A dense obstacle environment adaptive obstacle avoidance control method and mobile robot

Abstract

The application provides a dense obstacle environment adaptive obstacle avoidance control method, comprising: obtaining the outer contour geometric parameters of the current form of a mobile robot, establishing the vertex coordinate set and boundary line inequality group representation of the outer contour of the mobile robot, and updating the mathematical solution problem of the optimal distance weight of the obstacle point to the outer contour of the mobile robot; generating random sampling obstacle point coordinates in the preset working area around the current mobile robot, calculating the optimal distance weight of each sampling point to each boundary, and forming a training data set; iteratively training a distance neural network model by using the training data set; inputting the real-time perceived obstacle point coordinates into the trained distance neural network model, and outputting the predicted optimal distance weight in real time, which is used for outputting the obstacle avoidance direction. The application also provides a mobile robot. Furthermore, self-iteration is repeated in the whole task cycle, and the adaptive ability to the dense obstacle is continuously improved without stopping.

Description

Technical Field

[0001] This application belongs to the field of mobile robot technology, and more specifically, relates to an adaptive obstacle avoidance control method and a mobile robot in a dense obstacle environment. Background Technology

[0002] When a mobile robot navigates autonomously in a dense obstacle environment, one of the core aspects of obstacle avoidance is to quickly calculate the shortest distance between the obstacle and the robot itself, and then use the distance information to judge the collision risk and plan the obstacle avoidance path.

[0003] Traditional distance calculations mostly rely on iterative algorithms to find the shortest distance using fixed-shape polygons, or approximate estimations through grid map dilation. When the robot itself contains movable parts, or when the outer contour itself is irregular, the fitting error of fixed-shape polygons is large, and the computational cost of iterative solutions is high, making it difficult to meet real-time requirements.

[0004] Furthermore, when the outer contour shape of the robot changes during operation, traditional methods require re-modeling the contour and solving the distance, which cannot quickly adapt to the shape change and adjust the obstacle avoidance strategy, resulting in poor adaptability. Summary of the Invention

[0005] The purpose of this application is to provide an adaptive obstacle avoidance control method and a mobile robot for dense obstacle environments, so as to solve the technical problems of real-time performance and adaptability of mobile robots in the navigation and obstacle avoidance process in dense obstacle environments in the prior art.

[0006] To achieve the above objectives, the technical solution adopted in this application is: to provide an adaptive obstacle avoidance control method for dense obstacle environments, comprising the following steps:

[0007] Step S1: Obtain the geometric parameters of the outer contour of the current form of the mobile robot, establish the vertex coordinate set and boundary line inequality representation of the outer contour of the mobile robot, and update the mathematical solution problem of the optimal distance weight from the obstacle point to the outer contour of the mobile robot.

[0008] Step S2: Randomly generate the coordinates of sampling obstacle points within the preset working area around the mobile robot, calculate the optimal distance weights of each sampling point for each boundary, and form a training dataset;

[0009] Step S3: Iteratively train the distance neural network model using the training dataset;

[0010] Step S4: Input the real-time perceived obstacle point coordinates into the trained distance neural network model, and output the predicted optimal distance weights in real time to output the obstacle avoidance direction;

[0011] Step S5: Repeat steps S1-S4 above.

[0012] This application also provides a mobile robot, including a vehicle body, sensors, a motion controller, a computing unit, and side brushes and / or squeegees movably mounted on the vehicle body. The computing unit of the mobile robot is equipped with a distance neural network model trained by the above-mentioned adaptive obstacle avoidance control method for dense obstacle environments. Based on sensor data, the model is used to calculate in real time the shortest distance from each obstacle point in the surrounding environment to the outer contour of the mobile robot and the obstacle avoidance direction during autonomous navigation, so as to generate a safe movement trajectory for the motion controller.

