Humanoid robot terrain processing and multi-terrain gait control method

Abstract

The present application relates to the technical field of robot adaptive control, and provides a humanoid robot terrain processing and multi-terrain gait control method, which comprises the following steps: planning an inspection route, cutting a walking segment according to spatial position and functional area; collecting multi-source data by using a multi-source environment sensing unit, performing fusion processing on each walking segment, and extracting terrain features; according to a preset discrimination rule, mapping the terrain features of each walking segment into a terrain label with a road surface type identifier and a safety risk level, and converting the safety risk level into a gait level; according to the gait level, automatically selecting a target gait mode, and issuing gait control parameters matched with the gait mode to each joint controller; monitoring the foot bottom contact state and posture change in real time, comparing the terrain label and the gait level of the corresponding walking segment, dynamically adjusting the gait level of the corresponding walking segment and the subsequent gait level, and updating the gait control parameters.

Description

Technical Field

[0001] This invention relates to the field of robot adaptive control technology, specifically a method for terrain processing and multi-terrain gait control of a humanoid robot. Background Technology

[0002] Robot gait control refers to a complex system that manages and plans the movement patterns of legged robots, enabling them to walk and run smoothly. It involves coordinating the movement of multiple joints and actuators to adapt to different terrains, maintain balance, and perform various tasks.

[0003] Chinese patent publication number CN119078986A, published on December 6, 2024, entitled "A Gait Control Method for a Humanoid Robot," describes a method that includes: collecting the lift height of the forefoot; controlling the rotation of the joints of the hind legs to position the hind feet directly below the center of gravity of the humanoid robot; collecting the pressure distribution of the forefoot after it contacts an object; determining whether the forefoot is stable; if unstable, collecting the rotation angle of the forefoot; if the rotation angle is less than a threshold, switching between the forefoot and hind legs; otherwise, lifting the forefoot, adjusting the angles of the joints of the forefoot to change the object it contacts, and re-collecting the pressure distribution of the forefoot. By internally rotating the hind legs and bending the knees in the opposite direction, the stability of walking is improved. Furthermore, the forefoot can sense the pressure distribution on the contact surface. When the environmental information detected by the sensing unit is not very accurate, the pressure sensing of the forefoot enhances the adaptability of the gait to the environment, further improving the stability of the humanoid robot's walking.

[0004] In nighttime inspection scenarios at transportation hubs, uneven lighting, temporary obstacles, and slippery areas on the ground make it difficult to identify them in time. The combination of these complex factors can easily cause the robot to slip. In addition, the inspection route passes through different types of ground with significant changes in friction and height, requiring different gait patterns. There is a lack of specialized methods for handling typical station terrain.

[0005] To address the aforementioned issues, this invention proposes a method for terrain processing and multi-terrain gait control of a humanoid robot. This method pre-plans inspection routes, divides the routes into walking segments, collects multi-source data using multi-source sensors, extracts corresponding terrain features, generates terrain labels and risk levels, and then automatically selects gait modes from a pattern library and distributes them to each joint. Real-time corrections are made based on feedback from the operational status. Summary of the Invention

[0006] In view of the existing problems mentioned above, a method for terrain processing and multi-terrain gait control of humanoid robots is proposed.

[0007] The technical solution adopted by this invention to solve the above-mentioned technical problems is: a method for terrain processing and multi-terrain gait control of a humanoid robot, comprising: The inspection route is planned according to the site plan and inspection tasks, and the inspection route is discretized into multiple walking segments according to spatial location and functional area. When the humanoid robot moves along the inspection route, it uses a multi-source environmental perception unit to collect multi-source data, performs fusion processing on each walking segment, and extracts three types of terrain features: material, slope, and position. According to the preset discrimination rules, the terrain features of each walking segment are mapped to terrain labels with road surface type and safety risk level, and the safety risk level is converted into gait level, forming a mapping relationship between walking segments, terrain labels and gait level; Based on the gait level of the current walking segment, the target gait mode is automatically selected from the gait mode library, and gait control parameters matching the gait mode are sent to each joint controller to adaptively switch the gait between different walking segments. The system monitors the plantar contact status and posture changes in real time, compares the monitoring results with the terrain labels and gait levels of the corresponding walking segments, and dynamically adjusts the gait levels of the corresponding walking segments and their subsequent gait levels and updates the gait control parameters when a walking safety abnormality is detected.

