The surface flatness gradient is converted into a damping disturbance of the swing rod and the support surface, and a rough surface (large flatness gradient) will increase the damping, making the swing rod swing more slowly, and a smooth surface will reduce the damping. The water flow disturbance coefficient is converted into a lateral additional torque, simulating the force that causes the swing rod to sway left and right due to water flow impact. Through this mapping, an inverted pendulum dynamics equation containing multiple disturbance terms is constructed, so that the equation can reflect the comprehensive influence of the robot's own attitude and environmental disturbance.

[0070] Subsequently, a cleaning component contact reaction force correction term is generated to compensate for the impact of brush operation on the attitude. When the robot is cleaning, the brush is in contact with the surface of the step, which generates friction and pressure, and these forces will have additional effects on the attitude of the robot. For example, when the brush is rotating at high speed to clean the ceramic tile step, if it encounters a thick surface stain, the friction between the brush and the step will increase, which may cause the robot to tilt slightly backward. At this time, the real-time pressure data of the brush and the step is collected through the pressure sensor, and the friction coefficient between the brush material and the step surface (such as the friction coefficient between the nylon brush and the ceramic tile) is combined to calculate the size and direction of the contact reaction force, and then a correction term is generated and added to the dynamic equation to offset the interference of the brush force on the attitude balance, ensuring that the balance torque calculated by the model is more accurate.

[0071] Subsequently, the damping coefficient in the dynamic equation is dynamically adjusted according to the flatness gradient of the step surface. When the three-dimensional step model shows that the flatness gradient of the current area increases, it means that there are more protrusions or depressions on the surface of the step (such as the worn area of the concrete step after long-term use), at which time the damping coefficient will be increased to make the attitude change of the robot more gentle and enhance the anti-disturbance ability to the uneven surface. For example, when encountering a sudden protrusion, the model with a large damping coefficient will slow down the inclination speed of the body, facilitating timely adjustment. When the flatness gradient decreases (such as a smooth ceramic tile step), the damping coefficient is reduced to make the attitude adjustment of the robot more flexible, enabling it to quickly respond to subtle changes in inclination and balance the anti-disturbance and adjustment efficiency.

[0072] Finally, the model parameters are iteratively corrected through historical data and real-time data to improve the scene adaptability of the model. A large amount of attitude monitoring data collected during the historical cleaning process (such as the inclination angle change and angular velocity curve on a similar height difference step in the past) is compared with the real-time motion state data of the current robot (such as the current speed and acceleration) to calculate the deviation between the balance torque output by the model and the actual required balance torque. For example, the model predicts that 10 units of balance torque are required, but only 8 units are actually needed to maintain balance, indicating that there is a deviation. At this time, the gradient descent algorithm is used to optimize the moment of inertia and recovery coefficient of the robot to gradually reduce the deviation until the model prediction deviation is controlled within the preset range. For example, for the irregularly distributed steps in a private swimming pool, the model can accurately predict the balance torque required at different height differences through multiple iterative corrections, and finally a robot dynamics model fully adapted to the swimming pool step scene is constructed, providing a precise theoretical basis for the generation of subsequent real-time control strategies.

[0073] Further optionally, the real-time control strategy is compared with the expected step structure of the three-dimensional step model to obtain a predicted attitude deviation value of the robot. If the predicted attitude deviation value is greater than a safety threshold, a secondary optimization is triggered for the real-time control strategy until the output satisfies the stable control condition of the current step cleaning scene.

[0074] Specifically, in actual execution, the control device will first compare the generated real-time control strategy with the expected step structure of the current cleaning area in the three-dimensional step model, and obtain the predicted attitude deviation value of the robot under the guidance of the strategy through simulation calculation. Taking the multi-step area at the entrance of a public swimming pool as an example, the three-dimensional step model clearly labels that there is a large height difference between the second step and the third step, and there is a gentle convexity on the edge of the third step. The real-time control strategy currently set the track opening angle to be the regular value suitable for the gentle step, and the propulsion power is at a medium level. Through comparison and simulation, it can be known that when the robot climbs from the second step to the third step according to this strategy, the regular track opening angle is difficult to fully fit the vertical surface of the large height difference, and the medium propulsion power may cause the climbing speed to be too fast, thereby predicting that the robot will have a significant forward attitude deviation, and the deviation value is greater than the preset safety threshold, which has a risk of attitude instability.

[0075] In this case, the device will trigger secondary optimization for the real-time control strategy. During the optimization process, the parameters that cause the deviation to exceed the standard will be adjusted. First, according to the characteristics of the large height difference of the steps in the three-dimensional model, the real-time correction value of the track opening angle is appropriately increased to improve the contact area between the track and the step vertical surface, enhance the climbing grip, and offset the forward inclination trend. At the same time, combined with the convexity characteristics of the step edge, the propulsion power adaptation value is appropriately adjusted downward to reduce the climbing speed and avoid collisions or exacerbate attitude deviation due to excessive speed and convexity. After optimization, the new control strategy is compared with the expected structure of the three-dimensional step model again, and the predicted attitude deviation value is recalculated. If the predicted deviation value is still slightly higher than the safety threshold, the dynamic adjustment amount of the drive wheel spacing will be further fine-tuned to reduce the wheel spacing to improve the lateral stability of the robot and assist in offsetting the forward inclination risk; until after multiple fine-tuning, the predicted attitude deviation value corresponding to the new strategy is reduced to within the safety threshold, and can adapt to the structural characteristics of the step height difference and the edge convexity, and meet the stable control conditions of the current step cleaning scene, the final optimized real-time control strategy is output to the execution mechanism, ensuring that the robot completes the cleaning work in the area with a stable attitude.

