AI humanoid robot control method suitable for complex terrain

Abstract

This invention relates to the field of robotics and discloses a control method for AI humanoid robots applicable to complex terrain. The method includes: dividing the single-leg motion cycle of the humanoid robot into three phases: a pre-swing phase, a post-swing phase, and a support phase; in the pre-swing phase, generating a prediction instruction package based on plantar pressure distribution data and a local terrain depth map, the prediction instruction package including the expected ankle joint ground contact angle, the expected ground stiffness level, and the expected plantar slippage level; in the post-swing phase, loading the parameters in the prediction instruction package into the ankle joint controller based on the height of the swinging leg from the ground; in the support phase, comparing the actual contact data with the prediction instruction package, identifying the source of deviation, and updating the prediction model; the three phases of the left and right feet are executed asynchronously and in parallel, coordinated by phase state mutual exclusion rules and cross-foot trigger signals; this invention solves the time mismatch problem between control commands and physical feedback in robot motion control under complex terrain, achieving real-time adaptation to terrain changes.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more specifically, to a control method for AI humanoid robots suitable for complex terrain. Background Technology

[0002] Humanoid robots face significant challenges in motion control when navigating complex terrains such as gravel roads, soft sand, slippery surfaces, and uneven terrain. Existing humanoid robot control methods typically employ a linear "plan-execute-feedback" control process. This involves first planning a motion trajectory based on a pre-defined model, then executing that trajectory using joint controllers, and finally compensating and correcting through sensor feedback. This control paradigm performs well in known or structured terrain environments. However, when the robot encounters unmodeled, instantaneous terrain changes, this linear process suffers from a fundamental timing defect: by the time the robot detects a terrain change, the currently executing gait has already locked into an incorrect motion pattern, requiring compensation only in the next control cycle, resulting in a control response that lags behind the terrain change.

[0003] Chinese patent CN120422241A discloses a motion control method and system for intelligent robots. This method uses multi-sensor fusion and neural networks for motion prediction, employs energy consumption-stability multi-objective optimization to generate a reference trajectory, and achieves motion control through collaboration between a central controller and local controllers. While this approach focuses on global optimization and improved computational architecture, its control logic remains based on a sequential "planning-execution-feedback" structure, making it difficult to fundamentally solve the terrain adaptation problem at the moment of contact with the ground.

[0004] Chinese patent CN108858208A discloses an adaptive balance control method for humanoid robots in complex terrain. This method establishes a linkage model, calculates foot placement points, adjusts joint angles, uses zero-torque points to determine the balance state, and optimizes the foot movement trajectory using a PID balance control algorithm. While this scheme addresses terrain changes through foot placement adjustment and PID compensation, its adjustment mechanism is a "post-compensation" mode, meaning adjustments are only made after the foot touches the ground based on deviations, resulting in a response delay.

[0005] In summary, existing technologies have not effectively solved the time mismatch problem between control commands and physical feedback in robot motion control under complex terrain. Therefore, this paper proposes an AI humanoid robot control method suitable for complex terrain to address the above problems. Summary of the Invention

[0006] In order to overcome the above-mentioned defects of the prior art, the embodiments of the present invention provide an AI humanoid robot control method suitable for complex terrain. The technical problem to be solved is to eliminate the time mismatch between control commands and physical feedback when the humanoid robot walks in complex terrain.

[0007] To achieve the above objectives, the present invention provides an AI humanoid robot control method suitable for complex terrain, comprising the following steps: The gait cycle of a humanoid robot is divided into three phases: the early swing phase, the late swing phase, and the support phase. By decomposing the gait cycle into three consecutive time intervals, a time window is provided for terrain prediction and joint parameter preloading in advance during the swing phase.

[0008] In the early stage of the swing, a predictive command package is generated based on the plantar pressure distribution data at the moment the current foot leaves the ground and the local terrain depth map of the next landing point collected by the visual sensor. The predictive command package includes the expected ankle contact angle, the expected ground stiffness level, and the expected plantar slip level. By rapidly analyzing the contact state at the moment of departure and the terrain of the next landing point, predictive parameters for the next ground contact posture are formed.