[0008] As a preferred embodiment, the specific steps for discretizing the inspection route into multiple walking segments according to spatial location and functional area are as follows: Read the site plan, floor plan, and equipment layout plan, and convert the boundary information of each functional area into a site map with spatial coordinates and functional attributes; based on the necessary points, priorities, and time constraints given in the inspection task, use a multi-objective optimization algorithm to generate an initial inspection route from the starting point to the end point. The inspection route is represented as a polyline composed of continuous path points. It is jointly segmented according to the spatial distance between path points and the boundaries of functional areas. When the cumulative length of the path exceeds the preset basic segment distance, a new walking segment is automatically generated, and the start and end coordinates, the functional area to which it belongs, and the identification information required for terrain processing are recorded for each walking segment.

[0009] As a preferred embodiment, the specific steps for acquiring multi-source data using a multi-source environmental sensing unit are as follows: When the humanoid robot moves along the sequence of walking segments, the multi-source environmental perception unit continuously collects visual images, depth information, foot contact force and posture information in the spatial range corresponding to each walking segment, and converts the data collected at different times into the same ground coordinate system to form the local terrain point cloud and foot contact record of the corresponding segment. The multi-source environmental perception unit consists of a head vision sensor, a depth sensor, a plantar pressure sensor array, an inertial measurement unit, and a lidar sensor. The head vision sensor is used to acquire color image data of the ground in front to obtain visual information; the depth sensor is used to acquire point cloud data of the ground and obstacles to obtain geometric information; the plantar pressure sensor array is used to acquire the contact force magnitude, pressure distribution, and its time-varying curve for each step to obtain contact state information; the inertial measurement unit is used to acquire the robot's attitude angle, angular velocity, and linear acceleration data to obtain attitude disturbance information; and the lidar sensor is used to acquire three-dimensional distance information of the area in front to enhance the ability to recognize the contours of platform edges, stairwells, and obstacles.

[0010] As a preferred embodiment, the specific steps for extracting the three types of terrain features—material, slope, and location—are as follows: First, visual and geometric information are jointly analyzed. Road surface areas are classified using visual feature indicators. The classification results are then corrected by incorporating changes in foot contact force and high-frequency vibration components, yielding material characteristics representing the road surface material category. Finally, a plane fitting is performed on the local terrain point cloud. The plane fitting formula is as follows: , Where x, y, and z represent the point cloud coordinates in a unified ground coordinate system. b and c represent the fitting coefficients of the first, second, and third planes, respectively; The slope and aspect of the current ground segment are estimated based on the plane, forming the slope characteristics. The slope angle formula is: , By combining the locations of key structures marked on the station map, the height difference and horizontal distance between the fitted plane and the key locations are calculated to obtain the positional features describing the relative positional relationships of the segments. The formula is as follows: , Where k represents the index number of the key position. This represents the height difference feature of the ground relative to the key location k in a walking segment. These represent the three-dimensional coordinates of the key location k in the ground coordinate system.

[0011] As a preferred embodiment, the specific steps for mapping the terrain features of each walking segment to terrain labels with road surface type identification and safety risk level are as follows: The system reads three types of terrain features for each walking segment: material, slope, and location. The material features are compared with a pre-defined road surface material rule library to determine the road surface type for that segment, and a corresponding material risk coefficient is assigned to each type. The slope angle and height difference are then substituted into the risk assessment model to calculate the comprehensive safety risk score for that walking segment. The formula is as follows: , Where i represents the index of the walking segment, Indicates material weight. Indicates the slope weight. Indicates positional weight. This represents the material risk index for the i-th walking segment. This represents the slope risk index for the i-th walking segment. This represents the location risk index for the i-th walking segment; Based on the preset risk classification threshold, the comprehensive safety risk score is discretized into three levels: low risk, medium risk, and high risk. These levels are then combined with road surface type markings to form terrain labels, such as flat stone ground – low risk. According to the mapping relationship between safety risk levels and gait control strategies, low risk, medium risk, and high risk are respectively converted into normal gait level, cautious passage gait level, and key protection gait level. A unique terrain label and gait level are assigned to each walking segment, realizing the mapping between walking segments and terrain labels and gait levels.