[0013] The beneficial effects of the adaptive obstacle avoidance control method for dense obstacle environments provided in this application are as follows: Compared with existing technologies, by acquiring robot morphological parameters in real time and updating the optimization problem in step S1, the obstacle avoidance logic is transformed from fixed-shape programming to dynamic mathematical modeling, ensuring that subsequent training data remains mathematically synchronized with the latest shape, and rapidly reconstructing the optimal mapping rules of coordinates and weights when the shape changes. Step S2, combined with the current environment, automatically generates training data that reflects the geometric truth, so that each iteration of training accurately matches the current shape's adaptation requirements in the environment. The iterative training in step S3, combined with the loop in step S5, allows the neural network to quickly learn how to match the current shape boundary characteristics with the distribution of environmental obstacles with fewer resources. In steps S4-S5, the mobile robot iterates itself repeatedly throughout the entire task cycle, continuously improving its adaptive ability to dense obstacles and the real-time performance of navigation and obstacle avoidance without stopping.

[0014] The advantages of the mobile robot provided in this application are as follows: Compared with the prior art, the mobile robot applies the above-mentioned adaptive obstacle avoidance control method for dense obstacle environments, which can meet the requirements of low-latency real-time computing, and can also take into account both passage safety and passage efficiency in narrow and complex scenarios. Compared with traditional obstacle avoidance schemes, it has stronger adaptability and better robustness. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the adaptive obstacle avoidance control method for dense obstacle environments provided in this application.

[0017] Figure 2This is a schematic diagram illustrating the mathematical solution to the problem of updating the optimal distance weight based on the geometric parameters of the outer contour of the current form of a mobile robot, as provided in this application.

[0018] Figure 3 This is a schematic diagram illustrating the outer contour of the current form of the mobile robot provided in this application.

[0019] Figure 4 This is a schematic diagram illustrating the process of obtaining the true optimal distance weight based on a simulated contour region, as provided in this application.

[0020] Figure 5 This is a schematic diagram illustrating the association between the second largest rectangle and the main body, and the second smallest rectangle and the movable arm, provided for this application.

[0021] Figure 6 A schematic diagram of the multi-head network structure provided in this application.

[0022] Figure 7 This is a three-dimensional structural diagram of the mobile robot provided in this application. Detailed Implementation

[0023] To make the technical problems, technical solutions, and beneficial effects to be solved by this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0024] It should be noted that when a component is referred to as "fixed to" or "set on" another component, it can be directly set on the other component or indirectly set on the other component through an intermediate medium. When a component is referred to as "connected to" another component, it can be directly connected to the other component or indirectly connected to the other component through an intermediate medium.

[0025] It should be understood that the terms "length", "width", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0026] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0027] The adaptive obstacle avoidance control method for dense obstacle environments provided in the embodiments of this application will now be described.

[0028] The adaptive obstacle avoidance control method for dense obstacle environments is applicable to mobile robots, especially mobile robots whose shape changes in real time during movement. These robots change shape according to the needs of the task during movement, resulting in changes in their outer contour (edge ​​contour of the top-view projection area). Examples include humanoid robots, handling robots, firefighting robots, and cleaning robots.

[0029] Please refer to the following: Figure 1 The adaptive obstacle avoidance control method for dense obstacle environments includes the following steps:

[0030] Step S1: Obtain the geometric parameters of the outer contour of the current form of the mobile robot, establish the vertex coordinate set and boundary line inequality representation of the outer contour of the mobile robot, and solve the mathematical problem of updating the optimal distance weight from the obstacle point to the outer contour of the mobile robot.

[0031] Understandably, the core of step S1 lies in abstracting the robot's physical outer contour into a mathematical description, requiring only the robot's geometric parameters and no environmental information. The vertices of the outer contour can be arranged in either a counter-clockwise or clockwise order. A set of boundary line inequalities is used to determine whether any point is inside the robot. The mathematical solution to the optimal distance weights establishes a mapping relationship from the coordinates of obstacle points to the contribution of each boundary.

[0032] Step S2: Randomly generate the coordinates of sampling obstacle points within the preset working area around the mobile robot, calculate the optimal distance weight of each sampling point corresponding to each boundary, and form a training dataset.