[0009] In the later stage of the swing, the expected ankle joint contact angle, the expected ground stiffness level, and the expected foot slip level are loaded into the ankle joint controller in layers according to the height of the swinging leg from the ground, so that the ankle joint parameters are preset before the foot touches the ground, eliminating the time delay of adjustment after contact.

[0010] During the support phase, the plantar pressure distribution data at the moment of foot contact and the actual ankle contact angle are compared with the predicted instruction package to identify the source of deviation. The predicted model is then updated based on the identification results. By comparing the actual contact data with the predicted data, the predicted model can be corrected online.

[0011] The three phases of the left and right feet are executed asynchronously in parallel. The left and right feet are coordinated through phase state mutual exclusion rules and cross-foot trigger signals to ensure that the left and right feet alternately complete the swing and support actions, so as to achieve continuous connection of the gait cycle.

[0012] Furthermore, the early swing phase is the time interval from the moment the foot leaves the ground to the moment the foot rises to its highest point, used to complete terrain prediction during the lifting process after the foot leaves the ground; the late swing phase is the time interval from the moment the foot begins to descend from its highest point to the moment the foot touches the ground, used to complete joint parameter preloading during the descent; and the support phase is the time interval from the moment the foot touches the ground to the moment the foot leaves the ground again, used to verify the accuracy of the prediction and update the model during the support phase.

[0013] Furthermore, the generation of the prediction instruction package in step S2 specifically includes: extracting the slope value of the landing point area from the local terrain depth map. Surface texture density value at landing point and the magnitude of the change in landing point depth The terrain of the landing point is quantitatively represented by extracting three terrain features; Assign a value to the expected ankle contact angle. ,Will and Mapped to the expected ground stiffness level ,Will and Mapped to the expected plantar slip level The terrain features are transformed into predictive parameters required for joint control through mapping rules.

[0014] Furthermore, the expected ankle joint ground contact angle The ankle angle value is the value at which the foot strikes the ground next, used to pre-adjust the ankle posture to adapt to the terrain slope; the expected ground stiffness level The expected foot slip level is a graded value for the ground hardness at the next footing point, used to preset the ankle joint stiffness coefficient to adapt to the ground hardness. The graded value for the probability of foot slippage at the next landing point is used to preset the ankle joint damping coefficient and adjust the knee joint flexion angle to deal with the risk of slippage.

[0015] Furthermore, let the height of the swing leg from the ground be... The first threshold is The second threshold is ,and The hierarchical loading in step S3 specifically includes: when At that time, rotate the ankle joint to Set the ankle joint stiffness coefficient to The corresponding stiffness value is used to adjust the angle and preset the stiffness when the foot is far from the ground. when At that time, the ankle joint damping coefficient was set to The corresponding damping value will increase the knee flexion angle. The corresponding flexion increment completes the damping preset and flexion adjustment when the foot approaches the ground, so that all joint parameters are preset before contact with the ground.

[0016] Furthermore, let the actual ground contact angle of the ankle joint be... The expected ankle contact angle is Let the stiffness level corresponding to the actual plantar impact characteristics be . The expected ground stiffness level is Let the actual foot slip distance correspond to the slip level as . The expected plantar slippage level is Step S4, identifying the sources of deviation and updating the prediction model based on the identification results, specifically includes: calculating... , , The difference between the prediction and the actual situation is quantified through deviation calculation; let the angle deviation threshold be... The stiffness deviation threshold is The slip deviation threshold is This is used to distinguish between acceptable errors and errors that need to be corrected. when At that time, the slope judgment threshold in the prediction model is set to The accuracy of subsequent slope identification is improved by updating the threshold; when When the hardness judgment feature template in the prediction model is replaced with the template corresponding to the current plantar impact feature, the accuracy of subsequent hardness classification is improved by replacing the template. when When the slip risk judgment feature template in the prediction model is replaced with the template corresponding to the current plantar slip feature, the accuracy of subsequent slip risk assessment is improved by replacing the template.