[0076] Further optionally, in the above embodiments, the adjustment direction of the track opening angle, the adjustment direction of the driving wheel spacing, the switching gain of the sliding mode, the balance parameter, the gain coefficient, and the balance weight are integrated together, and the mapping relationship between the propulsion power and the attitude stability is combined to generate a real-time control strategy including the real-time correction value of the track opening angle, the dynamic adjustment amount of the driving wheel spacing, and the propulsion power adaptation value, including: establishing a collaborative correlation rule of each parameter, binding the adjustment direction of the track opening angle output by the inverted pendulum control model with the switching gain of the sliding mode, and when the switching gain of the sliding mode is increased due to the increase of the flatness gradient, the correction amplitude of the adjustment direction of the track opening angle is correspondingly expanded; the adjustment direction of the driving wheel spacing is bound with the balance parameter and the balance weight, and when the balance weight is increased in the high-risk area, the single adjustment amount of the adjustment direction of the driving wheel spacing is correspondingly reduced to avoid attitude mutation; a mapping relationship model of the propulsion power and the attitude stability is constructed, and the inclination angle in the real-time attitude monitoring data is taken as the core index, when the inclination angle is less than the safety threshold, the propulsion power adaptation value is set according to the cleaning efficiency priority principle, when the difference between the inclination angle and the safety threshold is less than the set difference, the propulsion power adaptation value is lowered according to the attitude stability priority principle, and the mapping relationship is corrected combined with the water flow disturbance coefficient, and when the water flow disturbance coefficient is increased, the upper limit of the propulsion power under the same inclination angle is correspondingly reduced; based on the collaborative correlation rule and the mapping relationship model, each parameter is quantitatively integrated: taking the adjustment direction of the track opening angle as the reference, the correction coefficient corresponding to the switching gain of the sliding mode is superimposed to obtain the real-time correction value of the track opening angle; taking the adjustment direction of the driving wheel spacing as the reference, the optimized value of the balance parameter and the distribution proportion of the balance weight are combined to obtain the dynamic adjustment amount of the driving wheel spacing; taking the corresponding propulsion power reference value of the attitude stability as the basis, the water flow disturbance coefficient is combined to obtain the propulsion power adaptation value; a parameter feasibility verification link is set, the real-time correction value of the track opening angle, the dynamic adjustment amount of the driving wheel spacing, and the physical limit parameters of the robot execution mechanism are compared, if the limit range is exceeded, the correction amplitude or adjustment amount is reduced in proportion, and the corresponding propulsion power adaptation value is simultaneously corrected to maintain the attitude balance; the real-time correction value of the track opening angle, the dynamic adjustment amount of the driving wheel spacing, and the propulsion power adaptation value that pass the verification are packaged as a standardized real-time control strategy, and the adjustment frequency of each parameter is included in the real-time control strategy, and the adjustment frequency is set according to the step complexity in the three-dimensional step model, the adjustment frequency is increased when the step complexity is high, and the adjustment frequency is reduced when the step complexity is low to reduce the calculation load.

[0077] Specifically, in the process of integrating various parameters and generating real-time control strategies, it is necessary to first establish the cooperative correlation rules between parameters to ensure that parameters of different control dimensions can adapt to each other and serve the attitude stability and work efficiency. For example, when the three-dimensional step model shows that the step surface flatness gradient of a certain area increases (such as the tile joint of a public swimming pool or the step stage with slight wear and tear), the sliding mode control module will increase the sliding mode switching gain to enhance the anti-disturbance ability of the device, and at this time, the output of the inverted pendulum control model will be bound with the gain to adjust the direction of the track angle. If the track angle originally needs to be slightly increased to adapt to the vertical inclination, the sliding mode switching gain will be increased, and the correction amplitude of the track angle will be correspondingly expanded, so that the contact area between the track and the step surface is larger, and the risk of slipping caused by poor flatness is further offset. In the high-risk area identified by the three-dimensional model (such as the sharp protrusions formed at the edge of the landscape pool step due to collision), the balance weight will be significantly increased, and at this time, the driving wheel spacing adjustment direction will be bound with the balance parameter and the weight, that is, even if the driving wheel spacing originally needs to be reduced to adapt to the narrow step, the single adjustment amount will be correspondingly reduced to avoid the sudden change of the wheel spacing, which causes the robot's center of gravity to deviate and causes the attitude to mutate.

[0078] Then, a mapping relationship model of propulsion power and attitude stability is constructed, and the inclination angle in real-time attitude monitoring data is taken as the core judgment index. For example, in the flat ceramic tile step area of the shallow water area of the private swimming pool, if the inclination angle of the robot is far less than the safety threshold (attitude stability), the propulsion power adaptive value will be set according to the cleaning efficiency priority principle, to ensure that the robot can complete the cleaning at a faster speed. If the robot is climbing a step with a large height difference, the inclination angle gradually approaches the safety threshold, and the difference between the inclination angle and the threshold is reduced to a certain extent, the propulsion power adaptive value will be immediately adjusted according to the attitude stability priority principle, and the influence of inertia on the attitude is reduced by reducing the moving speed. At the same time, the mapping relationship will be dynamically corrected in combination with the water flow disturbance coefficient. For example, at the entrance of the public swimming pool, even if the inclination angle of the robot is small, but the water flow disturbance coefficient increases due to the entry and exit of personnel, the upper limit of the propulsion power under the same inclination angle will be correspondingly reduced to avoid the superposition of water flow impact and high-power movement leading to attitude out of control. Based on the above-mentioned collaborative correlation rules and mapping relationship model, the specific control values are obtained by quantifying and integrating each parameter. For example, when the robot is working in the arc-shaped step area of the landscape swimming pool, the inverted pendulum control model outputs the adjustment direction of increasing the track angle, and the sliding mode switching gain is increased by a certain amplitude due to the moderate surface flatness gradient, and the real-time correction value of the track angle is calculated based on the increase direction and the correction coefficient corresponding to the sliding mode gain. If the drive wheel spacing adjustment direction is to be reduced, and the balance parameter of the current area needs to be enhanced after being optimized by fuzzy control, the balance weight is slightly increased due to the slight convexity, and the dynamic adjustment amount of the drive wheel spacing is calculated based on the reduction direction, the optimized value of the balance parameter and the weight distribution ratio.