[0012] As a preferred embodiment, the specific steps for automatically selecting a target gait pattern from the gait pattern library based on the gait level of the walking segment to which the current position belongs are as follows: Within each control cycle, the walking segment number is calculated based on the current pose, and the gait level corresponding to the walking segment is read from the established mapping. The current pose of the humanoid robot is transformed into the field map coordinate system consistent with the inspection route. The discrete path point sequence of the inspection route and the cumulative path length interval corresponding to each walking segment are read. Within each control cycle, the current pose is projected onto the broken line of the inspection route, the distance between it and all path points is calculated, the index of the path point with the smallest distance is selected, and the cumulative path length corresponding to that point is taken as the current walking mileage along the route. The walking mileage interval of all walking segments is searched for the segment that matches the walking mileage, and its number is taken as the current walking segment number. In a pre-built gait pattern library, basic gait patterns are retrieved by gait level to obtain basic gait control parameter templates corresponding to the normal gait level, the attention-based passage gait level, and the key protection gait level. Combining the comprehensive safety risk score and slope information of the walking segment, the basic parameter templates are fine-tuned to generate the target gait control parameter vector for the current segment. The formula is: , in This represents the target gait control parameter vector for the i-th walking segment. This represents the gait level of the i-th walking segment. This represents a basic parameter template. This represents the adjustment gain matrix. This represents the adjustment factor vector calculated from the risk score and slope. This represents the overall safety risk score for the i-th walking segment. This represents the slope angle of the i-th walking segment.

[0013] As a preferred embodiment, the specific steps for adaptively switching gait between different walking segments by issuing gait control parameters matching the gait pattern to each joint controller are as follows: The target gait control parameter vector is sent to each joint controller, and coordinated adjustments are made between stride length, foot lift height, center of gravity swing and double foot support time. When the robot crosses the boundary of adjacent walking segments, the target gait control parameters corresponding to the previous segment and the next segment are interpolated to achieve smooth gait adaptive switching between different walking segments.

[0014] As a preferred embodiment, the specific steps for real-time monitoring of foot contact status and posture changes, and comparing the monitoring results with the terrain tags and gait levels of the corresponding walking segments, are as follows: In each control cycle, the outputs of the plantar pressure sensor and the inertial measurement unit are read in real time to calculate the plantar contact integrity index, slip index, and posture deviation index of the current walking segment. The indexes are compared with the preset safety range of the terrain label and gait level corresponding to the walking segment. When some indexes are found to exceed the corresponding safety limit, the walking safety anomaly index of the segment is calculated, and the walking safety anomaly is determined accordingly. Once an anomaly is determined, the gait level of the current walking segment and subsequent segments is upgraded to a higher protection level, and the target gait control parameters of the corresponding segment are recalculated and generated.

[0015] Compared with the prior art, the present invention has the following advantages: 1. By fusing visual, depth, plantar pressure and posture information through a multi-source environmental perception unit, three types of terrain features—material, slope and position—are extracted at the granularity of walking segments. Furthermore, terrain labels and gait levels are generated, enabling the robot to automatically switch between different gait modes in different terrains and achieve adaptive walking.

[0016] 2. Different walking segments are graded and protected by comprehensive safety risk scoring. During operation, the integrity of foot contact, slippage index and posture deviation are monitored in real time. If walking safety abnormalities are detected, the gait level of the corresponding segment and subsequent segments is automatically upgraded, thereby reducing the probability of accidents such as slipping and falling without changing the overall inspection route. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0018] Figure 1 This is a flowchart illustrating the present invention; Figure 2 This is a comparison diagram of the effects of the present invention and the prior art, where gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0019] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0020] Example 1: To achieve the above objectives, please refer to Figures 1 to 2 This invention provides a method for terrain processing and multi-terrain gait control of a humanoid robot, the method comprising the following steps: Step S1: Plan the inspection route and divide the walking segment; Step S2: Multi-source data perception and extraction of terrain features; Step S3: Generate terrain labels and risk levels; Step S4: Automatically select gait mode and launch; Step S5: Real-time feedback of operating status for corrective action.