[0017] Furthermore, the asynchronous parallel execution of the three phases of the left and right feet in step S5 specifically includes: when the first foot is in the early or late swing phase, the second foot is in the support phase, and the robot always has a support foot to maintain balance through phase mutual exclusion. After the first foot completes the loading of the later stage of the swing, a ready signal is sent to the second foot. After receiving the ready signal, the second foot will advance the timing of entering the early stage of the swing by one step. The signal will start the prediction process of the other foot in advance, so as to achieve seamless connection between the left and right feet. When the first foot is detected during the support phase or or When the second foot receives the prediction correction signal, it sends a prediction correction signal to the second foot. After receiving the prediction correction signal, the second foot increases the conservative coefficient of the prediction parameter in the early stage of the next swing by a preset increment. By sharing the prediction error information through the signal, the other foot adopts a more conservative prediction strategy to avoid repeating the error.

[0018] Furthermore, the phase state mutual exclusion rule is as follows: when one foot is in the early or late swing phase, the other foot is in the support phase; when one foot is in the support phase, the other foot is in the early or late swing phase. This rule ensures that the swing leg and the support leg are mutually exclusive, maintaining the dynamic balance of the robot during walking.

[0019] Furthermore, it also includes step S6: recording the template replacement operation performed during each support period, assuming that the number of template replacements corresponding to a certain type of error is... The number of times threshold is ,when When an error occurs, a feature template is added to the feature template library corresponding to that type of error. By accumulating frequently occurring error features, the capacity of the feature template library is expanded, enabling the prediction model to adapt to more types of terrain conditions.

[0020] Furthermore, the parameters loaded into the ankle joint controller in step S3 are executed the instant the foot touches the ground, so that the preset joint parameters take effect immediately when the foot contacts the ground, without waiting for feedback signals after contact, thus eliminating the response delay of adjustment after contact in traditional methods.

[0021] The technical effects and advantages of this invention are as follows: This invention divides the single-leg gait cycle of a humanoid robot into three phases: the pre-swing phase, the post-swing phase, and the stance phase. This creates a time window for asynchronous pre-control. In the pre-swing phase, the system generates a prediction instruction package based on plantar pressure distribution data and a local terrain depth map. This instruction package contains only three lightweight parameters: the expected ankle contact angle, the expected ground stiffness level, and the expected plantar slip level. It is generated by quickly mapping and extracting three features: terrain slope, surface texture, and depth abrupt changes. This design avoids the complex global terrain reconstruction and complete gait planning calculations required in traditional methods, facilitating the completion of prediction operations within the limited time window of the pre-swing phase.

[0022] During the later stages of the swing, the system loads parameters from the prediction command package to the ankle joint controller in layers based on the height of the swinging leg from the ground. When the swinging leg is higher than the ground, the first layer of loading is executed, rotating the ankle joint to the expected angle and presetting the stiffness coefficient. When the swinging leg is lower than the ground, the second layer of loading is executed, presetting the damping coefficient and increasing the knee flexion angle. This layered loading mechanism allows joint parameters to be preset before the foot touches the ground, and can be executed the instant the foot touches the ground, helping to eliminate the delay in adjustment after ground contact in traditional methods.

[0023] During the support phase, the system collects actual ground contact data and compares it with the predicted command package to calculate angle deviation, stiffness deviation, and slip deviation. The source of error is identified based on the magnitude of the deviation: when the deviation exceeds a threshold, the slope judgment threshold is updated or the feature template is replaced; when the deviation is within the threshold range, no update operation is performed. This deviation attribution mechanism can distinguish between "model error" and "execution error," updating only the model error and avoiding misjudging random disturbances as environmental changes and incorrectly updating the model, thus achieving accurate online learning.

[0024] The three phases of the left and right feet are executed asynchronously in parallel. A phase state mutual exclusion rule ensures that only one foot is in the swing phase and the other in the support phase at any given time. Coordination between the left and right feet is achieved through ready signals and prediction correction signals: when one foot completes preloading, it sends a signal to the other foot, causing it to enter the early swing phase ahead of time; when one foot detects a prediction error, it sends a signal to the other foot, increasing the conservatism coefficient of the prediction parameters. This cross-foot coordination mechanism enables seamless connection between the left and right feet and instant sharing of prediction experience. Through this method, the present invention completes joint parameter adjustment before the foot touches the ground, which helps to improve the response delay problem of robot motion control in complex terrain. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the overall execution process of the method of the present invention; Figure 2 This is a flowchart of the swing pre-judgment instruction packet generation process of the present invention; Figure 3 This is a flowchart of the phased loading process in the later stage of the swing motion of the present invention; Figure 4 This is a flowchart of the support period deviation identification and model update process of the present invention; Figure 5 This is a flowchart illustrating the asynchronous coordination process between the left and right feet of this invention. Detailed Implementation