[0079] Subsequently, a parameter feasibility verification link is set to ensure that all control values meet the physical limits of the robot actuators. For example, if the calculated real-time correction value of the track angle exceeds the maximum angle limit of the track mechanical structure, the correction amplitude will be reduced in proportion to control the angle within the limit range, and the propulsion power adaptive value will be simultaneously reduced to compensate for the slight lack of track grip force by reducing the speed, to avoid slipping. If the dynamic adjustment amount of the drive wheel spacing is less than the minimum adjustable spacing of the wheel shaft, the adjustment amount is appropriately increased to ensure the normal operation of the wheel shaft, and the balance parameter is fine-tuned to enhance the attitude stability to offset the influence of the wheel spacing adjustment not meeting the expectation.

[0080] Finally, the real-time correction value of the track opening angle that passes the check, the dynamic adjustment amount of the driving wheel spacing, and the propulsion power adaptation value are packaged into a standardized real-time control strategy. The strategy also includes the adjustment frequency of each parameter. The frequency is set according to the complexity of the three-dimensional step model. For example, in the complex step area of a public swimming pool with large height differences and dense protrusions and depressions, the adjustment frequency will be significantly increased to ensure that the robot can update the control parameters according to the changes in the environment every time it moves a small distance. In the structured step area of a private swimming pool with a flat surface, the adjustment frequency will be appropriately reduced to reduce the computational load of the control device and avoid the impact of frequent parameter adjustment on cleaning efficiency. Ultimately, a real-time control strategy is formed that is suitable for the current environment and takes into account the feasibility of execution.

[0081] Step S105, based on the real-time control strategy, adjust the motion state of the robot to adapt to the step cleaning process in the target step area. In the above steps, the execution of the real-time control strategy realizes the dynamic adaptation of the robot motion state to the step cleaning scene, that is, through the real-time correction of the track opening angle and the driving wheel spacing, the robot is ensured to be in close contact with the step surface, such as adjusting the adhesion in the depression area. Through the dynamic distribution of the propulsion power, the cleaning efficiency and the posture stability are balanced, such as reducing the power in the high inclination area to maintain balance. This process not only improves the cleaning coverage and reduces the missed cleaning area caused by posture deviation, but also reduces the risk of collision between the robot and the step structure, which helps to solve the technical problems of traditional underwater cleaning equipment such as rough operation and easy damage to the steps or itself, and realizes efficient and safe underwater step cleaning.

[0082] In the above steps, the control device receives the real-time correction value of the track opening angle, the dynamic adjustment amount of the driving wheel spacing, and the propulsion power adaptation value output in the real-time control strategy in real time, and converts these parameters into specific instructions that can drive the execution mechanism to act, and through the coordinated operation of the execution mechanism, the dynamic adjustment of the robot's motion state is realized. Taking the ceramic tile step area with local depressions in a private swimming pool as an example, when the strategy instruction clearly needs to increase the track opening angle, the robot's track adjusting motor (or hydraulic drive device) will slowly change the opening amplitude of the track according to the correction value. The originally suspended track in the middle due to the depression will be adjusted by the opening angle to make the two sides of the track more closely adhere to the step surface of the depression area, not only improving the adhesion of the track and the step, avoiding the robot from slipping when moving and cleaning, but also ensuring that the bottom cleaning brush fully contacts the bottom of the depression, preventing the brush from failing to reach the depression area due to the suspended track, resulting in missed cleaning. In terms of driving wheel spacing adjustment, if the robot enters the arc-shaped narrow step area of the landscape pool, the dynamic adjustment amount in the strategy will guide the wheel shaft adjusting mechanism to gradually reduce the driving wheel spacing, so that the wheels fit the width profile of the arc-shaped step, avoiding the robot from colliding with the pool wall due to the excessive width of the wheel spacing. If the step surface is distributed with dense small protrusions, the wheel spacing adjustment speed will be slowed down accordingly, and the track opening angle will be slightly corrected to prevent the robot's center of gravity from shifting when the wheels quickly switch the spacing, further maintaining the stability of the posture. The adjustment of the propulsion power will closely adapt to the needs of the current cleaning scene, such as in the flat and long step area of a public swimming pool, the propulsion power adaptation value set by the strategy is higher, and the driving motor drives the robot to move at a constant speed with a higher speed, improving the cleaning coverage in a unit of time. When the robot enters the high-inclination-angle step climbing area (such as the high step from the deep water area to the shallow water area of the pool), the power will be adjusted downward according to the instruction, reducing the moving speed to reduce the risk of forward inclination caused by inertia, and cooperating with the increased track opening angle to ensure sufficient adhesion during climbing, avoiding the risk of rushing or sliding due to excessive power. During the entire adjustment process, the robot-mounted posture sensor and track contact pressure sensor will continuously feedback state data. If the sensor detects that the track adhesion force has not reached the expected value, the control device will fine-tune the track opening angle correction value based on the feedback signal. If it is found that the cleaning speed is too slow after the propulsion power is reduced, but the posture is always stable, the power will be appropriately increased to the balance interval of efficiency and stability, and finally through this feedback adjustment, the robot's motion state is always accurately adapted to the structural features (depression, arc, high inclination) of the current step and the cleaning needs, ensuring that the cleaning brush can fully cover all corners of the step, and avoiding collisions due to the inadaptation of the motion parameters and the structure of the step, achieving efficient and safe underwater step cleaning.

