A humanoid robot navigation method, device and equipment based on defect complementation

Abstract

The application discloses a humanoid robot navigation method and device based on defect complementarity, and an equipment. Defect points and data hollow areas are measured and identified through multi-sensor defect feature extraction and recognition, and a defect feature atlas is generated. A confidence decreasing verification mechanism is established based on defect mutual exclusion analysis, sensor weight distribution is determined, and a navigation gain parameter is generated from a negative error feature set. Data back information is extracted through reverse analysis of navigation data, environment implicit features are reconstructed, and a reverse navigation field is constructed. A hollow density map is generated by analyzing the spatial distribution density of data hollow areas, a phase transition area of a path is divided, and a phase transition rule set is generated. An adaptive navigation sequence is generated by fusing a curve path parameter, a three-dimensional path parameter and a navigation gain parameter, and a humanoid robot is driven to realize intelligent navigation control. The application converts sensor defects into navigation resources, and improves navigation performance in a complex environment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot navigation control, in particular to a humanoid robot navigation method, device and equipment based on defect complementation. BACKGROUND

[0002] With the rapid development of artificial intelligence and robot technology, the application demand of humanoid robots in the fields of service, medical treatment, security and the like is increasing day by day. In actual application, the humanoid robot needs to perform autonomous mobile tasks in complex indoor environments such as office buildings, commercial places and hospital corridors, which puts forward very high requirements for its environment perception and path planning capabilities. Modern humanoid robots usually adopt a multi-modal perception system to integrate RGB-D cameras, laser range finders, gyroscopes and accelerometers and the like to obtain environment information, but the performance of various sensing devices under different environmental conditions has obvious differences.

[0003] The current mainstream robot navigation technology is mainly based on direct fusion processing of sensor data, and integrates multi-source perception information through filtering algorithms and probability estimation methods. However, when dealing with sensor performance degradation, this kind of method often adopts data rejection or weight reduction strategy, and fails to fully utilize the environment feature information contained behind the performance degradation. In real environments with drastic changes in light, complex material reflection characteristics and frequent interference of dynamic objects, sensor performance degradation phenomenon is common, and the limitations of traditional fusion methods are increasingly prominent. Especially under the influence of special environmental elements such as transparent partitions, mirror walls and carpet joints, the navigation strategy simply relying on high-quality sensing data is difficult to ensure the continuous and stable operation of the system. Therefore, it is urgent to improve the navigation reliability of humanoid robots in complex environments. SUMMARY

[0004] The present application provides a humanoid robot navigation method, device and equipment based on defect complementation, aiming to deeply mine the complementary characteristics of multi-sensor defects, convert the sensor defects in the traditional sense into valuable navigation information sources, construct a reverse navigation field and a path phase transition rule, realize high-precision and high-reliability autonomous navigation control of humanoid robots in complex environments, and provide technical support for the wide application of humanoid robots in complex indoor scenes such as office buildings, commercial places and home environments.

[0005] The present application provides a humanoid robot navigation method, device and equipment based on defect complementation, aiming to deeply mine the complementary characteristics of multi-sensor defects, convert the sensor defects in the traditional sense into valuable navigation information sources, construct a reverse navigation field and a path phase transition rule, realize high-precision and high-reliability autonomous navigation control of humanoid robots in complex environments, and provide technical support for the wide application of humanoid robots in complex indoor scenes such as office buildings, commercial places and home environments.

[0006] Obtaining humanoid robot navigation data, performing defect feature extraction and identification on the navigation data to measure defect points and data hollow areas, and generating a defect feature map based on spatial distribution analysis of the measured defect points and the data hollow areas;

[0007] The defect feature atlas is used to establish confidence decreasing verification to obtain multi-level perception results, sensor weight distribution is determined according to the multi-level perception results, a negative error feature set is generated based on negative error values of each sensor extracted from the defect feature atlas based on the sensor weight distribution, and a navigation gain parameter is generated using the negative error feature set;

[0008] Data back information is extracted by inversely analyzing the navigation data through the navigation gain parameter, environment implicit features are reconstructed based on the data back information, and an inverse navigation field is constructed in combination of the environment implicit features and the navigation gain parameter;

[0009] A hollow density map is generated based on the spatial distribution density of the data hollow area analyzed based on the inverse navigation field, hollow clustering features and hollow connectivity modes are extracted from the hollow density map, a path phase change region is divided according to the density threshold set according to the hollow density map, and a path phase change rule set is generated based on the path phase change region, the hollow clustering features and the hollow connectivity modes;

[0010] An initial navigation path is generated according to the path phase change rule set and the inverse navigation field, a curve path parameter and a three-dimensional path parameter are generated by performing morphological transformation on the initial navigation path, and an adaptive navigation sequence is generated by fusing the curve path parameter, the three-dimensional path parameter and the navigation gain parameter.

[0011] The second aspect of the present application proposes a humanoid robot navigation device based on defect complementarity, comprising:

[0012] A defect analysis module is configured to obtain humanoid robot navigation data, perform defect feature extraction and identification on the navigation data to measure defect points and data hollow areas, and generate a defect feature atlas based on spatial distribution analysis of the measured defect points and the data hollow areas;

[0013] A gain generation module is configured to use the defect feature atlas to establish confidence decreasing verification to obtain multi-level perception results, determine sensor weight distribution according to the multi-level perception results, generate a negative error feature set based on negative error values of each sensor extracted from the defect feature atlas based on the sensor weight distribution, and generate a navigation gain parameter using the negative error feature set;

[0014] An inverse field construction module is configured to extract data back information by inversely analyzing the navigation data through the navigation gain parameter, reconstruct environment implicit features based on the data back information, and construct an inverse navigation field in combination of the environment implicit features and the navigation gain parameter;

[0015] A phase change rule module is configured to analyze spatial distribution density of the data empty area based on the reverse navigation field, generate an empty density map, extract empty cluster features and empty connectivity patterns from the empty density map, set a density threshold based on the empty density map to divide a path phase change area, and generate a path phase change rule set based on the path phase change area, the empty cluster features and the empty connectivity patterns.

[0016] A path generation module is configured to generate an initial navigation path based on the path phase change rule set and the reverse navigation field, perform shape transformation on the initial navigation path to generate curve path parameters and three-dimensional path parameters, and fuse the curve path parameters, the three-dimensional path parameters and the navigation gain parameters to generate an adaptive navigation sequence.

[0017] The third aspect of the present application provides a computer device comprising a memory, a processor and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the navigation method based on complementary defects of a humanoid robot disclosed in the first aspect when executing the program.

[0018] The present application has the following advantages: first, through defect feature map analysis and confidence decreasing verification mechanism, the active use and performance complement of sensor defects are realized. Confidence decreasing verification adopts a three-layer strategy, so that reliable perception can be achieved through multi-sensor cooperation even if the performance of a single sensor is poor, and the perception defects in the traditional sense are converted into valuable navigation information sources. The extraction of negative error feature set and the generation of navigation gain parameters effectively compensate for the systematic deviation of the sensor and improve the navigation accuracy of the humanoid robot in multi-material environments such as office buildings and commercial places. Second, a reverse navigation field is constructed to realize the deep mining of data back information and the reconstruction of environmental implicit features. Through reverse analysis of navigation data, information such as sensor failure position, abnormal distribution of measurement, data empty shape and other information is converted into the basis for judging potential passable areas. The reverse navigation field not only retains the navigation value of normal perception data, but also converts abnormal patterns and empty distribution into gain information for path planning, significantly improving the environmental adaptability. Finally, based on the empty density map, the environment is divided into strong phase change area, weak phase change area and normal area, and respectively adopts avoidance strategy, deceleration and enhanced perception strategy and high-efficiency passable mode. The path phase change rule set combines the empty cluster features and the connectivity patterns to realize fine navigation control. Through the Bezier control point optimization of the curve path parameters and the terrain undulation adaptation of the three-dimensional path parameters, the generated adaptive navigation sequence can adjust the motion control parameters in real time according to the environmental complexity, ensuring that the humanoid robot can safely and stably complete the navigation task in narrow passages, carpet joints and step edges

[0019] It should be understood that the general description and detailed description below are only exemplary and explanatory, and do not limit the present application. BRIEF DESCRIPTION OF DRAWINGS

[0020] The drawings here show specific examples of the technical solutions described in the present application, and constitute part of the specification together with the specific embodiments, for explaining the technical solutions, principles and effects of the present application.