[0021] This method is implemented in the order of S1–S5, and its overall process is as follows: The inspection route is planned according to the site plan and inspection tasks, and the inspection route is discretized into multiple walking segments according to spatial location and functional area. When the humanoid robot moves along the inspection route, it uses a multi-source environmental perception unit to collect multi-source data, performs fusion processing on each walking segment, and extracts three types of terrain features: material, slope, and position. According to the preset discrimination rules, the terrain features of each walking segment are mapped to terrain labels with road surface type and safety risk level, and the safety risk level is converted into gait level, forming a mapping relationship between walking segments, terrain labels and gait level; Based on the gait level of the current walking segment, the target gait mode is automatically selected from the gait mode library, and gait control parameters matching the gait mode are sent to each joint controller to adaptively switch the gait between different walking segments. The system monitors the plantar contact status and posture changes in real time, compares the monitoring results with the terrain labels and gait levels of the corresponding walking segments, and dynamically adjusts the gait levels of the corresponding walking segments and their subsequent gait levels and updates the gait control parameters when a walking safety abnormality is detected.

[0022] The specific steps for planning the inspection route and dividing the walking segment are as follows: Read the site plan, floor plan, and equipment layout diagram, and convert the boundary information of functional areas such as guard points, station hall, platform, stairs, passages, and equipment rooms into a site map with spatial coordinates and functional attributes; based on the necessary points, priorities, and time constraints given in the inspection task, use a multi-objective optimization algorithm to generate an initial inspection route from the starting point to the end point. Specifically, the multi-objective optimization algorithm considers multiple optimization objectives and constraints such as path length, time consumption, and safety margin at the same time, and performs a weighted comprehensive solution for each objective to obtain an inspection route scheme that is approximately optimal in an overall sense. The inspection route is represented as a polyline composed of continuous path points. It is jointly segmented according to the spatial distance between path points and the boundaries of functional areas. When the cumulative length of the path exceeds the preset basic segment distance, such as 3m, a new walking segment is automatically generated. The start and end coordinates, the functional area to which it belongs, and the identification information required for terrain processing are recorded for each walking segment.

[0023] Through the above steps, this invention does not simply use a number of discrete navigation points, but forms a sequence of walking segments with functional semantics, providing a structured basis for subsequently attaching shape tags and gait levels to the segments, thus achieving a clear distinction from existing technical solutions that only plan paths in geometric space.

[0024] The specific steps for multi-source data perception and terrain feature extraction are as follows: When the humanoid robot moves along the sequence of walking segments, the multi-source environmental perception unit continuously collects visual images, depth information, foot contact force and posture information in the spatial range corresponding to each walking segment, and converts the data collected at different times into the same ground coordinate system to form the local terrain point cloud and foot contact record of the corresponding segment. Specifically, the multi-source environmental perception unit comprises a head vision sensor, a depth sensor, a plantar pressure sensor array, an inertial measurement unit, and a lidar sensor. The head vision sensor acquires color image data of the ground in front, obtaining visual information; the depth sensor acquires point cloud data of the ground and obstacles, obtaining geometric information; the plantar pressure sensor array acquires the magnitude of the contact force, pressure distribution, and its time-varying curve for each step, obtaining contact state information; the inertial measurement unit acquires the robot's attitude angle, angular velocity, and linear acceleration data, obtaining attitude disturbance information; and the lidar sensor acquires three-dimensional distance information of the area in front, enhancing the ability to recognize the contours of platform edges, stairwells, and obstacles. First, visual and geometric information are jointly analyzed. Road surface areas are classified using indicators such as texture features, color histograms, and reflectivity. For example, when the energy and contrast of the texture features are low, the color histogram is concentrated in the neutral color channel, and the proportion of bright pixels in the reflectivity exceeds the first threshold, the area is classified as polished surface. When the contrast and roughness of the texture features are high, the color histogram is relatively uniformly distributed, and the proportion of bright pixels is below the second threshold, the area is classified as non-slip surface. When the color histogram shows a significant peak in the yellow channel, the texture is in a regular stripe pattern, and there are periodic micro-protrusions in the depth data, the area is classified as a tactile paving area. When the reflectivity exhibits localized high brightness and strong reflection, with an approximately circular shape and clear edge texture, the area is classified as a metal cover area. The classification results are then corrected by combining the variation amplitude of foot contact force and high-frequency vibration components to obtain material characteristics representing the category of road surface material in the segment. Finally, a plane fitting is performed on the local terrain point cloud. The plane fitting formula is: , Where x, y, and z represent the point cloud coordinates in a unified ground coordinate system. b and c represent the fitting coefficients of the first, second, and third planes, respectively; The slope and aspect of the current ground segment are estimated based on the plane, forming the slope characteristics. The slope angle formula is: , By combining key locations such as platform edges, stairwells, and equipment foundations marked on the station map, the height difference and horizontal distance between the fitted plane and these key locations are calculated to obtain the positional features describing the relative positional relationships of the segments. The formula is as follows: , Where k represents the index number of the key position. This represents the height difference feature of the ground relative to the key location k in a walking segment. These represent the three-dimensional coordinates of the key location k in the ground coordinate system.