[0083] Further optionally, after adjusting the motion state of the robot based on the real-time control strategy, the cleaning effect data and residual dirt information of the corresponding step region after the motion state is adjusted can also be collected through the image sensor and the dirt detection sensor integrated with the robot, the cleaning effect data including cleaning coverage and dirt removal rate, and the residual attitude deviation data and the actuator load data of the robot are collected, the cleaning effect data, the residual dirt information, the residual attitude deviation data, the actuator load data, and the preset cleaning standard and the attitude safety threshold of the corresponding region in the three-dimensional step model are compared, and it is judged whether the current cleaning region meets the preset cleaning requirement and whether the attitude is maintained stable. If the cleaning coverage is lower than the preset cleaning standard, or the residual dirt information is lower than the preset cleaning standard, the structure parameters and the water flow disturbance coefficient of the current region in the three-dimensional step model are extracted, and the deviation reason is analyzed combined with the residual dirt information, if the cleaning component is not fully contacted due to the surface flatness gradient, the secondary fine adjustment of the real-time correction value of the track opening angle in the real-time control strategy is triggered, and the contact pressure of the cleaning component and the step surface is increased. The structure parameters include step height difference, surface flatness gradient, protruding feature or recessed feature. If the residual attitude deviation data is greater than the attitude safety threshold, or the local load associated with the driving wheel in the actuator load data is higher than the preset regular load level, the adaptive control model based on the inverted pendulum balance theory is used to supplement the adjustment of the sliding mode switching gain and the propulsion power adaptation value, and the attitude anti-disturbance ability is strengthened. After the deviation correction is completed, the cleaning effect data, the residual dirt information, the residual attitude deviation data, and the actuator load data of the corresponding region are collected again for verification, until the cleaning effect data and the residual dirt information meet the preset cleaning requirement, the residual attitude deviation data meet the attitude safety threshold, and the actuator load data meet the preset regular load level.

[0084] If the cleaning effect data and the residual dirt information meet the preset cleaning requirement, the residual attitude deviation data meet the attitude safety threshold, and the actuator load data meet the preset regular load level, it means that the robot meets the preset cleaning requirement and the attitude is maintained stable. In this case, further optionally, the parameter combination of the current real-time control strategy is associated with the structure parameters and the water flow disturbance coefficient of the corresponding step region, and a mapping sample library is formed. The parameter combination includes the track opening angle correction value, the driving wheel spacing adjustment amount, and the propulsion power adaptation value. According to the current cleaning progress, the path planning of the subsequent cleaning region is updated combined with the topological relationship matrix of the three-dimensional step model, the structure parameters and the cleaning effect data of the cleaned region are fed back to the graph neural network, and the surface state feature of the cleaned region in the three-dimensional step model is updated and corrected.

[0085] Exemplarily, in the tile step area of the shallow water area of the public swimming pool, after the robot completes a segment of cleaning based on the real-time control strategy adjustment of the motion state, the integrated image sensor and the dirt detection sensor are started to collect data. The image sensor generates visual data related to cleaning coverage by shooting the underwater picture of the area, and can clearly capture that there is still dirt remaining in a small recess on the surface of the step, while the flat area around the recess is basically clean. The dirt detection sensor further confirms that the dirt removal rate in the recess does not meet the expectation through contact detection. At the same time, the attitude sensor records the attitude residual deviation data of the robot at this time, and finds that there is still slight tilting due to uneven force on the track in the recess, and the actuator load sensor also shows that the local load of the driving wheel is higher than the pre-set regular load level due to the adhesion of the edge of the recess area of the step. Then, the collected data is compared with the pre-set cleaning standard (such as the cleaning coverage needs to cover the entire area, and the dirt removal rate needs to reach a certain effect) of the area in the three-dimensional step model and the attitude safety threshold, and it is found that the cleaning coverage is lower than the pre-set cleaning standard, and the attitude residual deviation is greater than the attitude safety threshold, or the local load associated with the driving wheel in the actuator load data is higher than the pre-set regular load level, which indicates that the robot does not fully meet the stability requirement. By analyzing the deviation reasons, combined with the structure parameters (showing that this place is a recess feature with small depth) of the area in the three-dimensional step model and the water flow disturbance coefficient (there is slight water flow disturbance in this area due to the proximity to the water inlet of the swimming pool), it is judged that the low cleaning coverage is mainly due to the insufficient contact of the cleaning brush with the surface of the step caused by the recess, and the attitude residual deviation is related to the slight water flow disturbance.

[0086] For the problem of insufficient contact during cleaning, a secondary fine-tuning of the real-time control strategy for the real-time correction value of the track opening angle is triggered, which appropriately increases the track opening angle, making the track more conform to the concave profile and driving the cleaning brush to press against the bottom of the concave, thereby increasing the contact pressure to enhance the dirt peeling effect. For the attitude deviation caused by water flow disturbance, an adaptive control model based on the inverted pendulum balance theory is used to supplement the adjustment of the sliding mode switching gain, improve the response speed of the device to water flow disturbance, and at the same time, slightly reduce the propulsion power adaptation value to reduce the impact of water flow on the robot attitude. After completing the deviation correction, the robot cleans the concave area again and re-collects the cleaning effect and attitude data. The image sensor shows that the dirt in the concave has been removed, the cleaning coverage rate meets the standard, the dirt detection sensor confirms that the peeling rate meets the preset standard, the attitude residual deviation is reduced to within the safety threshold, and the actuator load also returns to normal. At this time, the parameter combination of the current real-time control strategy (including the adjusted track opening angle correction value, the driving wheel spacing adjustment amount, and the propulsion power adaptation value) is associated with the structural parameters of the region (concave features, surrounding step height difference), water flow disturbance coefficient, and is stored in the mapping sample library, providing a reference for the control strategy generation of similar scenarios in the future. At the same time, according to the current cleaning progress (1 / 3 of the shallow area 1 / 3 step area has been cleaned), combined with the topological relationship matrix of the three-dimensional step model (most of the remaining steps in the shallow area are flat structures, and there are steps with large height differences at the junction of the deep water area), the path planning of the subsequent cleaning area is updated, and the remaining flat steps in the shallow area are cleaned first, and then the area with large height differences in the deep water area is pushed. The structural parameters of the cleaned area (such as the actual depth of the concave and the surface flatness variation) and the cleaning effect data (such as the difference in dirt peeling rate in different areas) are fed back to the graph neural network, and the surface state features of the cleaned area in the three-dimensional step model are updated and corrected, such as updating the original model's "slight concave" to "slight concave with no residual dirt after cleaning", so that the model can more accurately reflect the actual state of the steps, providing more reliable environmental data support for subsequent cleaning operations.