[0021] Unless specifically stated or otherwise defined, the same reference signs in different drawings represent the same or similar technical features, and different reference signs may also be used to represent the same or similar technical features.

[0022] Figure 1 is a flow diagram of a human-shaped robot navigation method based on defect complementation according to the present application.

[0023] Figure 2 is a structural block diagram of a human-shaped robot navigation device based on defect complementation according to the present application.

[0024] Figure 3 is a structural diagram of a computer device according to the present application. DETAILED DESCRIPTION

[0025] In the following description, specific details are set forth in order to provide a thorough understanding of the embodiments of the present application. However, persons skilled in the art will understand that the present application can be practiced without these specific details, in other instances, well-known structures, devices, circuits, and methods have not been described in detail in order not to obscure the description of the present application.

[0026] It should be understood that when used in the specification and the appended claims of the present application, the term "comprising" indicates the presence of the described features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.

[0027] It should also be understood that the term "and / or" as used in the specification and the appended claims of the present application means any combination of one or more of the associated listed items and all possible combinations, and includes these combinations.

[0028] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0029] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0030] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0031] The technical solutions of the embodiments of this application will be described below.

[0032] like Figure 1 As shown, this embodiment of the invention provides a humanoid robot navigation method based on defect complementarity, including the following steps S110-S150:

[0033] Step S110: Obtain navigation data of the humanoid robot, extract and identify defect features based on the navigation data, measure defect points and data void areas, and generate a defect feature map based on spatial distribution analysis of the measured defect points and data void areas.

[0034] Specifically, the navigation perception data of the humanoid robot is synchronously collected by a multi-sensor fusion architecture. The visual sensor adopts a binocular camera configuration to obtain RGB-D image data at a frame rate of 30 fps, with a resolution of 1280 x 720 and a field of view angle of 120°. The laser radar sensor adopts a 16-line configuration, with a scanning frequency of 10 Hz, an angular resolution of 0.2°, and a maximum detection distance of 100 meters. The effective distance in an indoor environment is about 30 meters. The inertial sensor IMU has a sampling frequency of 200 Hz and includes a three-axis accelerometer and a three-axis gyroscope. For example, when the service robot "Xiaozhi" performs a navigation task in an office building, the visual sensor can normally recognize the reflected image in front of the glass elevator door, but the laser radar appears to be penetrating, resulting in ranging failure. In the dark area of the corridor, the visual depth information is missing, while the laser radar remains normal. In a dynamic crowd, the IMU is disturbed by collisions and produces transient abnormal readings. The data synchronization mechanism is based on hardware triggering and software timestamp calibration, ensuring that the time alignment error of different sensor data is less than 5 ms. Sensor calibration uses Zhang Zhengyou calibration method for camera intrinsic parameter calibration, and laser radar and camera extrinsic parameter calibration is achieved by matching the feature points of the checkerboard. Through the cooperative collection and preprocessing of multiple heterogeneous sensors, raw navigation data containing visual, laser, and inertial information are obtained.

[0035] Based on the collected multi-sensor navigation data, feature extraction algorithms are used to identify various measurement defects and data voids. Visual defect recognition mainly targets image blurring, occlusion, and lighting abnormalities, using a gradient threshold method to detect blurred areas. When the gradient amplitude G = √(Gx² + Gy²) < τ_blur, it is marked as a blurred defect point, where G is the gradient amplitude, Gx is the x-direction gradient, Gy is the y-direction gradient, and τ_blur is the blur judgment threshold (value 10). Laser radar defect recognition focuses on ranging abnormalities and reflection failures, and detects measurement defects by detecting distance mutations between adjacent points. When the distance change exceeds the statistical threshold, it is determined to be an abnormal point. Data voids are defined as areas where sensors cannot obtain effective measurements. Visual voids appear as invalid pixels in depth maps, and laser voids appear as discontinuous segments of scan lines. For example, in the glass partition area of the office building, the laser voids present a rectangular distribution of 2m x 1.5m, with a void area of 3m², while the visual sensor can normally detect the object contour behind the glass. The void detection algorithm uses connected component analysis to aggregate adjacent missing points into void regions, and marks the void when the void area A_hole > 0.05m². The defects of the inertial sensor are mainly noise and drift, and noise features are extracted through statistical analysis. When the drift rate exceeds 0.1° / s, it is marked as a defect state. The severity of the defect is divided into three levels: mild, moderate, and severe. The visual blur area ratio is less than 10% for mild, 10%-30% for moderate, and more than 30% for severe.

[0036] The identified defect points and hollow areas are analyzed for spatial distribution to construct a defect feature map reflecting the environmental perception characteristics. The spatial distribution analysis first divides the environment around the robot into a grid map, with a grid resolution of 0.1 m x 0.1 m and a coverage range of 20 m x 20 m around the robot center. For measuring defect points, the point density distribution ρ_point = N_defect / A_grid is calculated, where ρ_point is the defect point density, N_defect is the number of defect points in the grid, and A_grid is the grid area (0.01 m²), and the defect type distribution ratio is counted. For data hollow areas, the area distribution, shape characteristics, and connectivity are analyzed, and the minimum circumscribed rectangle is used to describe large hollows, recording the center position, principal axis direction, and coverage range. For example, in the glass partition area of an office building, the laser defect density reaches 15 / m², and the visual defect points are gradiently distributed in the light transition area, with a density decreasing from 80 points / m² to 10 points / m². In the central open area of the corridor, the defect density is only 2-3 / m². Cross-sensor defect correlation analysis uses the defect mutual exclusion index M_ij = 1-(ρ_i∩ρ_j) / (ρ_i∪ρ_j), where M_ij is the mutual exclusion of sensors i and j, ρ_i and ρ_j represent the defect density distribution of sensors i and j, ρ_i∩ρ_j is the defect overlap area, and ρ_i∪ρ_j is the defect union area. When M_ij > 0.7, it is considered that the two sensors have strong complementarity. The defect evolution mode is obtained through spatiotemporal trajectory analysis, and the appearance, disappearance, and transfer rules of defects in the moving process provide important reference for path planning.

[0037] In step S120, a multi-level perception result is obtained by using the defect feature map to establish a confidence decreasing verification, the sensor weight distribution is determined according to the multi-level perception result, the negative error feature set of each sensor is generated based on the sensor weight distribution from the defect feature map, and the navigation gain parameter is generated by using the negative error feature set.

[0038] Specifically, the multi-level perception results are obtained by confidence decreasing verification using defect feature maps. First, the defect distribution differences of different sensors at the same spatial position are analyzed, and the defect mutual exclusion index M_ij defined in the foregoing step S110 is calculated. When M_ij>0.7, it is considered that the sensors have strong complementarity, and a defect complementary pair is formed. Based on the identified complementary pair, a quantitative complementary relationship of vision-laser and laser-inertia is established, and a multi-level defect complementary matrix is constructed. Then, the complementary matrix is used to design confidence decreasing verification, starting from the single sensor with the highest confidence, and gradually adding complementary sensors to verify the quality improvement effect of perception. The initial confidence is calculated according to the defect density C_init=1-ρ_defect, wherein C_init is the initial confidence of the sensor, and ρ_defect is the defect density of the sensor, and the value range is between 0 and 1. The multi-sensor combination confidence is obtained by complementary weighted fusion, which comprehensively considers the basic performance and complementary strength of each sensor, and the weighting coefficient is dynamically determined according to the complementary matrix. The verification adopts a three-layer strategy: the first layer verifies the single sensor perception (threshold value 0.8), the second layer verifies the double sensor complementarity (threshold value 0.6), and the third layer verifies the three sensor cooperation (threshold value 0.4). For example, the confidence of the vision sensor is only 0.3 when the light is insufficient, and the confidence of the laser radar is 0.7 in the glass reflection area, but through the weighted fusion of the high complementary strength 0.85, the combination confidence is improved to 0.489, which meets the third layer verification requirement. The multi-level perception results record the sensor combination, confidence value, dominant sensor and failure mode of each layer, and form a structured perception quality evaluation report. The verification process ensures that even if the performance of a single sensor is poor, reliable perception can be achieved through multi-sensor cooperation through a decreasing threshold, fully embodying the core idea of defect complementarity.

[0039] In some embodiments, the determining sensor weight distribution according to the multi-level perception results comprises: extracting sensor response strength and stability indicators from the multi-level perception results; calculating the contribution degree of the sensor at different confidence levels based on the response strength; generating a dynamic weight coefficient based on the stability indicator and the contribution degree, and normalizing the dynamic weight coefficient to generate the sensor weight distribution.