[0025] In this way, the present invention does not make instantaneous judgments on a single frame image, but extracts three types of terrain features by multi-source fusion on the basis of "walking segments". This enables subsequent terrain labeling and gait control to achieve fine adaptation at the segment granularity, which is different from the existing simple shape recognition schemes that are based on a single sensor or single point height information.

[0026] The specific steps for generating terrain labels and risk levels are as follows: The system reads three types of terrain features for each walking segment: material, slope, and location. The material features are compared with a pre-defined road surface material rule library to determine the road surface type for that segment, and a corresponding material risk coefficient is assigned to each type. The slope angle and height difference are then substituted into the risk assessment model to calculate the comprehensive safety risk score for that walking segment. The formula is as follows: , Where i represents the index of the walking segment, Indicates material weight. Indicates the slope weight. Indicates positional weight. This represents the material risk index for the i-th walking segment. This represents the slope risk index for the i-th walking segment. This represents the location risk index for the i-th walking segment; Based on the preset risk classification threshold, the comprehensive safety risk score is discretized into three levels: low risk, medium risk, and high risk. These levels are then combined with road surface type markings to form terrain labels, such as flat stone ground – low risk. According to the mapping relationship between safety risk levels and gait control strategies, low risk, medium risk, and high risk are respectively converted into normal gait level, cautious passage gait level, and key protection gait level. A unique terrain label and gait level are assigned to each walking segment, realizing the mapping between walking segments and terrain labels and gait levels.

[0027] Through the above steps, this invention explicitly introduces two intermediate semantic layers at the fragment granularity level: terrain labels and gait levels. This transforms continuous, difficult-to-use raw terrain features into discrete-level information that can directly drive gait control, unlike existing technologies that simply trigger gait switching based on height thresholds for terrain.

[0028] The specific steps for automatically selecting gait mode are as follows: Within each control cycle, the walking segment number is calculated based on the current pose, and the gait level corresponding to the walking segment is read from the established mapping. Specifically, the current pose of the humanoid robot is transformed into the field map coordinate system consistent with the inspection route, and the discrete path point sequence of the inspection route and the cumulative path length interval corresponding to each walking segment are read. Within each control cycle, the current pose is projected onto the broken line of the inspection route, the distance between it and all path points is calculated, the path point index with the smallest distance is selected, and the cumulative path length corresponding to that point is taken as the current walking mileage along the route. The walking mileage interval of all walking segments is searched for the segment that matches the walking mileage, and its number is taken as the current walking segment number. In a pre-built gait pattern library, basic gait patterns are retrieved by gait level to obtain basic gait control parameter templates corresponding to the normal gait level, the attention-based gait level, and the key protection gait level. These templates include reference stride length, reference foot lift height, reference swing time, and the proportion of support. Combining the comprehensive safety risk score and slope information of the walking segment, the basic parameter templates are fine-tuned to generate the target gait control parameter vector for the current segment. The formula is as follows: , in This represents the target gait control parameter vector for the i-th walking segment. This represents the gait level of the i-th walking segment. This represents a basic parameter template. This represents the adjustment gain matrix. This represents the adjustment factor vector calculated from the risk score and slope. This represents the overall safety risk score for the i-th walking segment. Let represent the slope angle of the i-th walking segment; The target gait control parameter vector is sent to each joint controller to achieve coordinated adjustment of stride length, foot lift height, center of gravity swing and support time of both feet. When the robot crosses the boundary of adjacent walking segments, the target gait control parameters corresponding to the previous segment and the next segment are interpolated to achieve smooth gait adaptive switching between different walking segments.

[0029] Through the above steps, this invention automatically selects gait modes by combining gait level with segment features as an index, and continuously adjusts parameters within a unified control framework. This differs from existing technologies that only coarsely switch between a few fixed gaits based on a single terrain category and lack smooth transitions at segment boundaries.