[0087] Further optionally, if the actuator load data exceeds the physical constraint threshold or a sudden failure signal is detected after adjusting the motion state, an emergency control mechanism is triggered, based on the spatial features of the current position in the three-dimensional step model, the preset emergency posture adjustment parameters are called, the minimum balance torque is calculated through the inverted pendulum control model, the robot is controlled to switch to a low-power stable mode, and the environmental parameters and control parameters at the time of failure are recorded and fed back to the fault diagnosis module for strategy optimization. For example, in the step area between the deep water area and the shallow water area of a public swimming pool, the robot is slowly moving along the step surface for cleaning, the track suddenly gets stuck in the plastic foreign matter left in the step gap, and the actuator load sensor immediately detects that the load data of the track drive motor exceeds the physical constraint threshold, while the motor controller feeds back a sudden failure signal of slight stall. At this time, the device immediately triggers the emergency control mechanism, first retrieves the spatial feature data of the current position in the three-dimensional step model. The step width in this area is narrow, and there is a gentle transition area where the step facade connects with the pool bottom, without other protrusions or recesses. Based on these spatial features, the device quickly calls the preset emergency posture adjustment parameters, which are temporary adjustment standards pre-set for the track stuck scenario, including the minimum track opening angle, the safe distance between drive wheels, and other basic configurations. Subsequently, the minimum balance torque to maintain the current posture of the robot is calculated through the inverted pendulum control model to avoid tilting and overturning of the robot body due to track sticking. The model combines the current tilt angle and the structural features of the gentle transition area of the step to obtain a result that only a small torque is needed to maintain balance. The device controls the robot to switch to a low-power stable mode accordingly. The propulsion power is greatly reduced to prevent damage to the motor caused by continuous high load. At the same time, the track opening angle is adjusted to the minimum value set by the emergency parameters, reducing the degree of track sticking with the foreign matter, and the drive wheel spacing is slowly adjusted to slowly drive the track to escape from the foreign matter. During the entire process, the attitude sensor monitors the robot body state in real time to ensure that the robot always maintains a horizontal posture and gradually moves away from the stuck position to the gentle transition area of the step facade. During the emergency handling, the device automatically records the environmental parameters and control parameters at the time of failure. The environmental parameters include the step gap width of the current area, the surrounding water flow disturbance, and the approximate size features of the foreign matter (indirectly judged by the load change curve). The control parameters include the track opening angle, the drive wheel spacing, and the propulsion power value before the failure. These data are fed back to the fault diagnosis module in real time, and the module will analyze the causes of track sticking in the future, such as whether the mismatch between the current step gap width and the design gap of the track leads to easy sticking of foreign matter, and then optimize the subsequent control strategy. For example, in a similar step area with dense gaps, the track opening angle is slightly increased in advance to reduce the probability of foreign matter sticking and improve the robustness of the robot in dealing with sudden failures.

[0088] The embodiment of the present application can improve the environmental adaptability, posture stability and work efficiency of the robot in a complex underwater step environment, and solves the problem of insufficient adaptation of traditional underwater cleaning equipment to underwater step structures.

[0089] The control device of the pool underwater step cleaning robot provided by the embodiment of the present application comprises: a collection module for scanning the step structure in the target step area to construct a three-dimensional step model; a starting module for generating an initial control strategy according to the three-dimensional step model by a deep deterministic policy gradient algorithm before water entry cleaning, wherein the initial control strategy at least includes one parameter: a track opening angle, a driving wheel distance and a propulsion power, and the robot is started to enter the target step area based on the initial control strategy; a cleaning module for collecting real-time motion state data and posture monitoring data of the robot during the cleaning process; an adaptive control model based on the inverted pendulum balance theory for cooperative optimization of the real-time collected motion state data, posture monitoring data and step structure parameters in the three-dimensional step model to obtain a real-time control strategy of the robot; the adaptive control model is constructed by a sliding mode control, a fuzzy control and an inverted pendulum control model; and the motion state of the robot is adjusted based on the real-time control strategy to adapt to the step cleaning process in the target step area. In some embodiments, the control device of the pool underwater step cleaning robot can be applied to a terminal device. For example, the device can be carried in the underwater robot shown. Figure 2 It should be noted that, for the convenience and brevity of description, the specific working process of the control device of the pool underwater step cleaning robot described above can refer to the corresponding process in the foregoing embodiment of the control method of the pool underwater step cleaning robot, which will not be described here again.