[0026] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] Example 1 As attached Figures 1 to 5 As shown, this invention provides an AI humanoid robot control method suitable for complex terrain. This method divides the single-step cycle into three phases. In the early stage of the swing, the terrain of the next foot landing point is predicted. In the later stage of the swing, the predicted parameters are loaded into the ankle joint controller in layers. In the support phase, the accuracy of the prediction is verified by actual contact data and the prediction model is updated. At the same time, the left and right feet are executed asynchronously in parallel and coordinate with each other, thereby eliminating the response delay of adjustment after ground contact in traditional control methods.

[0028] 1. Gait phase segmentation The single-step phase cycle of the humanoid robot is divided into three phases: the early swing phase, the late swing phase, and the support phase.

[0029] The initial swing phase is the time interval from the moment the foot leaves the ground to the highest point of its upward movement. The moment the foot leaves the ground is determined by plantar pressure sensors: when the pressure values ​​of all sensing units are below a set threshold, the foot is considered to have left the ground. During this phase, the swinging leg begins to swing forward from the ground position and gradually rises until it reaches the highest point of its swing trajectory. The duration of this phase depends on the walking speed. The dynamic changes are calculated in real time by the gait planner based on the current speed; the faster the speed, the shorter the initial swing duration.

[0030] The late swing phase is the time interval from when the foot begins to descend from its highest point until it touches the ground. During this phase, the swinging leg begins to move downward from its highest point, gradually approaching the ground until the sole of the foot is about to touch the ground. The duration of this phase is similar to that of the early swing phase and is uniformly determined by the gait planner.

[0031] The stance phase is the time interval from the moment the foot touches the ground until it leaves the ground again. During this phase, the foot is in contact with the ground, supporting the robot's body weight until the foot leaves the ground again to enter the next gait cycle. The length of the stance phase depends on the walking speed and gait pattern. The end time of the stance phase is determined by multiplying the gait cycle duration by the proportion of the stance phase, which decreases as walking speed increases.

[0032] The three phases mentioned above constitute a complete gait cycle. Each cycle corresponds to the complete process of one foot leaving the ground and leaving the ground again. The left and right feet alternately perform the above phases to form continuous walking.

[0033] 2. Generation of prediction instruction packets in the early stage of swing. In the early stage of the swing, perform the following operations to generate a prediction instruction package.

[0034] The first step involves collecting plantar pressure distribution data at the moment the foot leaves the ground using plantar pressure sensors. This data records the pressure distribution of the foot at the moment of liftoff, including the location of the pressure center and the pressure value in each area.

[0035] The second step involves acquiring a local terrain depth map of the next landing point using a visual sensor. This depth map covers the area surrounding the next landing point and includes depth data of the landing point area. The position of the landing point area in the depth map is determined based on the current gait parameters (stride length, walking direction) and the trajectory of the swing leg. The predicted landing point position is used as the center of the depth map crop.

[0036] The third step is to extract three terrain features from the local terrain depth map: the slope value of the landing area. Surface texture density value at landing point and the magnitude of the change in landing point depth The landing area is defined as the projection area of ​​the foot, and its size is determined according to the actual size of the robot's foot.

[0037] Slope value The extraction process is as follows: The 3D coordinates of pixels within the landing point region of the depth map are fitted to a plane to obtain the normal vector of the fitted plane. The angle between this normal vector and the vertical direction is calculated; this angle is the slope value. The unit is degrees. A value of 0 degrees represents a level surface, and a larger value indicates a steeper slope.

[0038] Texture density value The extraction process is as follows: Divide the landing point area into several sub-regions, calculate the depth variance in each sub-region, and take the average of the variances of all sub-regions as the depth variance. value, The value is a dimensionless variance, where 0 indicates a perfectly flat surface and a larger value indicates a rougher surface.

[0039] Deep mutation magnitude value The extraction process is as follows: detect the depth difference between adjacent pixels in the depth map, and take the maximum absolute value of the depth difference between all adjacent pixels as the maximum value. Value, in meters. The larger the value, the more uneven the ground.

[0040] The fourth step is to map the above terrain features into three prediction parameters.