[0040] From the confidence data of multi-level perception results, the response strength and stability characteristics of each sensor are extracted. The response strength reflects the effective perception ability of the sensor in the current environment, and is calculated by considering the sensor confidence and noise level, where C_init(v) represents the initial confidence of the visual sensor v, and noise_level_v is the corresponding noise level coefficient. The response strength of the visual sensor is mainly affected by the lighting conditions, and can reach 0.9 under sufficient light, and decreases to 0.2 in dark environment; the response strength of the laser radar is related to the target reflection characteristics, and the response strength of the standard diffuse reflection surface is 0.8, and the mirror or transparent material decreases to 0.3; the response strength of the inertial sensor is relatively stable, and usually remains in the range of 0.7-0.9. The stability index is obtained by calculating the time variance of the confidence, and the smaller the variance, the higher the stability. A sliding time window is used for real-time updating. For example, when "Xiao Zhi" performs a patrol task, the visual stability in a static environment is as high as 0.85, and the IMU stability in a dynamic scene is better, up to 0.9. The contribution degree is defined as the ratio of the confidence improvement after the sensor is added to the overall confidence, which reflects the marginal contribution of each sensor to the overall perception quality. The joint distribution of response strength and stability presents a specific pattern, high response strength usually accompanies high stability, but in the working boundary condition of the sensor, high response and low stability may occur, which needs special processing.

[0041] Based on the extracted response strength, the actual contribution of each sensor at different confidence levels is calculated. The contribution degree takes the response strength as the modulation factor, and defines the improvement of the overall confidence after the sensor is added. The contribution of the kth layer is calculated by modulating the confidence improvement of the layer by the response strength, which includes the difference between the combined confidence containing the visual sensor and the combined confidence not containing the visual sensor. For example, in the visual-laser combination, the marginal contribution of the visual sensor is modulated when its response strength is 0.3, ensuring that low-performance sensors do not excessively affect the overall evaluation. The comprehensive contribution degree is obtained by hierarchical weighted summation, with hierarchical weights λ_1=0.5, λ_2=0.3, λ_3=0.2, ensuring that the basic perception ability is valued, while taking into account the multi-sensor cooperation effect. The time-varying characteristics of the contribution degree are tracked by a sliding window, with a window size of 20 sampling periods, about 2 seconds of time span, to update the contribution evaluation of each sensor in real time, adapting to the dynamic changes of the environment.

[0042] The dynamic weight coefficient is calculated according to the comprehensive stability index and the contribution degree information, and reflects the comprehensive performance of the sensor. The dynamic weight coefficient calculation formula is W_v=α×R_v+β×S_v+γ×G_v, wherein W_v is the dynamic weight coefficient of the visual sensor, R_v is the response intensity, S_v is the stability index, G_v is the comprehensive contribution degree, the weight factors α=0.4, β=0.3 and γ=0.3 balance the real-time performance and the long-term reliability. The influence of the complementary matrix is indirectly transmitted to the weight calculation through the contribution degree G_i, and the sensor combination with high complementarity is ensured to obtain higher weight distribution. The dynamic adjustment mechanism updates the weight in real time according to the environmental changes, and when the environmental feature change (such as light mutation and material switching) is detected, the weight is quickly adjusted, and the adjustment response time is less than 100 ms. The smooth updating of the weight is realized by using the exponential moving average method, and the smooth factor η=0.8 ensures the timeliness of the weight adjustment and avoids the system shock caused by the weight mutation. The dynamic weight coefficient obtained is normalized to ensure that the sum of the weights of all sensors is 1, and the mathematical consistency of the multi-sensor fusion is ensured.

[0043] According to the determined sensor weight distribution, the negative error feature set of each sensor is extracted from the foregoing defect feature spectrum. The negative error is defined as the error of the systematic small measurement value of the sensor, and the calculation formula is e_i=d_measured-d_true, wherein e_i is the negative error of the sensor i, d_measured is the measured value, d_true is the true value, and the obtained negative value reflects the small measurement degree. The visual negative error is mainly caused by the systematic deviation of the depth estimation, and the depth value is small in the sparse texture area, and the average negative error is about-5% to-10% of the true distance. The laser negative error appears on the low reflectivity surface, and the weak return signal causes the ranging to be small, and the negative error of the black surface can reach-15%. For example, when "Xiao Zhi" detects the black carpet of the office building, the laser ranging system is small by 0.3 meters, and the visual depth estimation is normal in the area. The comprehensive negative error weighted by the weight is reflected by the weighted sum after the multi-sensor fusion, wherein w_i is the normalized weight, and the negative error of the high-weight sensor is focused on. The spatial distribution feature is extracted by kernel density estimation, and a negative error heat map is generated to identify the high-risk area and distribution mode of the negative error. The negative error feature set includes statistical features (mean, variance, skewness, kurtosis), spatial features (distribution density, aggregation degree) and time sequence features (trend, periodicity), which provide comprehensive error information for subsequent navigation gain parameter generation.

[0044] In some embodiments, the generating the navigation gain parameter using the negative error feature set comprises: decomposing a positioning negative bias component and a ranging negative error component based on the negative error feature set; generating a path correction gain based on the positioning negative bias component; generating an obstacle avoidance safety margin based on the ranging negative error component; and fusing the path correction gain and the obstacle avoidance safety margin to generate the navigation gain parameter.

[0045] Firstly, two main error components affecting positioning and ranging are decomposed from the negative error feature set. The positioning negative bias component is mainly caused by visual depth underestimation and laser long-distance attenuation, which is extracted by analyzing the spatial cumulative characteristics of negative error. The decomposition method adopts principal component analysis (PCA), and the first principal component usually corresponds to the positioning bias, with a contribution rate of about 60-70%, and the second principal component corresponds to the ranging error, with a contribution rate of about 20-30%. The ranging negative error component directly affects the judgment of obstacle distance, and different error models are used for different distance segments by modeling the error characteristics. For example, when "Xiaozhi" navigates in the long corridor of the office building, the positioning negative bias shows a cumulative characteristic, with a cumulative error of about 0.2 meters per 10 meters, while the ranging negative error shows an instantaneous characteristic when avoiding obstacles at close range, with an error fluctuation in the range of ±0.1 meters. The statistical characteristics of the error components include mean, standard deviation and distribution form, and the normality assumption of the error distribution is verified by K-S test. Through error source analysis and feature decomposition, the positioning negative bias component and the ranging negative error component are accurately separated.

[0046] Then, based on the statistical characteristics of the positioning negative bias component, the path correction gain is designed to compensate for the systematic position underestimation. The path correction gain k_path=1-δ_p / d_nominal, where k_path is the path correction gain, δ_p is the positioning negative bias (negative value), and d_nominal is the nominal motion distance. The gain value is dynamically adjusted according to the size of the negative bias and the motion distance to ensure the accuracy and stability of the compensation. The fixed gain compensation strategy is used for short distance motion (<5m), with a gain value of about 1.05-1.1, and the adaptive gain is used for long distance motion to avoid overcompensation, with a gain upper limit of 1.15. The spatial variation characteristics of the gain are realized through scene recognition, with a larger gain in corridor environment and a smaller gain in open space. Different environmental types correspond to different gain strategy curves. For example, when "Xiaozhi" performs a long-distance navigation task, the path correction gain is automatically adjusted according to the detected positioning negative bias characteristics, and the cumulative error is controlled within 0.5 meters in a 30-meter corridor navigation, ensuring the accuracy of reaching the target position. The gain limiting mechanism prevents excessive correction in abnormal situations, with upper and lower limits of 1.2 and 0.9 respectively, maintaining stable operation of the navigation process.

[0047] Then, according to the distribution characteristics of the ranging negative error component, the obstacle avoidance margin parameter ensuring navigation safety is generated. The basic obstacle avoidance margin calculation formula is margin_static=k_safe*max(|μ_e|,d_min), wherein margin_static is the static obstacle avoidance margin, k_safe is the safety coefficient (typical value 1.5-2.0), μ_e is the ranging negative error mean, and d_min is the minimum margin (0.3m). The dynamic margin considers the robot motion speed, and the calculation formula contains the current speed v and the reaction time t_reaction (typical value 0.5s), ensuring that there is enough braking distance in an emergency. The comprehensive obstacle avoidance margin combines static and dynamic factors, and adopts the maximum value principle to ensure safety in different motion states. For example, when "Xiaozhi" detects the black carpet and other areas prone to ranging negative errors, the corresponding safety margin is automatically increased to 0.5m, and the standard margin of 0.3m is maintained in normal areas, effectively avoiding collision risks. The sensor confidence correction adjusts the margin size according to the real-time perception quality, and appropriately increases the safety distance when the confidence is low. The confidence and the margin are inversely proportional, ensuring that safety is prioritized.