[0033] It is understandable that in step S2, the essence of the training data is the geometric distance from a point to a polygon. This is a purely mathematical relationship that does not contain any environment-specific information. Therefore, the trained model naturally does not have the problem of deviation between simulation and reality.

[0034] Step S3: Iteratively train the distance neural network model using the training dataset.

[0035] Understandably, in step S3, the training dataset generated in step S2 is used to supervise the learning of the neural network, enabling the network to gradually learn to quickly predict the optimal distance weights of each boundary based on the coordinates of obstacle points. After training, the resulting model can process all obstacle points simultaneously in a batch parallel manner, changing the total time consumption from a relationship proportional to the number of points to a relationship independent of the number of points.

[0036] It is worth noting that the distance neural network model can be an existing distance calculation-related model, such as classic network structures like ResNet50, lightweight CNN, or MLP-Mixer. The loss function can be the existing mean squared error, which measures the error between the network's predicted weights and the true optimal weights. Furthermore, the selected existing distance calculation-related model can be an untrained initial model or a mature model that has undergone preliminary training or high-level training. Preferably, the distance neural network model is a classic ResNet50 network structure with mean squared error as the loss function, having undergone 2000 to 10000 iterations, thus possessing initial predictive ability while also considering the learning ability of new forms in the current environment.

[0037] Step S4: Input the real-time perceived obstacle point coordinates into the trained distance neural network model, and output the predicted optimal distance weights in real time to output the obstacle avoidance direction.

[0038] Understandably, in step S4, which is executed in actual operation, the robot obtains the coordinates of surrounding obstacle points in real time through sensors, calls the network model that has just been iteratively trained to perform rapid inference, and further converts the output weights into the obstacle avoidance direction required for obstacle avoidance decision. The obstacle avoidance direction describes which direction the robot should avoid when an obstacle approaches the robot, for use by the downstream motion controller.

[0039] Step S5: Repeat steps S1-S4 above.

[0040] Understandably, unlike traditional methods that treat the robot as a fixed shape, step S1, by acquiring the geometric parameters of the robot's current shape and updating the optimization problem in real time, transforms the underlying logic of the obstacle avoidance system from the existing one-time programming for a single shape to dynamic mathematical modeling that changes with shape. This mechanism ensures that subsequent training data generation remains precisely mathematically synchronized with the robot's latest shape. Whenever a shape change causes a change in the geometric relationship between the obstacle and the boundary, step S1 can instantly reconstruct the optimal mapping rules between coordinates and weights.

[0041] Step S2 combines the updated mathematical problem from Step S1 with the robot's current surrounding environment to automatically generate training data containing geometric truth. This ensures that the training material for each iteration accurately reflects the geometric adaptation requirements of the current form in the current environment. More importantly, the iterative training in Step S3, combined with the loop in Step S5, allows the distance neural network model to continuously learn the adaptability of the new form in the current environment, quickly learning to match the geometric boundary characteristics of the current form with the obstacle distribution of the current environment with fewer computational resources.

[0042] It is worth noting that this process does not require manual data collection and labeling. The training data based on pure mathematical relationships minimizes the computational load of each round of iterative training. Based on the processing capabilities of existing conventional mobile robots, each iteration takes less than 5 milliseconds, which means about 200 iterations can be completed per second, enabling the convergence of the new form in the current environment and ensuring the real-time requirements of the mobile robot.

[0043] In steps S4-S5, the mobile robot no longer relies solely on its factory performance during offline deployment, but instead iterates itself repeatedly throughout the entire task cycle, gradually improving its ability to adapt to dense obstacles in the current environment without shutting down.

[0044] Because randomly generated training data has uncertainty, such as the combined direction of each side weights exceeding the weight of a single side, or negative weights, which have no physical meaning, the trained model is prone to fitting bias, affecting the accuracy of subsequent obstacle avoidance direction prediction.