[0041] Will Directly assign the value to the expected ankle contact angle. ,Right now Expected ankle contact angle This indicates the angle that the ankle joint should maintain when the foot touches the ground next.

[0042] Will and Mapped to expected ground stiffness level The mapping rule is: when Greater than the first texture threshold and When it is less than the first mutation threshold, Set the value to 1 to represent a hard surface; when Less than the second texture threshold or When it exceeds the second mutation threshold, The value is 3, indicating a soft surface. Other cases A value of 2 indicates a medium-hardness surface.

[0043] The mapping rule was obtained through experimental calibration and reflects the correlation between ground hardness and surface texture and depth abrupt changes.

[0044] Will and Mapped to expected plantar slip level The mapping rule is: when Less than the slope threshold and When it is less than the mutation threshold, A value of 1 indicates low slip risk; when Greater than the slope threshold or When it exceeds the mutation threshold, A value of 3 indicates a high risk of slippage; the rest... A value of 2 indicates a moderate slip risk.

[0045] Fifth step, , , The prediction instruction is packaged into a pre-judgment instruction packet and stored in the buffer for later reading during the oscillation phase. At this point, the early oscillation phase ends, and the robot enters the later oscillation phase.

[0046] 3. Layered loading in the later stage of the oscillation During the later stages of the swing, a layered loading process is performed. This phase begins when the swinging leg reaches its highest point and starts to descend, and ends just before the foot touches the ground.

[0047] First, calculate the height of the swing leg from the ground. This height is calculated in real time using data from knee and ankle encoders combined with a kinematic model. Let the thigh length be... The length of the lower leg is Hip joint height It is calculated using the angles of the knee and ankle joints of the supporting leg and the lengths of the thigh and calf, using the following formula: ; in To support the knee joint angle, The supporting leg's ankle joint angle is given. This formula converts the joint angle of the supporting leg into the hip joint's height off the ground, a fundamental calculation in robot kinematics. Let the swing leg's knee joint angle be... The angle of the swinging leg ankle joint is Then the normal kinematics formula is: .

[0048] Let the first threshold be The second threshold is ,and First threshold The second threshold corresponds to the height at which the foot is about to enter the ground's influence zone. The critical height at which the ball of the foot approaches the ground.

[0049] When the swing leg is at the height of the ground At that time, the first-level loading is executed. Specific operations include: Ankle joint rotation is achieved by using an ankle joint motor to drive the ankle joint to rotate, thus increasing the ankle joint angle. The rotation speed is controlled to a set value to ensure that the rotation is completed before the foot touches the ground. The joint controller outputs control torque based on the difference between the current angle and the target angle. The larger the difference, the larger the output torque.

[0050] Set the ankle stiffness coefficient to The corresponding stiffness value, when When, the stiffness coefficient is set to a high stiffness value; when When, the stiffness coefficient is set to a medium stiffness value; when At this time, the stiffness coefficient is set to a low stiffness value. The stiffness value is achieved by adjusting the current loop parameters of the joint motor.

[0051] when At that time, the second-level loading is executed. Specific operations include: Set the ankle damping coefficient to The corresponding damping value. When When, the damping coefficient is set to a low damping value; when When, the damping coefficient is set to the medium damping value; when At this time, the damping coefficient is set to a high damping value. The damping value is achieved by adjusting the speed loop parameters of the joint motor.

[0052] Increase the knee flexion angle The corresponding buckling increment. When When the buckling increment is 0 degrees; when When, the buckling increment is the first increment value; when At this time, the flexion increment is the second increment value. The flexion angle is achieved by adjusting the target position of the knee joint motor, that is, adding the flexion increment value based on the current target position of the knee joint.

[0053] The parameters for the first and second levels of loading are executed the instant the foot contacts the ground. The moment of contact is detected by the sudden change in pressure value from the plantar pressure sensor; when the pressure value rises from zero to a set threshold, it is determined to be contact. The contact signal triggers the ankle joint controller to immediately operate with preset stiffness coefficients, damping coefficients, and rotation angles, while simultaneously switching the robot state to the support phase.

[0054] 4. Deviation identification and model update during the support period During the support phase, deviation identification and model updates are performed. This phase begins the moment the foot touches the ground and continues until the foot leaves the ground again.