[0048] Finally, the path correction gain and the obstacle avoidance safety margin are fused to generate the navigation gain parameter. The navigation gain parameter includes the path correction gain, the static obstacle avoidance margin, the dynamic obstacle avoidance margin, and the scene weight, forming a complete compensation and safety information system, wherein the scene weight value range is 0.2-1.0. The parameter fusion considers the mutual constraint relationship, and appropriately increases the safety margin when the path gain is too large, ensuring overall coordination. The scene weight is dynamically allocated according to the environmental complexity, and the weight in narrow space is 0.8-1.0, and the weight in open area is 0.2-0.4, balancing efficiency and safety demand. For example, when "Xiaozhi" performs complex environment navigation, the path correction gain is 1.08, the static margin is 0.4m, the dynamic margin is 0.2m, and the scene weight is 0.7, forming a coordinated cooperation, reflecting the differences in navigation strategies under different environmental conditions. The consistency check of the gain parameter prevents parameter conflicts, and ensures that each parameter is within a reasonable range through constraint condition verification, ensuring stable operation of the navigation process. The navigation gain parameter can effectively compensate for the systematic deviation of the sensor, and at the same time provide reliable collision protection through dynamic safety margin adjustment. Through the optimization and fusion of various parameters and consistency verification, a complete set of navigation gain parameters is finally generated.

[0049] In step S130, the navigation data is inversely analyzed by the navigation gain parameter to extract data back information, the environment implicit features are reconstructed based on the data back information, and the reverse navigation field is constructed by combining the environment implicit features and the navigation gain parameter.

[0050] Specifically, the navigation gain parameters are utilized to reverse parse the original navigation data, and to mine the neglected implicit information in the sensor data. The navigation gain parameters include path correction gain k_path, obstacle avoidance safety margin margin, and scene weight w_scene, which reflect the systematic bias characteristics of the sensor. The core idea of reverse parsing is to infer the real characteristics of the environment from the error patterns, for example, the continuous underestimation of visual depth implies the existence of environmental factors affecting depth estimation. The parsing process first compensates the original measurement value with gain, and the compensated measurement value d_compensated is calculated by the formula d_compensated=d_measured×k_path, where d_measured is the original measurement value, and k_path is used to compensate the systematic negative error. The data back information is defined as the set of abnormal patterns, missing areas and implicit clues contained in the original navigation data, including: sensor failure position and time, spatial distribution of measurement anomalies, shape characteristics of data voids, and complete record of historical trajectory. For example, in the "JiZhi" office building navigation, the data back information records the continuous ranging jump point of the laser radar on the glass wall, the depth missing pattern of the visual sensor in the dark light area, and the abnormal reading sequence of the IMU in the elevator. The time sequence reverse traces the historical data to analyze the trend of change, and identifies the time point and spatial position that leads to the decline of sensor performance.

[0051] In some embodiments, the reconstructing the environment implicit characteristics based on the data back information comprises: performing time sequence inversion on the data back information to obtain inverse characteristics of the historical trajectory; identifying data missing areas from the data back information and mapping them as potential passable areas; performing numerical change rate analysis on the data back information to infer the environment boundary position; and fusing the inverse characteristics, the potential passable areas and the environment boundary position to generate complete environment implicit characteristics.

[0052] For example, the performing time sequence inversion on the data back information to obtain inverse characteristics of the historical trajectory comprises: arranging the data back information in reverse order based on time stamp to generate a reverse time sequence, extracting adjacent time position difference based on the reverse time sequence to generate inverse motion vector; identifying direction mutation points and speed abnormal points based on the inverse motion vector, inferring environment constraint strength based on the distribution density of the direction mutation points and the speed abnormal points in space; and generating inverse characteristics of the historical trajectory based on the environment constraint strength.

[0053] Firstly, based on the data back information, a reverse time sequence is generated in reverse order of timestamp. The data back information contains the complete motion history and abnormal event records of the robot, and the reverse-ordered data sequence preserves the original timestamp. The position data at the time of sensor failure is extracted from the data back information, and these key time points constitute the nodes of the reverse sequence. Based on the reverse time sequence, the position difference between adjacent time points is extracted, and the reverse motion vector v_rev(t_i) is calculated, which contains three components of x-direction velocity, y-direction velocity and angular velocity, i.e. v_rev=[Δx / Δt,Δy / Δt,Δθ / Δt]^T. For example, the data back information of "Xiaozhi" shows that visual abnormalities occur continuously at a corridor corner, and the position change in this period is extracted through reverse sequence analysis to generate the corresponding reverse motion vector sequence. The spatial distribution of the vector field is obtained by weighted averaging of all vectors passing through the same grid, and the weight is inversely proportional to the time distance. Through the time sequence processing and vector calculation of the data back information, the reverse motion vector reflecting the motion characteristics is successfully generated. Then, based on the reverse motion vector, the abnormal features in the motion process are identified. The direction mutation detection calculates the included angle of adjacent vectors, and marks the direction mutation point when the included angle exceeds the set threshold. The velocity anomaly includes sudden acceleration and sudden stop, and the velocity anomaly point is identified by calculating the acceleration. The identified direction mutation points and velocity anomaly points are spatially clustered to form abnormal regions. The number of abnormal points in unit area is calculated to generate the spatial distribution density. For example, the reverse motion vector analysis of "Xiaozhi" found that multiple direction mutation points occurred in a certain area, indicating that there is strong environmental constraint at this position. According to the density distribution and feature analysis of abnormal points, the environmental constraint strength of different regions is inferred, and the higher the density, the stronger the constraint. The constraint type inference combines abnormal features and environmental prior knowledge, such as narrow passages causing direction mutations and dynamic obstacles causing velocity anomalies. Through abnormal point identification and density analysis, the inference of environmental constraint strength is completed. Finally, based on the environmental constraint strength, the reverse features of historical trajectory containing multi-dimensional information are generated. The reverse feature vector F_reverse contains four dimensions of constraint strength, path preference, risk level and traffic efficiency, i.e. F_reverse=[f_constraint,f_preference,f_risk,f_efficiency], and each component is normalized to the range of [0,1]. The path preference is obtained by counting the number of historical passes in each region, reflecting the habitual path of the robot; the risk level is calculated according to the frequency and severity of abnormal events, which comprehensively evaluates the navigation risk of each region; the traffic efficiency is obtained by analyzing the historical speed data, calculating the ratio of actual traffic speed to theoretical maximum speed. The spatial representation of the feature uses a grid map, and each grid stores the corresponding four-dimensional feature vector. The time decay mechanism weights the historical data to ensure that recent data have a greater impact. Through multi-dimensional analysis and comprehensive evaluation, the generation of reverse features of historical trajectory is completed.

[0054] From the data back information, missing areas are identified and their passable properties are inferred. The data back information records the valid detection range and blind area distribution of all sensors, and missing areas do not always mean the presence of obstacles. Based on the shape features and failure patterns in the data back information, three inference principles are used to determine passability: the continuity principle assesses whether the blank area connects two known passable areas; the symmetry principle uses the regularity of the environment layout; and the functionality principle considers the functional attributes of space. For example, Xiao Zhi's data back information shows a rectangular hollow area at the entrance of a conference room. By analyzing the area connecting two passable corridors and conforming to the geometric characteristics of the door, it is inferred to be a potentially passable area. The inference result generates a passable probability, and boundary smoothing processing converts discrete judgments into continuous probability fields. The verification mechanism checks whether the historical trajectories in the data back information have ever passed through these areas, improving the credibility of the inference.

[0055] Using the original sensor measurement values in the data back information, numerical change rate analysis is performed to infer environmental boundaries. The data back information saves the original measurement values of vision and laser sensors before failure, and the spatial gradient of these numerical values is used to identify implicit boundaries. Boundary detection uses a multi-sensor fusion gradient operator, and when the gradient value exceeds an adaptive threshold, it is determined to be a boundary point. Based on the distribution characteristics of abnormal patterns in the data back information, physical boundaries (walls, obstacles), material boundaries (different ground materials), and light boundaries (light-dark boundaries) are distinguished. For example, the laser ranging mutation sequence recorded in Xiao Zhi's data back information reveals the precise profile position of the glass wall, which is information that is ignored in normal perception. The continuity of the boundary is formed by connecting adjacent boundary points to form a complete profile, and multi-scale analysis detects boundaries at different resolutions to capture complete information from the overall structure to local details.