[0045] For this purpose, please refer to Figure 2 In one specific embodiment, step S1 specifically includes:

[0046] Define the set of vertex coordinates of the outer contour of the mobile robot as follows: ;

[0047] Let (x, y) be the coordinates of any point on the plane, and let the system of inequalities be used. Represents the boundary lines of the outer contour of the mobile robot, where l is the number of sides, (a k b k c k ) represents the boundary line parameter, a k With b k Let c represent the components of the normal direction of the k-th boundary line on the x-axis and y-axis, respectively. k This represents the offset of the k-th boundary line;

[0048] Obstacle points The problem of finding the optimal distance weights to the outer contour of the mobile robot is modeled as follows: under non-negative constraints w k ≥0 and normalization constraints Below, request Maximize the weight w k The w k This is the optimal distance weight for the k-th boundary.

[0049] Understandably, the aforementioned set of inequalities transforms obstacle avoidance distance calculation from an empirical problem dependent on environmental perception into a deterministic optimization problem determined solely by geometric parameters. The non-negativity constraint in the optimization problem ensures that the weights, representing contributions, cannot be negative; the normalization constraint prevents the combined direction resulting from the weights on each side from being too long. This guarantees the correctness of the distance calculation, ensuring that the trained model will not exhibit fitting bias and improving the accuracy of subsequent obstacle avoidance direction predictions.

[0050] For single-axis or dual-axis mobile robots with simple movements, the outer contours of each shape can be obtained by decomposing each shape and then performing shape simulation processing on the edge contour of the top-view projection area of ​​each shape. However, for multi-axis robots, such as existing humanoid robots which typically have more than 32 axes, their shapes cannot be exhaustively listed. However, the solution in this application relies on the outer contour of any shape to establish the vertex coordinate set and boundary line inequality representation of the mobile robot's outer contour. It is evident that the method of obtaining its geometric parameters by pre-setting the outer contours of each shape is not suitable for multi-axis mobile robots.

[0051] For this purpose, please refer to Figure 3 In one specific embodiment, obtaining the outer contour of the current form of the mobile robot includes the following steps:

[0052] Acquire images of the robot from different perspectives, and stitch these images together to obtain a panoramic image;

[0053] The panoramic image is segmented to obtain the top-down projection area of ​​the mobile robot;

[0054] Obtain the first minimum rectangle of the top-view projection area, and map the mask of the top-view projection area into the first minimum rectangle to obtain several redundant areas;

[0055] In each redundant region, a first maximum rectangle is obtained, and the mask of all the first maximum rectangle regions is mapped to the first minimum rectangle to obtain the simulated contour region.

[0056] It is understandable that multi-axis robots have an inexhaustible range of shapes, meaning they cannot be derived from preset outlines. The specific embodiments described above enable the automatic acquisition of robot outlines in any shape, and the acquisition process does not rely on data feedback from joint motion parameters, making it suitable for multi-axis mobile robots.

[0057] By employing two consecutive rectangular bounding box acquisition and mapping strategies, the simulated contour region can be geometrically regularized while minimizing its area. Regularized region geometry ensures that it can be used to establish a system of boundary line inequalities, while minimizing the area can improve the robot's flexibility.

[0058] Preferably, after obtaining the top-view projection region of the mobile robot, morphological processing is performed to make the edges of the segmentation result relatively regular and complete. If the obtained redundant region is too small, it is directly included in the simulated contour region to reduce the amount of computation.

[0059] Compared to existing technologies, this method solves the problem that multi-axis mobile robots cannot pre-enumerate all shapes and contours, making it difficult to obtain accurate geometric parameters of the current shape in real time. At the same time, this block rectangle fitting simulation contour method does not significantly increase the number of vertices and sides of the outer contour, controlling the computational load of subsequent mathematical solutions and network inference, and still meeting the requirements of low-latency real-time computing.

[0060] In an alternative solution, the top-view projection area of ​​the mobile robot can also be obtained by first acquiring a top view using a camera positioned at a certain height on top of the mobile robot with a top-view perspective, and then segmenting the top view to obtain the top-view projection area of ​​the mobile robot.

[0061] For multi-axis robots, due to their diverse shapes, when they are unfolded (such as a cleaning robot with front side brushes unfolded in working state, or a humanoid robot with its arms unfolded), the simulated contour area obtained by the above method as the outer contour of the current shape may have an actual area that is too large. This will lead to the calculated safe distance being too large, making the robot too conservative and unable to pass through narrow spaces.