[0055] The first step is to collect actual contact data. An ankle encoder is used to read the actual ground contact angle of the ankle joint at the moment the foot strikes the ground. The peak impact force is collected at the moment of ground contact using a plantar pressure sensor, and the actual stiffness level is obtained by mapping the magnitude of the peak impact force. The mapping rule is: When the peak impact force is greater than the first impact threshold, Take 1; When the peak impact force is between the second impact threshold and the first impact threshold, Take 2; When the peak impact force is less than the second impact threshold, Take 3.

[0056] The peak impact force is negatively correlated with ground hardness; the harder the ground, the greater the peak impact force. The pressure center's movement distance is detected by plantar pressure sensors during the support phase. The pressure center's coordinates are calculated by weighted average of the pressure values ​​from each sensor unit, using the following formula: , ; in For the first The pressure value of each sensing unit Its coordinates are used. The total distance the pressure center moves during the support period is taken as the slip distance, and the actual slip level is obtained by mapping the slip distance. The mapping rule is: when the slip distance is greater than the first slip threshold, Take 3; when the slip distance is between the second slip threshold and the first slip threshold, Take 2; when the sliding distance is less than the second sliding threshold, Take 1. The greater the sliding distance, the more severe the foot slippage.

[0057] The second step is to calculate the deviation value. Angle deviation. The unit is degrees. Stiffness deviation. Dimensionless. Slip deviation. Dimensionless. Let the angular deviation threshold be... The stiffness deviation threshold is The slip deviation threshold is .

[0058] The third step is to update the prediction model based on the deviation value.

[0059] when When the slope identification threshold is set to 0, it indicates an error in slope recognition. The slope judgment threshold in the prediction model is then set to 0. This means updating the slope judgment threshold with the current deviation value, so that the same terrain features can obtain a more accurate slope value in subsequent predictions.

[0060] when If the ground hardness classification is incorrect, it indicates an error in the ground hardness classification. The soft / hardness judgment feature template in the prediction model is replaced with the template corresponding to the current foot impact feature. The impact features at the current ground contact moment (including peak impact force, impact rise time, and impact waveform) are recorded as the new judgment template. Several impact feature vectors are stored in the feature template library; the new template replaces the existing template with the lowest similarity.

[0061] when If the error occurs, it indicates that there is an error in the slip risk assessment. The slip risk judgment feature template in the prediction model is replaced with the template corresponding to the current plantar slip characteristics, and the slip characteristics (including slip distance, slip speed, and pressure center trajectory) of the current support period are recorded as the new judgment template.

[0062] when , and At that time, no update operation is performed.

[0063] The updated prediction model (slope judgment threshold, feature template) is used when generating the prediction instruction package in the early stage of the next swing.

[0064] 5. Asynchronous parallel execution and coordination of the left and right feet The three phases of the left and right feet are executed asynchronously and in parallel. When the first foot is in the early or late swing phase, the second foot is in the support phase; when the first foot is in the support phase, the second foot is in the early or late swing phase. That is, at any given moment, only one foot is in the swing phase, and the other foot is in the support phase, ensuring that the robot always has a support foot to maintain balance.

[0065] Phase states are managed by a state machine, which is updated once per control cycle. The state transition conditions are as follows: when the swing leg completes the later stage of swing and the ground contact detection is triggered, it transitions from the later stage of swing to the support stage; when the support leg reaches the end time of the support stage, it transitions from the support stage to the early stage of swing.

[0066] Coordination between the left and right feet is achieved through a cross-foot trigger signal, which includes two types of signals: First, the ready signal. After the first foot completes the loading of the later swing phase, a ready signal is sent to the second foot. Upon receiving this signal, the second foot advances its entry into the early swing phase by one step length. The step length is dynamically adjusted according to walking speed, using the following formula: ,in As the reference step size, To adjust the coefficient, This is the walking speed. This mechanism allows the second foot to begin the anticipatory operation of the next cycle in advance, achieving seamless connection between the left and right feet.

[0067] Second, anticipate correction signals. When the first support level is detected... or or At that time, a prediction correction signal is sent to the second foot. After receiving the signal, the second foot adjusts the conservative coefficient of the prediction parameters for the next swing phase. Add a preset increment. The initial value of the conservatism coefficient is 1.0. When At that time, when generating the prediction instruction packet, , , Make corrections: Adjusted to , Adjusted to , Adjusted to When no prediction error is detected in multiple consecutive gait cycles, Gradually decrease until it returns to 1.0.