[0056] The temporal reverse features, potential passable areas and environmental boundary information are integrated to construct a complete environmental implicit feature representation. The fusion framework adopts a multi-level structure, fully utilizing the complementarity of various features. The four dimensions of the temporal reverse features play different roles in the fusion: the constraint strength directly affects the obstacle judgment weight; the path preference adjusts the path attraction; the risk level is used for safety margin adjustment; and the passable efficiency affects the speed planning suggestion. The feature fusion function considers the contribution and mutual relationship of each feature to generate a unified implicit feature representation F_hidden, which contains the multi-dimensional implicit information of the environment. The conflict resolution mechanism handles the contradictions of different feature sources, and determines the final feature through confidence comparison and timeliness judgment. For example, when "Xiaozhi"'s reverse feature shows that a certain area is high-risk, while the potential passable analysis determines that it is feasible, the decision is arbitrated through environmental boundary information. Spatial interpolation fills in the feature blank area, maintaining the continuity of the feature field. Multi-resolution representation supports different precision requirements of applications, and semantic enhancement gives the implicit feature a clear meaning.

[0057] The reconstructed environmental implicit features and navigation gain parameters are combined to construct a reverse navigation field guiding the robot motion. The reverse navigation field is based on the potential field theory, but integrates data-driven implicit features and gain modulation. Each component of the navigation gain parameter participates in the construction of the field, including the path correction gain, safety margin and scene weight. The field function Φ(x, y) considers three components: implicit feature potential, gain modulation potential and safety potential field, where Φ(x, y) represents the potential energy value at position (x, y), w_scene is the scene weight, Φ_hidden is the implicit feature potential generated based on F_hidden, Φ_gain is the path gain modulation potential, and Φ_safety is the safety potential field. The implicit feature potential is generated according to the content of F_hidden, with passable areas as negative potential, invisible obstacles as positive potential, and historical preferred paths with additional attraction. The gain modulation potential compensates for path negative errors, and the safety potential field integrates static and dynamic margins to generate repulsion zones around obstacles. For example, in "Xiaozhi"'s office building navigation, the implicit feature potential at the door of the glass conference room is negative (passable), combined with the path correction gain and safety margin to form a comprehensive navigation field. The gradient of the field provides direction guidance for motion, ensuring the balance between target approach and obstacle avoidance. The scene weight w_scene is adjusted in real time according to the environment type, with narrow spaces increasing the influence of implicit features and open areas reducing constraints.

[0058] In step S140, a hollow density map is generated based on the spatial distribution density of the data hollow area in the reverse navigation field, the hollow cluster feature and the hollow connected mode are extracted from the hollow density map, the density threshold is set according to the hollow density map to divide the path phase change area, and the path phase change rule set is generated based on the path phase change area, the hollow cluster feature and the hollow connected mode.

[0059] In some embodiments, the analyzing the spatial distribution density of the data hole region based on the inverse navigation field to generate a hole density map comprises: locating a three-dimensional coordinate of the data hole region in the inverse navigation field; calculating a number of holes in a unit volume based on the three-dimensional coordinate to generate a local density value; spatially interpolating the local density value to generate a continuous density distribution; and mapping the continuous density distribution to a two-dimensional grid map to generate a hole density map.

[0060] First, the three-dimensional spatial distribution of the high-density region is accurately located based on the hole density map. The grid positions with density values exceeding a threshold are extracted from the hole density map, and the two-dimensional grid coordinates are converted into three-dimensional spatial coordinates. The coordinate transformation considers the current pose of the robot and the height information of the environment, and the three-dimensional coordinates of the hole points are represented as P_hole = [x, y, z]T. For example, the hole density map of "Xiaozhi" shows that the density at a certain corridor corner reaches 15 / m2, and the corresponding three-dimensional coordinates are concentrated near the ground (z ≈ 0.1 m) and the waist height (z ≈ 1.2 m). The accuracy of coordinate positioning is improved through multi-sensor cross-validation, and coordinate aggregation combines adjacent hole points to reduce data redundancy. Based on the efficient coordinate extraction of the hole density map, a set of three-dimensional coordinates of the data hole region is obtained.

[0061] Next, based on the located three-dimensional coordinates, the number of holes in a unit volume in space is calculated in combination with the density distribution data. The hole density map ρ_hole(x, y) is used as the basic data to extend the two-dimensional density to three-dimensional volume density, where ρ_hole represents the hole density value at position (x, y). The local density calculation uses the kernel density estimation method, and the voxelization method divides the space into 0.1m3 cubic units. The density value of each voxel is calculated by interpolating the ρ_hole value of the corresponding two-dimensional grid. Density normalization ensures comparability in different regions, and vertical layer analysis is based on the identified density distribution characteristics. For example, based on the high-density characteristics of the corner region in the "Xiaozhi" hole density map, the three-dimensional volume density of this region is calculated to be 12 / m3. The dynamic density update mechanism supports real-time environmental changes, and the newly detected holes are fused and updated with the existing density map.

[0062] Then, the calculated local density values are spatially interpolated to generate a continuous and smooth density distribution field. The interpolation process directly utilizes the continuity basis of the void density map and uses radial basis function (RBF) interpolation for three-dimensional expansion. The determination of the interpolation weight coefficient considers the density gradient information of adjacent grids in the density map, ensuring that the interpolation result is consistent with the original density distribution. The boundary processing uses an extrapolation strategy based on the gradient characteristics of the edge of the density map to estimate the density values outside the boundary. The interpolation quality is evaluated through cross-validation with the original density data to ensure that the expanded three-dimensional density field retains the original spatial features. Multi-scale interpolation is performed at different resolutions, with coarse scales inheriting overall trends and fine scales preserving local features. Continuity constraints ensure the first derivative of the density field to be continuous, avoiding unnatural jumps. Based on the RBF interpolation and smoothing processing driven by the density map, a spatially continuous density distribution field is generated.

[0063] Finally, the generated three-dimensional density distribution is projected onto a two-dimensional plane to update and optimize the void density map. The projection method uses maximum projection to retain the maximum density information in the vertical direction. The two-dimensional grid resolution is set to 0.2m to ensure compatibility of the data format. The optimized density map adds supplementary three-dimensional information, improving the accuracy of density estimation. Image enhancement processing improves the distinguishability of density differences, and edge detection identifies areas with sharp changes in density. For example, after three-dimensional processing, the density map accuracy of the "Xiaozhi" office building corner area is improved by 15%, more accurately reflecting the real distribution of the void. The visualization color scheme uses a heat map form, with red representing high density and blue representing low density.

[0064] Based on the generated hole density map, the clustering features and connectivity patterns of holes are extracted. The clustering analysis uses the DBSCAN algorithm to classify adjacent grids with ρ_hole exceeding the threshold value into the same class, and the density threshold is directly determined based on the distribution of ρ_hole. The clustering parameters are adjusted according to the density statistical characteristics: the average density is used as the reference, and the standard deviation is used to determine the clustering radius. The feature vector F_cluster of each cluster is [centroid, area, density_mean, shape_factor], where density_mean directly comes from the ρ_hole value of the corresponding region. The shape factor is described by calculating the aspect ratio and circularity of the cluster, and the long strip-shaped cluster (aspect ratio > 3) usually corresponds to a corridor or a channel. The connectivity analysis uses the graph theory method, taking the cluster as the node and the adjacent relationship as the edge to construct the connected graph G=(V,E), where V is the node set representing each cluster, and E is the edge set representing the connection relationship between clusters. For example, the hole density map of "Xiao Zhi" office building shows that the corridor area forms a chain-like connectivity pattern, and the open hall presents a network-like connectivity feature. The connectivity strength is measured by the shortest distance between clusters, and a distance less than 1m is considered as strong connectivity. The time evolution analysis tracks the dynamic changes of clusters and identifies stable patterns and transient patterns. According to the density map driven DBSCAN clustering and connectivity analysis, the hole clustering features and connectivity patterns are accurately extracted.