[0062] For further details, please refer to Figure 4 Within the simulated contour region, the second maximum bounding box is obtained. The mask of the second maximum bounding box region is mapped onto the simulated contour region to obtain several edge regions. The second minimum bounding box of each edge region is then obtained.

[0063] The second largest rectangle and several second smallest rectangles are independently used as the outer contour of the current form of the mobile robot, and the optimal distance weight is solved. The optimal distance weight of the outer contour with the shortest distance is taken as the true optimal distance weight.

[0064] It is understandable that since the simulated contour region is obtained entirely by continuously obtaining rectangular boxes, the strategy of continuously obtaining rectangular boxes is unified again in the above steps to obtain a second maximum rectangular box and several second minimum rectangular boxes for the simulated contour region. The whole process is the same in that it uses the method of obtaining rectangular boxes. The difference is that the specific strategies for obtaining rectangular boxes are different, so as to realize the background removal and splitting process.

[0065] Consequently, the outer contour of each independent mobile robot's current form obtained in this way is a regular rectangle shape, eliminating irregular lines, curves, and diagonal lines. At the same time, it reduces the number of times the optimal distance weight is calculated and the total amount of computation, perfectly matching the actual optimal distance weight calculation process, comprehensively improving the solution efficiency and reducing latency.

[0066] For further details, please refer to Figure 5 After obtaining the true optimal distance weights, the following steps are also included:

[0067] The relative positional relationship between the main body of the mobile robot and each movable arm is predefined. Taking the second largest rectangle as the main body, the second smallest rectangle is associated with the movable arm in a one-to-one correspondence according to the relative positional relationship.

[0068] When it is predicted that the shortest distance corresponding to the actual optimal distance weight will continue to be less than the threshold of the shortest distance in the obstacle avoidance direction within a preset future time period, the mobile robot actively controls the corresponding movable arm to retract and simultaneously updates the geometric parameters of the second minimum rectangle corresponding to the movable arm, and re-solves the optimal distance weight based on the new geometric parameters.

[0069] Understandably, conventional robots can typically be broken down into a main body and multiple movable arms. For example, a large cleaning robot can be divided into a chassis (main body) and side brushes that extend to the front on both sides when in operation (movable arms). The left side brush is located on the front left side of the chassis, and the right side brush is located on the front right side of the chassis, forming an inherent relative positional relationship. Similarly, a humanoid robot can be divided into a torso (main body) and limbs. The left robotic arm is located on the left side of the torso, the right robotic arm on the right side, the left leg robotic arm on the left front or left rear side of the torso, and the right leg robotic arm on the right front or right rear side of the torso, forming an inherent relative positional relationship.

[0070] Therefore, the process of obtaining the second largest rectangle and several second smallest rectangles is essentially a detection process of the main body of the mobile robot and its multiple movable arms. Furthermore, when the shortest distance corresponding to the true optimal distance weight is less than the threshold for the shortest distance in the obstacle avoidance direction, the mobile robot can directly control the corresponding movable arm to retract, thereby re-acquiring the outer contour of the current shape and calculating the optimal distance weight until the shortest distance meets the obstacle avoidance requirements. This ensures the robot's ability to pass through narrow spaces while avoiding task stagnation due to overly conservative obstacle avoidance, further improving obstacle avoidance flexibility and passage efficiency in complex scenarios.

[0071] Furthermore, the generation of the training dataset in step S2 specifically includes:

[0072] A rectangular working area is pre-defined around the mobile robot. N sampling obstacle points (x) are generated uniformly and randomly within the area. i ,y i ), where i = 1, 2, ..., N;

[0073] For each sampling point, solve for the optimal distance weights of each boundary under non-negativity and normalization constraints. As labels, a training dataset of pure geometric relationships is formed, determined solely by the geometric parameters of the mobile robot.