[0030] The specific steps for real-time feedback of operational status and error correction are as follows: In each control cycle, the outputs of the plantar pressure sensor and the inertial measurement unit are read in real time to calculate the plantar contact integrity index, slip index, and posture deviation index of the current walking segment. These indices are compared with the preset safety range of the terrain label and gait level corresponding to the walking segment. When some indices are found to exceed the corresponding safety limit, the walking safety anomaly index of the segment is calculated, and a walking safety anomaly is determined accordingly. Once an anomaly is determined, the gait level of the current walking segment and subsequent segments is upgraded to a higher protection level. At the same time, the target gait control parameters of the corresponding segment are recalculated and generated, thereby reducing the local walking risk without changing the overall inspection route.

[0031] Through the above steps, this invention does not simply rely on offline-set gait levels, but instead performs a closed-loop comparison between real-time foot contact state and posture changes and terrain labels and gait levels. It dynamically adjusts gait levels and control parameters at the granularity of walking segments, realizing online adaptive correction of risks that are not modeled in the field. This is different from the control methods in the prior art where gait levels are fixed once set and lack feedback correction for environmental perception errors.

[0032] like Figure 2 This diagram compares the effects of a humanoid robot terrain processing and multi-terrain gait control method. The horizontal axis lists key performance indicators, and the vertical axis represents exemplified performance scores, ranging from 0 to 100, with higher values ​​indicating better performance. The aim is to visually demonstrate the expected improvement of this invention in key capabilities compared to typical existing technologies. Specifically, multi-terrain adaptability characterizes the humanoid robot's ability to maintain normal walking on different road surfaces such as stone, non-slip surfaces, tactile paving, and metal covers; non-slip safety characterizes the robot's ability to suppress slips and instability when walking on high-risk terrains such as wet, slippery, and sloping surfaces; gait transition smoothness characterizes the continuity of parameter transitions when the robot completes gait level transitions at the boundaries of adjacent walking segments; and inspection task completion rate characterizes the proportion of walking segments inspected according to predetermined task requirements along a complete inspection route.

[0033] Example 2: Based on the above embodiment 1, a method for terrain processing and multi-terrain gait control of a humanoid robot in a nighttime inspection scenario at a transportation hub is specifically as follows: Step 1: During the initial deployment phase at the site, guide the humanoid robot to walk slowly along the expected inspection route, collect pose and obstacle information in real time, automatically generate a topological grid map containing semantic nodes, connect the recorded pose trajectories in chronological order to form the original inspection route, and adjust the semantic nodes on the trajectory according to the inspection task requirements to obtain the target inspection route that does not depend on the site plan and equipment layout diagram; discretize the target inspection route into several path points, and when the path curvature exceeds a threshold or the walking distance from the previous semantic node exceeds a set value, divide it into new walking segments at the corresponding positions, and bind the start and end path point numbers, adjacent semantic node identifiers and functional area labels to each walking segment for subsequent terrain processing and gait level configuration; Step 2: Select several key moments within each walking segment, uniformly register the color images from the head vision sensor and the depth maps from the depth perception sensor into a multi-channel terrain grid, and input it into the pre-trained terrain recognition model. The model directly outputs the road surface material type, average slope, and relative position to key structures such as platform edges and stairwells for that segment. Then, using the segment as a time window, statistically analyze the data from the foot pressure sensor array and inertial measurement unit, calculate the contact stability index, and fuse it with the material results output by the terrain recognition model to correct the recognition error of high-risk materials such as slippery and loose materials. This provides three types of terrain features for each walking segment: material features, slope features, and position features. Step 3: Read the three types of terrain features (material, slope, and location) for each walking segment, and input them into a feature vector. This vector is then used to input a pre-trained offline risk assessment model. The risk assessment model outputs a continuous risk score and a road surface type probability distribution for each walking segment. The risk assessment model uses a gradient boosting tree, which is learned from labeled samples that correlate features with abnormal postures in a large amount of historical walking data. This allows the weights of material, slope, and height difference on risk to be automatically determined by the data. The continuous risk score output by the risk assessment model is divided into three levels—low risk, medium risk, and high risk—according to a preset threshold. This score is then combined with the road surface type label with the highest probability to generate a terrain label. Based on a preset risk level and gait control strategy mapping table, low, medium, and high risks are respectively assigned to normal gait level, cautious passage gait level, and key protection gait level. A unique terrain label and gait level are automatically assigned to each walking segment. Step 4: Based on the inspection route and the terrain labels and gait levels of each walking segment, combined with simulation and trial operation data, an offline gait pre-scheduling table is generated. Reference step length, foot lift height, support ratio, and transition parameters with adjacent segments are preset for each walking segment. The humanoid robot accumulates the walking distance within the current segment using an odometer. When the accumulated distance reaches the segment length threshold, it switches to the next walking segment number and directly reads the corresponding gait level and reference gait control parameters from the gait pre-scheduling table. Based on the stability indicators of the most recent steps within the segment, the reference parameters are adjusted proportionally to obtain the target gait control parameters for the current segment. At the segment boundary, the target parameters of the preceding and following segments are interpolated according to the pre-stored transition parameters to achieve smooth gait transitions. Step 5: Set a fixed number of steps within each walking segment and read the outputs of the plantar pressure sensor and inertial measurement unit in real time. Encode each step's contact interruption, excessive horizontal slip, and out-of-tolerance attitude angle or angular velocity as anomaly markers. Count the frequency of each type of anomaly marker in the current segment as the window slides to obtain a segment running stability score. Based on the combination of the stability score and the initial comprehensive risk score of the segment, look up the corresponding target gait level and parameter contraction coefficient from the pre-trained correction strategy library. When the stability score indicates that the current segment has entered an unstable state, only the gait level of the segment and the immediately following segments is increased, and the step length is reduced proportionally to the contraction coefficient, the support phase ratio is increased, and the body swing amplitude is limited. At the same time, after the task is completed, the stability score is written into the experience library to update the default gait level of the segment in subsequent tasks.