[0065] In some embodiments, according to the hole density map, a density threshold is set to divide the path phase change region, including: calculating the global average density and local peak density of the hole density map; setting a reference threshold based on the global average density and a dynamic threshold based on the local peak density; marking the region with a density exceeding the dynamic threshold as a strong phase change area, and the region with a density between the reference threshold and the dynamic threshold as a weak phase change area; merging the strong phase change area and the weak phase change area to generate a complete path phase change region.

[0066] First, the statistical characteristic indicators of the generated hole density map are calculated. The global average density p global is directly averaged over all p hole values in the density map, reflecting the overall hole distribution level. The local peak density is detected by applying a sliding window on the density map, with a window size of 2m x 2m, and recording the maximum p hole value within each window. The peak detection algorithm identifies local maximum points on the density map, requiring the peak point density to be greater than the density of the surrounding 8-neighborhood. For example, the density map statistics of the "Xiaozhi" office building show that the global average density p global = 4.2 / m2, and the local peak density of the corridor corner p peak local = 22 / m2. The density histogram is generated based on p hole data, showing a bimodal distribution characteristic. Spatial autocorrelation analysis evaluates the degree of aggregation of the density map, and temporal stability is evaluated by calculating the correlation coefficient of consecutive frame density maps. Outlier processing eliminates transient interference points in the calculation, and the 3σ criterion is used to identify outliers.

[0067] Next, according to the calculated density statistical characteristics, an adaptive density threshold system is set. The baseline threshold T base = p global + k _ σ x s global, where p global is the calculated global average density, s global is the corresponding standard deviation, and k _ σ is the standard deviation coefficient (typical value 1.5-2.0). The dynamic threshold T dynamic = a x p peak local + (1-a) x T base, where p peak local is the identified local peak density, and a is the weight coefficient (value 0.6-0.8). The spatial variation of the threshold is smoothed by Gaussian filtering, and the time adaptive mechanism adjusts the threshold according to historical statistical data. The influence of environmental type on the threshold is introduced by scene recognition, and the threshold tolerance is improved in complex environments. Threshold verification is performed by calculating the classification accuracy, with the actual distribution of the density map as the reference standard. Sensitivity analysis evaluates the influence of threshold changes on regional division, ensuring the robustness of classification.

[0068] Then, the density map is regionally divided based on the set T_base and T_dynamic thresholds. The strong phase transition region is defined as the region where p_hole > T_dynamic, indicating a high concentration of holes, and the navigation strategy needs to be changed. The weak phase transition region is defined as the region where T_base < p_hole < T_dynamic, indicating a moderate degree of perception degradation. Region labeling uses connected component analysis to ensure that adjacent grids of the same type are grouped into a region. Morphological operations optimize the region boundaries, with an opening operation to remove small noise regions and a closing operation to fill small holes within the region. Region feature extraction includes geometric properties such as area, perimeter, and compactness. Boundary smoothing uses the Douglas-Peucker algorithm to simplify the polygon boundary while preserving the main shape features. For example, based on the threshold division of the "Xiaozhi" office building density map, the corridor corner is labeled as a strong phase transition region, and the transition region is a weak phase transition region. Phase transition quantification reflects the degree of deviation from the normal state of the region, and the transition zone identifies the buffer region between strong and weak phase transition regions.

[0069] Finally, the identified strong and weak phase transition regions are integrated to generate a unified path phase transition region representation. Region merging uses the proximity principle, with regions of the same type within a distance of less than 0.3m being merged into a larger region to avoid excessive fragmentation affecting path planning efficiency. The merging process uses a recursive algorithm, first identifying a seed region, and then gradually absorbing surrounding regions of the same type that meet the conditions. Hierarchical representation decomposes large phase transition regions into multiple sub-regions, which are further divided based on the density distribution characteristics within the region to support fine-grained path planning. Sub-region division considers factors such as density gradient direction, region shape complexity, and boundary curvature to ensure that each sub-region has relatively consistent navigation characteristics. Region attribute labeling includes key information such as phase transition type, average density, area size, and shape descriptor to provide data support for subsequent rule generation. Boundary optimization uses a combination of morphological filtering and curve fitting to maintain the basic morphology of the region while eliminating the path planning difficulties caused by jagged boundaries. For example, in the "Xiaozhi" office building navigation, multiple small strong phase transition regions in the corridor region are merged into a continuous large region with a length of about 15 meters and a width of 2 meters, simplifying the complexity of path planning. Temporal and spatial consistency checks ensure the stability of the phase transition region in continuous navigation, avoiding path jitter caused by frequent region changes. Through density threshold division, region merging optimization, and hierarchical representation, accurate division and structured expression of the path phase transition region are completed.

[0070] The rules set guiding the navigation behavior is formulated by integrating the phase transition region, the cavity clustering feature and the connectivity mode. The rule generation adopts the condition-action format: IF (condition) THEN (action). The basic rules include: strong phase transition region avoidance rule (completely avoiding the high-density cavity region), weak phase transition region deceleration rule (reducing the speed and increasing the perception time), and chain connectivity following rule (adjusting the path in the direction of the cavity chain). The rule priority is determined according to the trade-off between safety and efficiency, and the safety-related rules have the highest priority. The conflict resolution mechanism handles the situation where multiple rules are triggered at the same time, and adopts the priority sorting and rule fusion strategy. The rule parameters include trigger threshold, action intensity, duration, etc., which support fine adjustment. For example, based on the phase transition region and clustering feature of the "XiaoZhi" office building, a rule set containing 15 specific rules is generated, covering the navigation strategies under different environmental conditions. The context-related rules consider the current task and environmental state, and the rule learning mechanism automatically extracts new rules or adjusts existing rules by analyzing successful and failed navigation cases.

[0071] In step S150, the initial navigation path is generated according to the path phase transition rule set and the reverse navigation field, and the initial navigation path is morphologically transformed to generate curve path parameters and three-dimensional path parameters, and the curve path parameters, three-dimensional path parameters and navigation gain parameters are fused to generate an adaptive navigation sequence.

[0072] Specifically, the initial navigation path of the robot is generated by multi-constraint fusion using the path phase transition rule set and the reverse navigation field. The path generation adopts an improved A algorithm combined with potential field theory, and the reverse navigation field Φ(x, y) provides the global potential energy distribution, and the path phase transition rule set provides local behavior constraints. The cost function F(n) of the initial path search is designed as F(n) = G(n) + H(n) + P(n), where F(n) is the total cost of node n, G(n) is the actual cost from the starting point to the current node, H(n) is the heuristic estimate from the current node to the target, and P(n) is the potential energy penalty term based on the reverse navigation field. The potential energy penalty term calculation formula is P(n) = w_field × Φ(x, y) + w_rule × R(n), where w_field is the navigation field weight, w_rule is the rule weight, and R(n) is the penalty value of the path phase transition rule. The specific implementation of the rule constraint is realized by dynamically adjusting the node expansion strategy in the path search process: when encountering a strong phase transition region, the search radius is increased to avoid the region, and when entering a weak phase transition region, the movement step is reduced to improve the accuracy. For example, "XiaoZhi" in the office building navigation, when the path approaches the strong phase transition region (ρ_hole = 22 / m²) of the corridor corner, the A algorithm automatically detours and selects an alternative path in the open area. Path smoothing adopts B-spline curve fitting to ensure the feasibility and continuity of the generated path. The safety margin check verifies that the distance from each point on the path to the nearest obstacle meets the safety requirements.

[0073] In some embodiments, the morphological transformation of the initial navigation path generates a curve path parameter and a three-dimensional path parameter, including: determining a section in the initial navigation path corresponding to the path phase change region based on the path phase change rule set; inserting a Bezier control point based on the start and end points of the section; fitting a smooth curve parameter as a curve path parameter based on the Bezier control point; extracting terrain undulation information based on the environmental implicit features of the initial navigation path passing region, and generating a height change sequence as a three-dimensional path parameter based on the terrain undulation information.

[0074] Based on the path phase change rule set, the section of the initial navigation path that needs special processing is accurately identified. Path section analysis is performed by checking whether each sampling point on the path is located in a strong phase change region or a weak phase change region, and the sampling interval is set to 0.1 m to ensure detection accuracy. The identification of the phase change section uses continuity judgment: when more than 3 consecutive sampling points are located in the same type of phase change region, it is marked as a complete section. The section features include start and end point coordinates, section length, phase change type and intensity, etc. For example, when "Xiaozhi" passes through the office building corridor corner, a strong phase change section with a length of 1.5 m is detected, with start point coordinates (10.2, 5.4) and end point coordinates (11.5, 6.1). Section classification divides the path into three categories: normal section, weak phase change section and strong phase change section, and different types of sections use different processing strategies. Transition section identification identifies the connection between the phase change region and the normal region to ensure the smoothness of the path transformation. Spatial weight allocation is determined according to the phase change intensity: strong phase change section weight 0.8, weak phase change section weight 0.6, normal section weight 0.4. Section importance evaluation considers phase change intensity, section length and passing difficulty to provide a basis for subsequent optimization. Based on the accurate analysis of the path phase change rule set, the accurate identification and classification of the phase change section in the initial navigation path are completed.