[0090] The terminal device 300 includes a processor 301 and a memory 302, and the processor 301 and the memory 302 are connected through a bus 303, such as an I2C bus. Specifically, the processor 301 is configured to provide computing and control capabilities to support the operation of the entire terminal device. The processor 301 can be a central processing unit, and the processor 301 can also be other general-purpose processors, digital signal processors, application-specific integrated circuits, field programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, and the like. Those skilled in the art can understand that the structures shown in the above embodiments are only block diagrams of part of the structures related to the embodiments of the present application, and do not constitute a limitation on the terminal device to which the embodiments of the present application are applied. Specifically, the server can include more or fewer components than those shown in the figure, or combine certain components, or have a different component arrangement. The processor is configured to run a computer program stored in the memory and implement the control method of the underwater pool step cleaning robot according to any one of the embodiments of the present application when the computer program is executed. It should be noted that those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the terminal device described above can refer to the control method of the underwater pool step cleaning robot described above, and will not be described here.

[0091] The embodiments of the present application also provide a storage medium for computer readable storage, and the storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the steps of any one of the control methods of the underwater pool step cleaning robot provided in the specification of the embodiments of the present application.

Claims

1. A control method of a pool underwater step cleaning robot, characterized by, The method comprises: scanning the step structure in the target step region to build a three-dimensional step model; before water entry cleaning, generating an initial control strategy according to the three-dimensional step model through a deep deterministic policy gradient algorithm, the initial control strategy comprising at least one of the following parameters: a track opening angle, a driving wheel spacing, and a propulsion power, and starting the robot into the target step region based on the initial control strategy; in the cleaning process, real-time collection of motion state data and attitude monitoring data of the robot; based on an adaptive control model of an inverted pendulum balance theory, cooperative optimization of the real-time collected motion state data, attitude monitoring data, and step structure parameters in the three-dimensional step model is performed to obtain a real-time control strategy of the robot; the adaptive control model is constructed by a sliding mode control, a fuzzy control, and an inverted pendulum control model; based on the real-time control strategy, the motion state of the robot is adjusted to adapt to the step cleaning process in the target step region.

2. The control method of the pool underwater step cleaning robot according to claim 1, characterized in that, The scanning of the step structure in the target step region to build a three-dimensional step model comprises: scanning the target step region by a multi-source sensor integrated with the robot to obtain structure parameters of the target step region; the structure parameters comprise the height, width, and vertical surface inclination angle of each step in the target step region; obtaining flatness information of each step surface in the target step region by a ranging sensor; based on the structure parameters and the flatness information, modeling the spatial topological relationship of the steps by a graph neural network to build a three-dimensional step model of the target step region.

3. The control method of the pool underwater step cleaning robot according to claim 2, characterized in that, The modeling of the spatial topological relationship of the steps by the graph neural network based on the structure parameters and the flatness information to build a three-dimensional step model of the target step region comprises: encapsulating the structure parameters of each step and the flatness information of the corresponding region as nodes of the graph neural network, and the structure parameter features of each node comprising one of the following: step height, step width, step vertical surface inclination angle, and step surface flatness quantization value; defining edge features between the nodes according to the spatial positional relationship between the steps, the edge features comprising one of the following: height difference, horizontal distance, and connecting surface slope of adjacent steps; performing feature aggregation of the structure parameter features and the edge features by a graph neural network composed of a graph convolution layer, an attention mechanism layer, and a fully connected layer, learning the topological correlation law between the steps, and outputting a topological relationship matrix comprising a global spatial layout; fusing the topological relationship matrix and the structure parameter features of each node to generate a three-dimensional step model comprising step geometric parameters, surface state, and spatial connection relationship, and dynamically adjusting the grid density of the three-dimensional step model by a quadtree algorithm, and based on the optimized grid, combining the structure parameters of each node to build a step three-dimensional geometric contour by a triangulation algorithm, embedding a multi-dimensional texture of the grid unit corresponding to each node by luminance remapping algorithm to reflect the surface flatness difference between the nodes, to obtain the final output three-dimensional step model.

4. The control method of the pool underwater step cleaning robot according to claim 3, characterized in that, The generation of the initial control strategy according to the three-dimensional step model through the deep deterministic policy gradient algorithm comprises: The step height difference, facade inclination angle, surface flatness gradient, and spatial connection relationship are extracted from the three-dimensional step model, and the water flow disturbance coefficient obtained by fluid dynamics simulation is combined to form the state space input; the water flow disturbance coefficient is obtained by fluid dynamics simulation based on the pre-collected historical water flow velocity data and historical step spatial structure data; The track opening angle adjustment range, drive wheel spacing variation interval, and propulsion power threshold are set as continuous action space, and the physical constraint model of the actuator is embedded in the Actor-Critic dual network architecture, so that the control parameters output by the Actor network meet the mechanical movement limit; A dynamic reward function is set, a penalty coefficient is set for the high-risk area identified in the three-dimensional step model, an excessive reward is given to the action of avoiding in advance, and the reward weight is adjusted adaptively according to the step complexity; the high-risk area includes the step edge with a convex height greater than a convex obstacle threshold; A three-dimensional step model is used to generate a diversified virtual scene for pre-training, and a fine-tuning is performed in combination with the fault cases in the historical cleaning data, the output accuracy of the Actor network is optimized by the deep deterministic policy gradient algorithm, and the Critic network dynamically evaluates the action value and corrects the policy bias in the reverse direction; The trained Actor-Critic dual network architecture is applied to the three-dimensional step model to generate an initial control strategy including a track opening angle dynamic adjustment curve, a drive wheel spacing matching scheme, and a propulsion power gradient distribution table, and the initial control strategy is pre-checked and corrected in combination with the convex features and concave features in the three-dimensional step model; After generating the initial control strategy, the posture stability of the initial control strategy in the step transition area is verified through model simulation, and if the predicted inclination angle exceeds the safety threshold, the parameter adaptive correction of the initial control strategy is triggered.