[0068] 6. Online evolution of predictive models This implementation also includes an online model evolution mechanism. During each template replacement operation in the support period, the source of error corresponding to that operation is recorded. Let the number of template replacements corresponding to a certain type of error be denoted as . The number of times threshold is .

[0069] when This indicates that this type of error occurs frequently, and the existing feature template library is insufficient. Add a feature template to the feature template library corresponding to this type of error.

[0070] To address slope recognition errors, terrain features at the current ground contact time are added as a new slope judgment reference template. These terrain features include... , , The vector formed by these vectors.

[0071] To address hardness classification errors, the impact characteristics at the current moment of impact are added as a new hardness judgment reference template. The impact characteristics include peak impact force, rise time, and waveform characteristics.

[0072] To address slip risk assessment errors, slip characteristics from the current support period are added as a new reference template for slip risk judgment. Slip characteristics include slip distance, slip velocity, and pressure center trajectory.

[0073] The feature template library initially contains several standard templates, and the number of templates gradually increases as walking experience accumulates. When the number of templates reaches the upper limit, the least recently used strategy is used to replace the least frequently used templates. That is, the last usage time of each template is recorded, and when a replacement is needed, the template with the earliest last usage time is selected for replacement.

[0074] When generating the prediction instruction package in the early stage of subsequent swing, the feature matching process uses both the original template and the newly added template to calculate the similarity. The similarity between the current feature and the features of each template is calculated by Euclidean distance, and the template with the highest matching degree is selected as the judgment basis.

[0075] 7. Overall Control Process At the start of a gait cycle, the swing leg enters the pre-swing phase. The system collects plantar pressure distribution data and local terrain depth maps at the moment the current foot leaves the ground, and extracts... , , Terrain features, generating content including , , The predicted instruction package is stored in the cache.

[0076] Subsequently, the swing leg enters the later stage of the swing. The system monitors the height of the swing leg from the ground in real time. ,when The first level of loading is executed, rotating the ankle joint to... Set the stiffness coefficient to The corresponding value; when When performing the second level of loading, the damping coefficient is set to... The corresponding value is adjusted and the knee flexion angle is increased.

[0077] The moment the foot touches the ground, the swinging leg enters the support phase. (System reads) , , ,calculate , , ,and , , The comparison is performed, and the slope judgment threshold in the prediction model is updated or the feature template is replaced based on the comparison results. The updated prediction model is used when generating the prediction instruction package in the early stage of the next swing.

[0078] Simultaneously, the left and right feet execute asynchronously in parallel. When one foot completes the late-stage loading of the swing, it sends a ready signal to the other foot, causing the other foot to enter the early-stage swing ahead of time; when one foot detects a prediction error, it sends a prediction correction signal to the other foot, causing the other foot to increase the conservatism coefficient of the prediction parameters.

[0079] When the template replacement count for a certain type of error Exceed When this happens, the system adds new feature templates to the feature template library corresponding to that type of error, thus enabling model evolution.

[0080] Using the above method, the robot can adjust its joint parameters before its feet touch the ground when walking on complex terrain, achieving real-time adaptation to terrain changes. This method can be applied to the stable walking control of humanoid robots in various complex terrain conditions such as gravel roads, soft sand, slippery roads, and uneven ground.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A control method for an AI humanoid robot suitable for complex terrain, characterized in that, Includes the following steps: S1: Divide the single-step phase cycle of the humanoid robot into three phases: the early swing phase, the late swing phase, and the support phase; S2: In the early stage of the swing, a prediction instruction package is generated based on the plantar pressure distribution data at the moment the current foot leaves the ground and the local terrain depth map of the next foot landing point collected by the visual sensor. The prediction instruction package includes the expected ankle joint contact angle, the expected ground stiffness level and the expected plantar slip level. S3: In the later stage of the swing, the expected ankle joint contact angle, the expected ground stiffness level and the expected foot slip level are loaded into the ankle joint controller in layers according to the height of the swinging leg from the ground. S4: During the support phase, compare the plantar pressure distribution data and the actual ankle contact angle at the moment the foot touches the ground with the prediction instruction package, identify the source of deviation, and update the prediction model based on the identification results; S5: The three phases of the left and right feet are executed asynchronously in parallel, and the left and right feet are coordinated by the phase state mutual exclusion rule and the cross-foot trigger signal.