[0075] Based on the identified segment features, Bezier control points are inserted at key locations to achieve path smoothing. The control point insertion strategy is determined by the segment type: strong phase transition segments insert control points 0.3 m outside the start and end points, weak phase transition segments insert control points 0.2 m outside, and normal segments insert control points 0.1 m outside. The position of the control points is optimized by gradient descent perpendicular to the path direction, with the objective function considering path length, curvature continuity, and obstacle avoidance. The weight distribution of the Bezier control points uses the chord length parameterization method to ensure that the curve passes through the specified path points. For example, in the strong phase transition segment of the "Jiachi" corridor corner, four control points are inserted to form a cubic Bezier curve: the control point before the start point (9.9, 5.1), the start point (10.2, 5.4), the end point (11.5, 6.1), and the control point after the end point (11.8, 6.4). Control point constraints include a minimum turning radius limit R_min ≥ 0.8 m and a maximum curvature limit κ_max ≤ 1.2 m⁻¹ to prevent generating sharp turns that the robot cannot execute. The iterative optimization process minimizes the total path cost by adjusting the control point positions, with the convergence condition being that the cost change is less than 0.01 for five consecutive iterations. Control point validation checks whether it is within the feasible region to avoid collisions with obstacles. Multi-level optimization gradually adjusts the control points from coarse to fine to ensure global optimality.

[0076] Based on the optimized Bezier control points, curve parameters describing the path geometry are fitted. The parameter equation of the cubic Bezier curve is P(t) = (1-t)³P0 + 3(1-t)²tP1 + 3(1-t)t²P2 + t³P3, where t ∈ [0, 1] is the parameter and P0 to P3 are the control point coordinates. The curve path parameters include parameterized representation, curvature distribution κ(t), and tangent direction θ(t). The curvature calculation formula is κ(t) = |P'(t) × P''(t)| / |P'(t)|³, where P'(t) and P''(t) are the first and second derivatives, respectively. The tangent direction θ(t) = arctan(dy / dx) provides orientation information for each point on the path. The discretization of the path parameters uses adaptive sampling, with increased sampling density in areas with large curvature changes and reduced sampling points for straight segments. For example, the Bezier curve of the "Jiachi" corner segment has a maximum curvature κ_max = 0.8 m at t = 0.5, corresponding to a minimum turning radius of 1.25 m, which satisfies the robot kinematic constraints. Parameter validation is performed by checking curvature continuity and path feasibility to ensure that the generated curve parameters meet the physical constraints. The velocity profile matches the curvature distribution, with reduced speed v = v_max√(R_min / R(t)) in high-curvature areas and increased efficiency on straight segments. Arc length parameterization s(t) = ∫0ᵗ|P'(τ)|dτ provides a natural parameter representation of the path, facilitating subsequent motion planning. Through Bezier curve fitting and parameterization, a complete set of curve path parameters is generated. -1 , corresponding to a minimum turning radius of 1.25 m, which satisfies the robot kinematic constraints. Parameter validation is performed by checking curvature continuity and path feasibility to ensure that the generated curve parameters meet the physical constraints. The velocity profile matches the curvature distribution, with reduced speed v = v_max√(R_min / R(t)) in high-curvature areas and increased efficiency on straight segments. Arc length parameterization s(t) = ∫0ᵗ|P'(τ)|dτ provides a natural parameter representation of the path, facilitating subsequent motion planning. Through Bezier curve fitting and parameterization, a complete set of curve path parameters is generated.

[0077] Based on the initial navigation path through the area, the terrain undulation information is inferred from the aforementioned environment implicit features F_hidden, generating three-dimensional path parameters. The environment implicit features F_hidden contain four dimensions of temporal inverse features: constraint strength, path preference, risk level, and traffic efficiency, which are inherently related to the terrain undulation. Terrain inference uses multidimensional correlation analysis: areas with constraint strength f_constraint>0.7 usually correspond to height changes such as steps or slopes; path preference f_preference<0.3 indicates a historical avoidance tendency, often corresponding to uneven terrain; risk level f_risk>0.6 is related to terrain changes, with higher risk at locations such as step edges; and traffic efficiency f_efficiency<0.5 indicates that the actual speed is lower than the theoretical value, with uphill and downhill being the main reason. For example, when "Xiaozhi" path passes through the edge of the carpet, F_hidden shows a feature combination of high constraint, low preference, high risk, and low efficiency, inferring a height change of 0.02m. Terrain type recognition is based on feature combination patterns, and the undulation direction is determined by spatial gradients. The height change sequence H(s)=[h0,h1,...,h n ] is arranged according to the path arc length parameter, with a resolution of 0.05m. The vertical velocity dh / ds reflects the slope, and the pitch angle pitch(s)=arctan(dh / ds) describes the attitude adjustment. Parameter interpolation ensures continuity, and safety verification ensures within the physical limit range. Through F_hidden feature inference, complete three-dimensional path parameters are generated.

[0078] Based on the generated curve path parameters, three-dimensional path parameters and navigation gain parameters, multi-dimensional fusion is performed to generate an adaptive navigation sequence. A hierarchical structure is adopted: the geometric layer fuses the curve parameters and three-dimensional parameters to form a spatial path, the dynamics layer combines the navigation gain parameters to generate a motion sequence, and the control layer converts the motion sequence into execution instructions. The geometric fusion function Path_3D(s)=[x(s),y(s),h(s)]^T combines the two-dimensional curve path and the height variation sequence, where s is the arc length parameter. The four components of the navigation gain parameter (path correction gain, static obstacle avoidance margin, dynamic obstacle avoidance margin and scene weight) play a modulating role in fusion: the path correction gain compensates for systematic position errors, the static safety margin affects the distance between the path and the obstacle, the dynamic margin adjusts the safety distance according to the speed, and the scene weight balances different environmental strategies. The sequence update adopts a sliding window mechanism, and the window size is set to 2m. The parameters of the front path segment are updated in real time. For example, when "Xiaozhi" detects a new strong phase change region in front, the adaptive mechanism automatically adjusts the speed profile and safety margin of the region to ensure safe passage. The time parameterization of the navigation sequence is achieved through speed planning, considering the dynamic constraints of the robot and the task efficiency requirements. The speed profile V(s) is dynamically adjusted according to the path curvature κ(s), the terrain slope dh / ds and the safety margin in the aforementioned navigation gain parameters: V(s)=min(V_max,V_curve(κ),V_terrain(dh / ds),V_safety(margin)), where V(s) is the planned speed at arc length parameter s, V_max is the maximum allowed speed, V_curve(κ) is the speed limit based on curvature restriction (related to the aforementioned minimum turning radius R_min), V_terrain(dh / ds) is the speed limit based on terrain slope, and V_safety(margin) is the speed limit based on static and dynamic safety margins. Acceleration planning ensures the smoothness of speed changes, and the maximum acceleration and deceleration do not exceed the physical capabilities of the robot. The time stamp T(s)=∫[0→s]ds / V(s) provides the expected arrival time of each point on the path. Through the above multi-dimensional fusion and adaptive adjustment, the final generated navigation sequence can effectively cope with sensor defects and environmental uncertainties, and realizes stable, safe and efficient autonomous navigation control of humanoid robots in complex environments.

[0079] To perform the above-mentioned method embodiment corresponding to the humanoid robot navigation method based on defect complementation, to realize the corresponding functions and technical effects. Referring to Figure 2 , Figure 2 A structure block diagram of a humanoid robot navigation device 200 based on defect complementation provided by an embodiment of the present application is shown. For ease of illustration, only the parts related to the present embodiment are shown. The humanoid robot navigation device 200 based on defect complementation provided by the present embodiment includes:

[0080] The defect analysis module 201 is used to acquire navigation data of the humanoid robot, extract and identify defect features based on the navigation data, measure defect points and data void areas, and perform spatial distribution analysis based on the measured defect points and data void areas to generate a defect feature map.