[0034] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A method for terrain processing and multi-terrain gait control of a humanoid robot, characterized in that, include: The inspection route is planned according to the site plan and inspection tasks, and the inspection route is discretized into multiple walking segments according to spatial location and functional area. When the humanoid robot moves along the inspection route, it uses a multi-source environmental perception unit to collect multi-source data, performs fusion processing on each walking segment, and extracts three types of terrain features: material, slope, and position. According to the preset discrimination rules, the terrain features of each walking segment are mapped to terrain labels with road surface type and safety risk level, and the safety risk level is converted into gait level, forming a mapping relationship between walking segments, terrain labels and gait level; Based on the gait level of the current walking segment, the target gait mode is automatically selected from the gait mode library, and gait control parameters matching the gait mode are sent to each joint controller to adaptively switch the gait between different walking segments. In a pre-built gait pattern library, basic gait patterns are retrieved by gait level to obtain basic gait control parameter templates corresponding to the normal gait level, the attention-based passage gait level, and the key protection gait level. Combining the comprehensive safety risk score and slope information of the walking segment, the basic parameter templates are fine-tuned to generate the target gait control parameter vector for the current segment. The formula is: ， in This represents the target gait control parameter vector for the i-th walking segment. This represents the gait level of the i-th walking segment. This represents a basic parameter template. This represents the adjustment gain matrix. This represents the adjustment factor vector calculated from the risk score and slope. This represents the overall safety risk score for the i-th walking segment. Let represent the slope angle of the i-th walking segment; The system monitors the plantar contact state and posture changes in real time, compares the monitoring results with the terrain label and gait level of the corresponding walking segment, and dynamically adjusts the corresponding walking segment and its subsequent gait level when a walking safety abnormality is detected, and recalculates and generates the target gait control parameters for the corresponding segment.

2. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for discretizing the inspection route into multiple walking segments according to spatial location and functional area are as follows: Read the site plan, floor plan, and equipment layout plan, and convert the boundary information of each functional area into a site map with spatial coordinates and functional attributes; based on the necessary points, priorities, and time constraints given in the inspection task, use a multi-objective optimization algorithm to generate an initial inspection route from the starting point to the end point. The inspection route is represented as a polyline composed of continuous path points. It is jointly segmented according to the spatial distance between path points and the boundaries of functional areas. When the cumulative length of the path exceeds the preset basic segment distance, a new walking segment is automatically generated, and the start and end coordinates, the functional area to which it belongs, and the identification information required for terrain processing are recorded for each walking segment.

3. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for acquiring multi-source data using a multi-source environmental sensing unit are as follows: When the humanoid robot moves along the sequence of walking segments, the multi-source environmental perception unit continuously collects visual images, depth information, foot contact force and posture information in the spatial range corresponding to each walking segment, and converts the data collected at different times into the same ground coordinate system to form the local terrain point cloud and foot contact record of the corresponding segment. The multi-source environmental perception unit consists of a head vision sensor, a depth sensor, a plantar pressure sensor array, an inertial measurement unit, and a lidar sensor. The head vision sensor is used to acquire color image data of the ground in front to obtain visual information; the depth sensor is used to acquire point cloud data of the ground and obstacles to obtain geometric information; the plantar pressure sensor array is used to acquire the contact force magnitude, pressure distribution, and its time-varying curve for each step to obtain contact state information; the inertial measurement unit is used to acquire the robot's attitude angle, angular velocity, and linear acceleration data to obtain attitude disturbance information; and the lidar sensor is used to acquire three-dimensional distance information of the area in front to enhance the ability to recognize the contours of platform edges, stairwells, and obstacles.

4. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for extracting the three types of terrain features—material, slope, and location—are as follows: First, visual and geometric information are jointly analyzed. The road surface area is classified by visual feature indicators. The classification results are corrected by combining the change amplitude of foot contact force and high-frequency vibration components to obtain the material characteristics that characterize the material category of the road surface segment. Plane fitting is performed on the local terrain point cloud. The plane fitting formula is as follows: ， Where x, y, z represent the point cloud coordinates in a unified ground coordinate system, and a, b, c represent the fitting coefficients of the first, second, and third planes, respectively; The slope and aspect of the current ground segment are estimated based on the plane, forming the slope characteristics. The slope angle formula is: ， By combining the locations of key structures marked on the station map, the height difference and horizontal distance between the fitted plane and the key locations are calculated to obtain the positional features describing the relative positional relationships of the segments. The formula is as follows: ， Where k represents the index number of the key position. This represents the height difference feature of the ground relative to the key location k in a walking segment. These represent the three-dimensional coordinates of the key location k in the ground coordinate system.

5. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for mapping the terrain features of each walking segment into terrain labels with road surface type identification and safety risk level are as follows: The system reads three types of terrain features for each walking segment: material, slope, and location. The material features are compared with a pre-defined road surface material rule library to determine the road surface type for that segment, and a corresponding material risk coefficient is assigned to each type. The slope angle and height difference are then substituted into the risk assessment model to calculate the comprehensive safety risk score for that walking segment. The formula is as follows: ， Where i represents the index of the walking segment, Indicates material weight. Indicates the slope weight. Indicates positional weight. This represents the material risk index for the i-th walking segment. This represents the slope risk index for the i-th walking segment. This represents the location risk index for the i-th walking segment.

6. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 5, characterized in that, The specific steps for mapping the terrain features of each walking segment to terrain labels with road surface type identification and safety risk level also include: Based on the preset risk classification threshold, the comprehensive safety risk score is discretized into three categories: low risk, medium risk, and high risk. These categories are then combined with road surface type markings to form terrain labels. According to the mapping relationship between safety risk levels and gait control strategies, low risk, medium risk, and high risk are respectively converted into normal gait level, cautious passage gait level, and key protection gait level. A unique terrain label and gait level are assigned to each walking segment, realizing the mapping between walking segments and terrain labels and gait levels.

7. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for automatically selecting a target gait pattern from the gait pattern library based on the gait level of the walking segment to which the current position belongs are as follows: Within each control cycle, the robot calculates the walking segment number based on its current pose and reads the gait level corresponding to the walking segment from the established mapping. The robot's current pose is then converted to the field map coordinate system consistent with the inspection route. The discrete path point sequence of the inspection route and the cumulative path length interval corresponding to each walking segment are read. Within each control cycle, the current pose is projected onto the broken line of the inspection route, and the distance between it and all path points is calculated. The path point index with the smallest distance is selected, and the cumulative path length corresponding to that point is taken as the current walking mileage along the route. The robot searches for segments with matching walking mileage intervals among all walking segments and uses their numbers as the current walking segment number.

8. The method for terrain processing and multi-terrain gait control of a humanoid robot according to claim 1, characterized in that, The specific steps for adaptively switching gait between different walking segments are as follows: (This involves sending gait control parameters matching the gait mode to each joint controller.) The target gait control parameter vector is sent to each joint controller, and coordinated adjustments are made between stride length, foot lift height, center of gravity swing and double foot support time. When the robot crosses the boundary of adjacent walking segments, the target gait control parameters corresponding to the previous segment and the next segment are interpolated to achieve smooth gait adaptive switching between different walking segments.

Images