2. The AI ​​humanoid robot control method for complex terrain according to claim 1, characterized in that, The initial swing phase is the time interval from the moment the foot leaves the ground to the moment the foot rises to its highest point; the later swing phase is the time interval from the moment the foot begins to descend from its highest point to the moment the foot touches the ground; and the support phase is the time interval from the moment the foot touches the ground to the moment the foot leaves the ground again.

3. The AI ​​humanoid robot control method for complex terrain according to claim 1, characterized in that, The generation of the prediction instruction package in step S2 specifically includes: Extract the slope value of the landing point area from the local terrain depth map. Surface texture density value at landing point and the magnitude of the change in landing point depth ; Will Assign a value to the expected ankle contact angle. ,Will and Mapped to the expected ground stiffness level ,Will and Mapped to the expected plantar slip level .

4. The AI ​​humanoid robot control method for complex terrain according to claim 3, characterized in that, The expected ankle contact angle The expected ground stiffness level is the ankle joint angle value when the foot strikes the ground next. The expected plantar slip level is the graded value for the ground hardness at the next footing point. The graded value is used to indicate the probability of the foot sliding at the next landing point.

5. The AI ​​humanoid robot control method for complex terrain according to claim 1, characterized in that, Let the height of the swing leg from the ground be... The first threshold is The second threshold is ,and The hierarchical loading in step S3 specifically includes: when At that time, rotate the ankle joint to Set the ankle joint stiffness coefficient to The corresponding stiffness value; when At that time, the ankle joint damping coefficient was set to The corresponding damping value will increase the knee flexion angle. The corresponding buckling increment.

6. The AI ​​humanoid robot control method for complex terrain according to claim 1, characterized in that, Let the actual ground contact angle of the ankle joint be... The expected ankle contact angle is Let the stiffness level corresponding to the actual plantar impact characteristics be . The expected ground stiffness level is Let the actual foot slip distance correspond to the slip level as . The expected plantar slippage level is Step S4, which involves identifying the sources of deviation and updating the prediction model based on the identification results, specifically includes: calculate , , ; Let the angle deviation threshold be The stiffness deviation threshold is The slip deviation threshold is ; when At that time, the slope judgment threshold in the prediction model is set to ; when When the time comes, the soft and hard judgment feature template in the prediction model will be replaced with the template corresponding to the current plantar impact feature; when When the slip risk judgment feature template in the prediction model is replaced with the template corresponding to the current plantar slip feature, the template will be replaced with the template corresponding to the current plantar slip feature.

7. The AI ​​humanoid robot control method for complex terrain according to claim 6, characterized in that, The asynchronous parallel execution of the three phases of the left and right feet in step S5 specifically includes: When the first foot is in the early or late swing phase, the second foot is in the support phase. After the first foot completes the loading of the later stage of the swing, it sends a ready signal to the second foot. After receiving the ready signal, the second foot will advance the timing of entering the early stage of the swing by one step. When the first foot is detected during the support phase or or When the second foot receives the prediction correction signal, it sends a prediction correction signal to the second foot. After receiving the prediction correction signal, the second foot increases the conservative coefficient of the prediction parameter in the early stage of the next swing by a preset increment.

8. The AI ​​humanoid robot control method for complex terrain according to claim 7, characterized in that, The phase state mutual exclusion rule is as follows: when one foot is in the early or late swing phase, the other foot is in the support phase; when one foot is in the support phase, the other foot is in the early or late swing phase.

9. The AI ​​humanoid robot control method for complex terrain according to claim 6, characterized in that, It also includes step S6: recording the template replacement operation performed during each support period, assuming the number of template replacements corresponding to a certain type of error is . The number of times threshold is ,when When the error occurs, add a feature template to the feature template library corresponding to that type of error.

10. The AI ​​humanoid robot control method for complex terrain according to claim 5, characterized in that, The parameters loaded into the ankle joint controller in step S3 are executed the moment the foot touches the ground.

Images