[0081] The gain generation module 202 is used to establish confidence-decreasing verification to obtain multi-level perception results using the defect feature map, determine sensor weight allocation based on the multi-level perception results, extract negative error values ​​of each sensor from the defect feature map based on the sensor weight allocation to generate a negative error feature set, and generate navigation gain parameters using the negative error feature set.

[0082] The reverse field construction module 203 is used to perform reverse parsing on the navigation data through the navigation gain parameters to extract back information of the data, reconstruct the implicit features of the environment based on the back information of the data, and construct a reverse navigation field by combining the implicit features of the environment and the navigation gain parameters.

[0083] The phase transition rule module 204 is used to generate a hole density map based on the spatial distribution density of the data hole region analyzed by the reverse navigation field, extract hole clustering features and hole connectivity patterns from the hole density map, divide the path phase transition region according to the density threshold set by the hole density map, and generate a path phase transition rule set based on the path phase transition region, the hole clustering features and the hole connectivity patterns.

[0084] The path generation module 205 is used to generate an initial navigation path based on the path phase transition rule set and the reverse navigation field, perform morphological transformation on the initial navigation path to generate curved path parameters and three-dimensional path parameters, and fuse the curved path parameters, the three-dimensional path parameters and the navigation gain parameters to generate an adaptive navigation sequence.

[0085] The aforementioned humanoid robot navigation device 200 based on defect complementarity can implement the humanoid robot navigation method based on defect complementarity described in the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0086] like Figure 3 As shown, the third embodiment of the present invention also provides a computer device, including a memory 301, a processor 302, and a computer program stored in the memory 301 and executable on the processor 302, characterized in that the processor 302 executes the program to implement the steps of the humanoid robot navigation method based on defect complementarity described in the first embodiment of the present invention.

[0087] The above examples are intended to illustrate and deduce the technical solutions of the present application, and to completely describe the technical solutions, objects and effects of the present application. The purpose is to make the public more thoroughly and comprehensively understand the disclosed content of the present application, and does not limit the protection scope of the present application.

[0088] The above examples are not based on an exhaustive enumeration of the present application, and there can be many other unlisted embodiments. Any substitution and improvement made without violating the concept of the present application shall fall within the protection scope of the present application.

Claims

1. A method for navigating a humanoid robot based on complementary defects, characterized in that, The method comprises the following steps: acquiring humanoid robot navigation data, extracting defect features according to the navigation data, measuring defect points and data hollow areas, analyzing the spatial distribution based on the measured defect points and the data hollow areas, and generating a defect feature map; using the defect feature map to establish a confidence-decreasing verification to obtain multi-level perception results, determining sensor weight distribution according to the multi-level perception results, extracting negative error values of each sensor from the defect feature map based on the sensor weight distribution to generate a negative error feature set, and using the negative error feature set to generate navigation gain parameters; extracting data back information by reverse analysis of the navigation data based on the navigation gain parameters, reconstructing environmental implicit features based on the data back information, and constructing a reverse navigation field by combining the environmental implicit features and the navigation gain parameters; generating a hollow density map based on the spatial distribution density of the data hollow areas analyzed by the reverse navigation field, extracting hollow clustering features and hollow connectivity patterns from the hollow density map, setting a density threshold to divide path phase change areas according to the hollow density map, and generating a path phase change rule set based on the path phase change areas, the hollow clustering features, and the hollow connectivity patterns; generating an initial navigation path according to the path phase change rule set and the reverse navigation field, generating curve path parameters and three-dimensional path parameters by morphological transformation of the initial navigation path, and generating an adaptive navigation sequence by fusing the curve path parameters, the three-dimensional path parameters, and the navigation gain parameters.

2. The method of claim 1, wherein, The method of determining sensor weight distribution according to the multi-level perception results comprises the following steps: extracting sensor response intensity and stability indicators from the multi-level perception results; calculating the contribution degree of sensors at different confidence levels based on the response intensity; generating dynamic weight coefficients based on the stability indicators and the contribution degree, and generating sensor weight distribution by normalizing the dynamic weight coefficients.

3. The method of claim 1, wherein, The method of generating navigation gain parameters using the negative error feature set comprises the following steps: decomposing the negative error feature set into positioning negative bias components and ranging negative error components; generating path correction gain based on the positioning negative bias components; generating obstacle avoidance safety margin based on the ranging negative error components; generating navigation gain parameters by fusing the path correction gain and the obstacle avoidance safety margin.

4. The method of claim 1, wherein, The method of reconstructing environmental implicit features based on the data back information comprises the following steps: obtaining inverse features of historical trajectories by time reversal of the data back information; identifying data missing areas from the data back information and mapping them as potential passable areas; analyzing the data back information to infer the position of the environmental boundary; generating complete environmental implicit features by fusing the inverse features, the potential passable areas, and the environmental boundary position.

5. The method of claim 1, wherein, The method of generating a hollow density map based on the spatial distribution density of the data hollow areas analyzed by the reverse navigation field comprises the following steps: locating the three-dimensional coordinates of the data hollow areas in the reverse navigation field; calculating the number of hollows in a unit volume based on the three-dimensional coordinates to generate local density values; generating continuous density distribution by spatial interpolation of the local density values; The continuous density distribution is mapped to a two-dimensional grid map to generate a cavity density map.

6. The method of claim 1, wherein, A density threshold is set according to the cavity density map to divide a path phase change region, including: A global average density and a local peak density of the cavity density map are calculated; A reference threshold is set based on the global average density, and a dynamic threshold is set based on the local peak density; Regions with a density exceeding the dynamic threshold are marked as strong phase change regions, and regions with a density between the reference threshold and the dynamic threshold are marked as weak phase change regions; The strong phase change regions and the weak phase change regions are combined to generate a complete path phase change region.

7. The method of claim 1, wherein, The initial navigation path is morphologically transformed to generate a curve path parameter and a three-dimensional path parameter, including: A section corresponding to the path phase change region in the initial navigation path is determined based on the path phase change rule set; A Bezier control point is inserted based on the start and end points of the section; A smooth curve parameter is fitted and generated as a curve path parameter based on the Bezier control point; Terrain fluctuation information is extracted based on the environmental implicit features of the region through which the initial navigation path passes, and a height change sequence is generated as a three-dimensional path parameter based on the terrain fluctuation information.

8. The method of claim 4, wherein, The reverse features of the historical trajectory are obtained by time reversal of the data back information, including: A reverse time sequence is generated in reverse order based on the data back information according to the time stamp, and a reverse motion vector is extracted based on the position difference between adjacent time points; Direction mutation points and speed anomaly points are identified based on the reverse motion vector, and the environmental constraint strength is inferred based on the distribution density of the direction mutation points and the speed anomaly points in space; The reverse features of the historical trajectory are generated based on the environmental constraint strength.

9. A humanoid robot navigation device based on complementary defects, characterized in that, It includes: A defect analysis module for obtaining humanoid robot navigation data, extracting and identifying defect features from the navigation data, measuring defect points and data cavity regions, and generating a defect feature map based on spatial distribution analysis of the measured defect points and the data cavity regions; A gain generation module for establishing a confidence decreasing verification using the defect feature map to obtain a multi-level perception result, determining a sensor weight distribution based on the multi-level perception result, extracting a negative error value of each sensor from the defect feature map based on the sensor weight distribution to generate a negative error feature set, and generating a navigation gain parameter using the negative error feature set; A reverse field construction module for reverse analysis of the navigation data to extract data back information based on the navigation gain parameter, reconstructing environmental implicit features based on the data back information, and constructing a reverse navigation field combining the environmental implicit features and the navigation gain parameter; A phase change rule module for analyzing the spatial distribution density of the data cavity region based on the reverse navigation field to generate a cavity density map, extracting a cavity clustering feature and a cavity connectivity mode from the cavity density map, setting a density threshold according to the cavity density map to divide a path phase change region, and generating a path phase change rule set based on the path phase change region, the cavity clustering feature and the cavity connectivity mode. A path generation module is configured to generate an initial navigation path according to the path phase change rule set and the reverse navigation field, perform morphological transformation on the initial navigation path to generate a curved path parameter and a three-dimensional path parameter, and fuse the curved path parameter, the three-dimensional path parameter, and the navigation gain parameter to generate an adaptive navigation sequence.

10. A computer device, comprising: A computer program product includes a memory and a processor; the memory is configured to store a computer program; the processor is configured to execute the computer program and implement the method according to any one of claims 1 to 8 when the computer program is executed.

Images