5. The control method of the pool underwater step cleaning robot according to claim 4, characterized in that, The method based on the initial control strategy to start the robot into the target step area comprises: Based on the surface flatness gradient of the entrance area in the three-dimensional step model and the water flow disturbance coefficient, a safe entrance point that meets the preset safe entrance condition is selected, and a path planning from the initial parking position of the robot to the entrance point is generated; the preset safe entrance condition includes that the water flow disturbance coefficient is lower than the set low disturbance threshold, and the surface flatness gradient is lower than the set convex obstacle threshold; The robot actuator pre-checking is started, the track is adjusted to the opening angle initial value in the initial control strategy, the drive wheel is matched to the spacing reference value, and whether the actuator load meets the physical constraint is detected through the torque sensor; if the load exceeds the threshold, the corresponding parameter is fine-tuned; The robot is controlled to enter the water and enter the path execution stage, the water flow velocity data is collected in real time, and the water flow disturbance deviation value is calculated based on the real-time collected water flow velocity data; the propulsion power is dynamically adjusted according to the propulsion power gradient distribution table of the initial control strategy and the water flow disturbance deviation value, so as to avoid the posture deviation caused by water impact; When the distance between the robot and the entry point is less than a set distance, the pre-adaptation of the track angle dynamic adjustment curve is triggered in advance based on the convex features and concave features in the three-dimensional step model, and if the robot is at the edge of a step with a convex height greater than a convex obstacle threshold in the entry area, the track gripping force is increased; After the robot enters the initial section of the target step area, real-time motion state data and attitude data are collected, and compared with the expected structure parameters of the three-dimensional step model. If the attitude deviation is less than or equal to a preset threshold, it is confirmed that the robot has stably entered, and if the attitude deviation is greater than the preset threshold, preliminary attitude correction is performed based on an inverted pendulum control model.

6. The control method of the pool underwater step cleaning robot according to claim 1, wherein, The adaptive control model based on the inverted pendulum balance theory cooperatively optimizes the real-time motion state data, attitude monitoring data, and step structure parameters in the three-dimensional step model to obtain a real-time control strategy for the robot, including: The speed and acceleration in the real-time motion state data, the inclination angle and angular velocity in the attitude monitoring data, the step height difference, the surface flatness gradient, the convex features, and the concave features in the three-dimensional step model structure parameters are spatio-temporally aligned, and abnormal data are removed by fusing the flow disturbance coefficient to generate an input data set; A robot dynamics model is established based on the inverted pendulum balance theory as an inverted pendulum control model. The inclination angle in the input data set is taken as a balance deviation reference, the step height difference is combined to calculate the balance torque required for the robot to maintain attitude balance through the robot dynamics model, and the adjustment direction of the track angle and the drive wheel spacing is generated based on the balance torque; The flow disturbance coefficient and the uncertainty disturbance coefficient caused by the unevenness of the steps are obtained through a sliding mode control module, and the sliding mode switching gain is dynamically adjusted according to the flow disturbance coefficient, the uncertainty disturbance coefficient, and the flatness gradient. The uncertainty disturbance coefficient is set based on historical step spatial structure data; The step complexity and attitude deviation degree are fuzzy reasoned by a fuzzy control module, and the balance parameters of the inverted pendulum control model and the gain coefficient of the sliding mode control are adaptively optimized. The balance weight is increased for high-risk areas. The step complexity is determined by the height difference and the convex density. The adjustment direction of the track angle and the drive wheel spacing, the sliding mode switching gain, the balance parameter, the gain coefficient, the balance weight, the mapping relationship between the propulsion power and the attitude stability are cooperatively integrated to generate a real-time control strategy including the real-time correction value of the track angle, the dynamic adjustment amount of the drive wheel spacing, and the propulsion power adaptation value.

7. The control method of the pool underwater step cleaning robot according to claim 6, wherein, The robot dynamics model is established based on the inverted pendulum balance theory as an inverted pendulum control model, including: The input parameters of the robot dynamics model are determined, including the physical parameters of the robot, real-time attitude monitoring data, structure parameters extracted from the three-dimensional step model, and flow disturbance coefficients. The physical parameters of the robot include moment of inertia, damping coefficient, and restoring coefficient. The real-time attitude monitoring data includes inclination angle and angular velocity. The three-dimensional step model structure parameters include step height difference and surface flatness gradient. The flow disturbance coefficient is obtained from fluid dynamics simulation. The cleaning operation state of the robot is equivalent to an inverted pendulum device, an inclination angle is mapped to an angle at which the inverted pendulum deviates from a balance position, an angular velocity is mapped to an angular rate of the inverted pendulum, a step height difference is converted into an external tilting disturbance torque of the inverted pendulum, a surface flatness gradient is converted into a contact damping disturbance of the inverted pendulum, and a water flow disturbance coefficient is converted into a lateral additional torque of the inverted pendulum, so as to construct an inverted pendulum dynamics equation containing multiple disturbance terms; A cleaning component contact reaction force correction term is generated based on the inverted pendulum dynamics equation, the contact reaction force correction term is calculated from a friction coefficient of a robot brush and a step surface and real-time pressure sensor data, and the contact reaction force correction term is used to compensate for an additional action force of the brush on the robot posture during the cleaning operation; The damping coefficient in the inverted pendulum dynamics equation is dynamically adjusted according to the flatness gradient of the step surface in the three-dimensional step model, and when the flatness gradient increases, the damping coefficient is correspondingly increased, and when the flatness gradient decreases, the damping coefficient is correspondingly decreased; The parameters of the inverted pendulum dynamics equation are iteratively corrected by using posture monitoring data and motion state data of the robot collected in a historical cleaning process, a deviation between a balance torque output by a robot dynamics model and an actual required balance torque is calculated, a gradient descent algorithm is used to optimize the moment of inertia and the restoring coefficient, so that a model prediction deviation of the robot dynamics model is controlled within a preset range, and a robot dynamics model adapted to the pool step scene is constructed as an inverted pendulum control model.