[0074] Understandably, since the generated labels reflect the constant geometric relationship of distance weights from spatial points to fixed polygons, the dataset naturally does not contain any environment-specific information. Therefore, the trained model does not have any deviation between simulation and reality and can be directly deployed and used in any scenario.

[0075] In the specific embodiment of step 2 above, nonnegativity constraints and normalization constraints are introduced to ensure the correctness of distance calculation in the training data. However, fitting bias may still occur when the training data is input into the architecture of existing mainstream distance neural network models.

[0076] For this purpose, please refer to Figure 6 In one specific embodiment, the distance neural network model described in step S3 includes alternately stacked constrained bounded layers and non-negative constrained layers:

[0077] The constrained bounded layer sequentially performs linear transformation, layer normalization, and hyperbolic tangent function activation on the input to make the output strictly bounded, corresponding to the normalization constraint.

[0078] The non-negativity constraint layer sequentially performs linear transformations and corrects the activation of linear units on the input, and makes the output naturally non-negative by truncating negative values, which corresponds to the non-negativity constraint.

[0079] J-round constraint bounded layer and non-negative constraint layer are stacked alternately;

[0080] The number of network layers is determined by the convergence cycle of the mathematical problem being solved, and the width of each layer is determined by the dimension of the intermediate variables.

[0081] Understandably, by designing a network structure with alternating layers of bounded constraints and non-negative constraints, each layer directly corresponds to the specific constraints of the mathematical problem in step S1, forming a correspondence with the dual constraints of the training data. This transforms the network from an uninterpretable black box, giving each layer a clear physical meaning, facilitating debugging and improvement. Furthermore, the physical constraints are guaranteed by the network structure itself rather than through additional post-processing, fundamentally avoiding erroneous outputs that violate physical laws, and significantly improving the safety and reliability of the obstacle avoidance system.

[0082] Furthermore, the specific formulas for nonnegativity constraints and normalization constraints have been described in the specific embodiments of step S1 above, and will not be detailed here.

[0083] For further details, please refer to Figure 6 The distance neural network model is a multi-head network structure:

[0084] Each mobile robot is described by its own set of boundary line inequalities by decomposing its outer contour sub-body.

[0085] The multi-head network structure includes a shared layer and multiple output heads. The shared layer includes a structure of alternating stacked constrained and non-negative constraint layers, used to extract obstacle point features. Multiple independent output heads are connected to the back, and each output head corresponds to an outer contour sub-body. The optimal distance weights of each boundary prediction of the outer contour sub-body are output in the manner of the non-negative constraint layer.

[0086] During inference, the distance from the obstacle point to each outer contour sub-body is calculated separately, and the minimum value is taken as the shortest distance to the mobile robot. The weight of the corresponding outer contour sub-body is used as the effective obstacle avoidance information.

[0087] It is understood that the splitting method can be an existing geometric splitting method (such as triangulation) or the method for continuously obtaining rectangular boxes provided in this application, which aims to decompose a large outline into multiple outer outline sub-bodies.

[0088] The method of predicting and taking the shortest distance for each segment of the contour after splitting not only adapts to the contour representation method of splitting into multiple blocks, giving full play to the advantages of the block contour reducing the area enclosed by the outer contour and improving the passage capacity, but also does not require too much additional inference time. The design of sharing the front features of the network also controls the overall number of parameters and computation, while still ensuring low-latency real-time inference.

[0089] Furthermore, the loss function used in step S3 when training the distance neural network model includes at least the first term, weight error L. w : Measures the mean squared error between the network's predicted weights and the label weights; the formula is:

[0090]

[0091] Where, N b This indicates the number of samples used in each training session. Let represent the standard optimal distance weight calculated based on the i-th sample and the k-th edge. Let represent the network prediction weights calculated based on the i-th sample and the k-th edge. This represents the squared error between the standard optimal distance weights and the predicted network weights. This means that all N in the batch will be included. b The errors of each sample are summed. This means summing the squared errors of all l edges of the same sample.