8. The control method of the pool underwater step cleaning robot according to claim 6, wherein, The track opening angle, the adjustment direction of the drive wheel spacing, the sliding mode switching gain, the balance parameter, the gain coefficient, and the balance weight are collaboratively integrated, and a mapping relationship between the propulsion power and the posture stability is combined to generate a real-time control strategy containing a real-time correction value of the track opening angle, a dynamic adjustment amount of the drive wheel spacing, and an adaptive value of the propulsion power, including: A collaborative correlation rule of each parameter is established, the track opening angle adjustment direction and the sliding mode switching gain output by the inverted pendulum control model are bound, when the sliding mode switching gain is increased due to an increase in the flatness gradient, the correction amplitude of the track opening angle adjustment direction is correspondingly expanded, and the drive wheel spacing adjustment direction and the balance parameter and the balance weight are bound, when the balance weight is increased in a high-risk area, the single adjustment amount of the drive wheel spacing adjustment direction is correspondingly reduced; A mapping relationship model of the propulsion power and the posture stability is constructed, the inclination angle in the real-time posture monitoring data is taken as a core index, when the inclination angle is less than a safety threshold, the adaptive value of the propulsion power is set according to the cleaning efficiency priority principle, when a difference between the inclination angle and the safety threshold is less than a set difference, the adaptive value of the propulsion power is down-regulated according to the posture stability priority principle, and the mapping relationship is corrected in combination with the water flow disturbance coefficient, and when the water flow disturbance coefficient is increased, the upper limit of the propulsion power under the same inclination angle is correspondingly reduced. Based on the synergistic association rules and the mapping relationship model, each parameter is quantitatively integrated: taking the track angle adjustment direction as the reference, the real-time correction value of the track angle is obtained by superimposing the correction coefficient corresponding to the sliding mode switching gain; taking the driving wheel spacing adjustment direction as the reference, the dynamic adjustment amount of the driving wheel spacing is obtained by combining the optimized value of the balance parameter and the distribution proportion of the balance weight; taking the corresponding reference value of the attitude stability as the basis, the adaptive value of the propulsion power is obtained by combining the water flow disturbance coefficient; The parameter feasibility verification link is set, the real-time correction value of the track angle, the dynamic adjustment amount of the driving wheel spacing and the physical limit parameters of the robot executive mechanism are compared, if it exceeds the limit range, the correction amplitude or adjustment amount is reduced in proportion, and the adaptive value of the propulsion power is corrected synchronously; The real-time correction value of the track angle, the dynamic adjustment amount of the driving wheel spacing and the adaptive value of the propulsion power that pass the verification are packaged into a standardized real-time control strategy, the real-time control strategy includes the adjustment frequency of each parameter, the adjustment frequency is set according to the step complexity in the three-dimensional step model, the adjustment frequency is increased when the step complexity is high, and the adjustment frequency is reduced when the step complexity is low.

9. The control method of the pool underwater step cleaning robot according to claim 1, wherein, After adjusting the motion state of the robot based on the real-time control strategy, it further includes: Through the image sensor and the dirt detection sensor integrated by the robot, the cleaning effect data and residual dirt information of the corresponding step region after the current adjustment motion state are collected, the cleaning effect data includes cleaning coverage and dirt peeling rate, and the attitude residual deviation data and the executive mechanism load data of the robot are collected, and the cleaning effect data, residual dirt information, attitude residual deviation data and executive mechanism load data are compared with the preset cleaning standard and attitude safety threshold, and the preset regular load level of the corresponding region in the three-dimensional step model; If the cleaning coverage is lower than the preset cleaning standard, or the residual dirt information is lower than the preset cleaning standard, the structure parameters and water flow disturbance coefficient of the current region in the three-dimensional step model are extracted, the deviation reason is analyzed combined with the residual dirt information, if the surface flatness gradient leads to insufficient contact of the cleaning component, the secondary fine adjustment of the real-time correction value of the track angle in the real-time control strategy is triggered, and the contact pressure of the cleaning component and the step surface is increased; the structure parameters include step height difference, surface flatness gradient, convex feature or concave feature; If the attitude residual deviation data is greater than the attitude safety threshold, or the local load of the driving wheel associated in the executive mechanism load data is higher than the preset regular load level, the adaptive control model based on the inverted pendulum balance theory is used to supplement the adjustment of the sliding mode switching gain and the propulsion power adaptive value, and the attitude anti-interference ability is strengthened; After completing the deviation correction, the cleaning effect data, residual dirt information, attitude residual deviation data and executive mechanism load data of the corresponding region are collected again for verification, until the cleaning effect data and residual dirt information meet the preset cleaning requirements, the attitude residual deviation data meet the attitude safety threshold, and the executive mechanism load data meet the preset regular load level.

10. A control device of a pool underwater step cleaning robot, characterized by, The device comprises: The acquisition module is used for scanning the step structure in the target step region to construct a three-dimensional step model; The starting module is used for generating an initial control strategy according to the three-dimensional step model through a deep deterministic policy gradient algorithm before water entry cleaning, wherein the initial control strategy at least comprises one parameter of a track opening angle, a driving wheel spacing and a propulsion power, and the robot is started to enter a target step region based on the initial control strategy; The cleaning module is used for collecting motion state data and attitude monitoring data of the robot in real time during the cleaning process; based on an adaptive control model of an inverted pendulum balance theory, the motion state data, the attitude monitoring data and step structure parameters in the three-dimensional step model are cooperatively optimized to obtain a real-time control strategy of the robot; the adaptive control model is obtained by constructing a sliding mode control, a fuzzy control and an inverted pendulum control model; and the motion state of the robot is adjusted based on the real-time control strategy to adapt to the step cleaning process in the target step region.

Images