[0092] Understandably, by directly aligning the network's predicted weights with the standard optimal distance weights obtained from geometric calculations in the labels, the network's learning objective becomes clear, meeting the training requirements of a pure geometric relation dataset. This directly constrains the fitting direction, effectively reduces training errors, improves model prediction accuracy, and ensures the reliability of subsequent obstacle avoidance distance calculation results.

[0093] Please see Figure 7 The mobile robot 100 provided in this application embodiment will now be described. The mobile robot 100 includes a vehicle body 101, sensors 102, a motion controller, a computing unit, and side brushes 103 and / or squeegees 104 movably mounted on the vehicle body 101. The computing unit of the mobile robot 100 is equipped with a distance neural network model trained by the above-described adaptive obstacle avoidance control method for dense obstacle environments. Based on the data from the sensors 102, the model is used in real time during autonomous navigation to calculate the shortest distance from each obstacle point in the surrounding environment to the outer contour of the mobile robot 100 and the obstacle avoidance direction, so that the motion controller can generate a safe movement trajectory.

[0094] Understandably, the mobile robot 100 applies the aforementioned adaptive obstacle avoidance control method for dense obstacle environments. This method not only solves the problems of multi-axis robot morphology not being able to be pre-defined and exhaustively enumerated, and difficulty in obtaining the current contour geometric parameters in real time, but also ensures the accuracy of obstacle avoidance prediction through the network structure design incorporating physical constraints and multi-task loss functions. At the same time, it controls the amount of computation throughout the process, which can meet the requirements of low-latency real-time computing. It can also take into account both passage safety and passage efficiency in narrow and complex scenarios. Compared with traditional obstacle avoidance solutions, it has stronger adaptability and better robustness.

[0095] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. An adaptive obstacle avoidance control method for dense obstacle environments, characterized in that, Includes the following steps: Step S1: Obtain the geometric parameters of the outer contour of the current form of the mobile robot, establish the vertex coordinate set and boundary line inequality representation of the outer contour of the mobile robot, and update the mathematical solution problem of the optimal distance weight from the obstacle point to the outer contour of the mobile robot. The process of obtaining the outer contour of the current form of the mobile robot includes the following steps: first, obtaining the top-view projection area of ​​the mobile robot; obtaining the first minimum rectangle of the top-view projection area, mapping the mask of the top-view projection area onto the first minimum rectangle to obtain several redundant areas; obtaining the first maximum rectangle in each redundant area, mapping the mask of all the first maximum rectangle areas onto the first minimum rectangle to obtain the simulated contour area. Define the set of vertex coordinates of the outer contour of the mobile robot as follows: ; Let (x, y) be the coordinates of any point on the plane, and let the system of inequalities be used. Represents the boundary lines of the outer contour of the mobile robot, where l is the number of sides, (a k b k c k ) represents the boundary line parameter, a k With b k Let c represent the components of the normal direction of the k-th boundary line on the x-axis and y-axis, respectively. k This represents the offset of the k-th boundary line; Obstacle points The problem of finding the optimal distance weights to the outer contour of the mobile robot is modeled as follows: under non-negative constraints w k ≥0 and normalization constraints Below, request Maximize the weight w k The w k This is the optimal distance weight for the k-th boundary; The second maximum bounding box is obtained within the simulated contour region. The mask of the second maximum bounding box region is mapped onto the simulated contour region to obtain several edge regions. The second minimum bounding box of each edge region is then obtained. The second largest rectangle and several second smallest rectangles are independently used as the outer contour of the current form of the mobile robot, and the optimal distance weight is solved. The optimal distance weight of the outer contour with the shortest distance is taken as the true optimal distance weight. Step S2: Randomly generate the coordinates of sampling obstacle points within the preset working area around the mobile robot, calculate the optimal distance weights of each sampling point for each boundary, and form a training dataset; Step S3: Iteratively train the distance neural network model using the training dataset; Step S4: Input the real-time perceived obstacle point coordinates into the trained distance neural network model, and output the predicted optimal distance weights in real time to output the obstacle avoidance direction; Step S5: Repeat steps S1-S4 above.

2. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 1, characterized in that, Obtaining the top-view projection area of ​​the mobile robot includes the following steps: Acquire images of the robot from different perspectives, and stitch these images together to obtain a panoramic image; The panoramic image is segmented to obtain the top-down projection area of ​​the mobile robot.

3. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 1, characterized in that, After obtaining the true optimal distance weights, the following steps are also included: The relative positional relationship between the main body of the mobile robot and each movable arm is predefined. Taking the second largest rectangle as the main body, the second smallest rectangle is associated with the movable arm in a one-to-one correspondence according to the relative positional relationship. When it is predicted that the shortest distance corresponding to the actual optimal distance weight will continue to be less than the threshold of the shortest distance in the obstacle avoidance direction within a preset future time period, the mobile robot actively controls the corresponding movable arm to retract and simultaneously updates the geometric parameters of the second minimum rectangle corresponding to the movable arm, and re-solves the optimal distance weight based on the new geometric parameters.

4. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 2 or 3, characterized in that, Step S2, which involves forming the training dataset, specifically includes: A rectangular working area is pre-defined around the mobile robot. N sampling obstacle points (x) are generated uniformly and randomly within the area. i ,y i ), where i = 1, 2, ..., N; For each sampling point, solve for the optimal distance weights of each boundary under non-negativity and normalization constraints. As labels, a training dataset of pure geometric relationships is formed, determined solely by the geometric parameters of the mobile robot.

5. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 4, characterized in that, The distance neural network model described in step S3 includes alternating stacked constrained bounded layers and non-negative constrained layers: The constrained bounded layer sequentially performs linear transformation, layer normalization, and hyperbolic tangent function activation on the input to make the output strictly bounded, corresponding to the normalization constraint. The non-negativity constraint layer sequentially performs linear transformations and corrects the activation of linear units on the input, and makes the output naturally non-negative by truncating negative values, which corresponds to the non-negativity constraint. J-round constraint bounded layer and non-negative constraint layer are stacked alternately; The number of network layers is determined by the convergence cycle of the mathematical problem being solved, and the width of each layer is determined by the dimension of the intermediate variables.

6. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 5, characterized in that, The distance neural network model is a multi-head network structure: Each mobile robot is described by its own set of boundary line inequalities by decomposing its outer contour sub-body. The multi-head network structure includes a shared layer and multiple output heads. The shared layer includes a structure of alternating stacked constrained and non-negative constraint layers, used to extract obstacle point features. Multiple independent output heads are connected to the back, and each output head corresponds to an outer contour sub-body. The optimal distance weights of each boundary prediction of the outer contour sub-body are output in the manner of the non-negative constraint layer. During inference, the distance from the obstacle point to each outer contour sub-body is calculated separately, and the minimum value is taken as the shortest distance to the mobile robot. The weight of the corresponding outer contour sub-body is used as the effective obstacle avoidance information.

7. The adaptive obstacle avoidance control method for dense obstacle environments as described in claim 6, characterized in that, The loss function used in step S3 when training the distance neural network model includes at least the first term, weight error L. w : Measures the mean squared error between the network's predicted weights and the label weights; the formula is: Where, N b This indicates the number of samples used in each training session. Let represent the standard optimal distance weight calculated based on the i-th sample and the k-th edge. Let represent the network prediction weights calculated based on the i-th sample and the k-th edge. This represents the squared error between the standard optimal distance weights and the predicted network weights. This means that all N in the batch will be included. b The errors of each sample are summed. This means summing the squared errors of all l edges of the same sample.

8. A mobile robot, characterized in that, The mobile robot includes a vehicle body, sensors, a motion controller, a computing unit, and side brushes and / or squeegees that are movably mounted on the vehicle body. The computing unit of the mobile robot is equipped with a distance neural network model trained by the adaptive obstacle avoidance control method for dense obstacle environments as described in claim 1. Based on sensor data, the model is used in real time during autonomous navigation to calculate the shortest distance from each obstacle point in the surrounding area to the outer contour of the mobile robot and the obstacle avoidance direction, so that the motion controller can generate a safe movement trajectory.

Images