A humanoid robot gait naturalness intelligent evaluation training method and device

Abstract

The application provides a humanoid robot gait naturalness intelligent evaluation training method and device, comprising: acquiring corresponding reference force exertion timing records from a historical force exertion timing database for the working condition label, and determining the feasible range of force exertion timing; when the step speed variation amplitude exceeds a preset safety threshold and the friction resistance level is lower than a preset stability level, adjusting the force exertion timing delay; updating the initial force exertion timing using the adjusted force exertion timing delay, and iteratively comparing and correcting the feasible range of the force exertion timing to obtain an optimal take-off force exertion time; generating a force exertion timing dynamic adjustment scheme according to the comparison between the optimal take-off force exertion time and the metatarsal flexion torque peak time; and performing propulsion power and slipping risk evaluation on the force exertion timing dynamic adjustment scheme, and outputting the verified force exertion timing scheme after confirming that the requirements are met.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and apparatus for intelligent evaluation and training of the naturalness of humanoid robot gait. Background Technology

[0002] In the field of humanoid robot research, intelligent assessment and optimization of gait naturalness is a crucial direction for improving robot motion performance, directly impacting the robot's adaptability and practical value in complex environments. A natural and stable gait is not only a key indicator of a robot's human-like walking ability but also the foundation for its safe and efficient operation in everyday life scenarios. However, existing methods for gait optimization often neglect the dynamic balance between different walking speeds and ground conditions. Many solutions focus only on gait adjustment under a single working condition, failing to fully consider the comprehensive impact of external environmental changes on robot walking stability. Especially when ground friction varies significantly, traditional gait adjustment strategies often prove inadequate, leading to instability during robot walking. A deeper technical challenge lies in the precise control of ankle force exertion timing—a core factor. Ankle force exertion timing refers to the specific moment when the ankle joint generates thrust to propel the robot forward; the choice of this timing directly determines propulsion efficiency and walking stability. Improper timing of ankle force application can trigger a series of chain reactions. For example, during rapid walking, applying force too early may cause slippage due to insufficient ground friction, affecting overall balance; conversely, applying force too late will miss the optimal propulsion opportunity, reducing walking efficiency. This contradiction in timing is particularly pronounced under different walking speeds and ground conditions, becoming a pressing problem in gait optimization. Specifically, in real-world scenarios, assuming a robot needs to switch from slow to fast walking on a smooth surface, the timing of ankle force application becomes exceptionally complex. Applying force too early may cause slippage, loss of balance, or even a fall; while delaying force application may result in insufficient propulsion, affecting walking speed and rhythm, making it difficult to achieve the desired movement effect. Therefore, dynamically determining the ankle force application timing under different walking speeds and ground friction conditions to balance propulsion efficiency and walking stability has become a critical issue that urgently needs to be addressed. Summary of the Invention

[0003] This invention provides an intelligent evaluation and training method for the naturalness of gait in humanoid robots, mainly comprising: The plantar flexion torque, plantar pressure, and forward acceleration of the swing leg are collected to determine the stride speed variation range and friction resistance level. Based on the stride speed variation range and friction resistance level, the working condition label is obtained by support vector machine classification. The initial force application timing is determined based on the working condition label and the forward acceleration of the swing leg. For the aforementioned working condition label, the corresponding reference power generation time record is obtained from the historical power generation time database to determine the feasible range of power generation time. When the change in step speed exceeds a preset safety threshold and the frictional resistance level is lower than a preset stable level, adjust the delay in the timing of force application. The initial force exertion timing is updated using the adjusted force exertion timing delay, and iteratively compared and corrected with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time; Based on the comparison between the optimal push-off moment and the peak moment of plantar flexion torque, a dynamic adjustment scheme for the timing of force exertion is generated; The dynamic adjustment scheme for the timing of force application is evaluated for propulsion force and slippage risk. Once the requirements are confirmed to be met, the verified timing of force application scheme is output.

[0004] Furthermore, plantar flexion torque, plantar pressure, and forward acceleration of the swing leg are collected to determine the range of gait speed variation and frictional resistance level, including: Scan the plantar flexion torque signal along the time axis, locate the inflection point where the amplitude changes from rising to falling, and record the peak time corresponding to the inflection point; Read the pressure of the plantar pressure sensor in the heel area, and mark the time when the heel leaves the ground when the pressure in the heel area drops below a preset release threshold. The gait speed change amplitude is calculated based on the difference in cycle duration between the peak moment and the heel-off moment in adjacent gait cycles. The tangential and normal forces during the contact phase between the foot and the ground are collected, and the friction resistance level is determined based on these forces.

[0005] Furthermore, determining the initial force application timing based on the aforementioned working condition label and the forward acceleration of the swing leg includes: Obtain the timestamp of the peak forward acceleration of the swing, and calculate the difference between the timestamp and the time of heel lift-off to obtain the push-off response delay. If the push-off response delay is greater than a preset delay threshold, then half of the push-off response delay is added as compensation based on the heel lift-off time to obtain the initial power generation timing. If the push-off response delay is not greater than the preset delay threshold, then the moment when the heel leaves the ground is taken as the initial force application timing.

[0006] Furthermore, for the aforementioned operating condition label, corresponding reference power-generating timing records are obtained from the historical power-generating timing database to determine the feasible range of power-generating timings, including: For the aforementioned operating condition label, retrieve all reference power generation time records corresponding to the aforementioned operating condition label from the historical power generation time database; Extract each power exertion moment from all retrieved reference power exertion moment records, and arrange them in ascending order to obtain the reference moment sequence; Read the minimum force exertion time at the very beginning and the maximum force exertion time at the very end of the reference timing sequence, and use the minimum force exertion time as the lower bound and the maximum force exertion time as the upper bound to determine the feasible range of force exertion timing.

[0007] Furthermore, the initial power-generating timing is updated using the adjusted power-generating timing delay, including: The adjusted power-generating timing delay is superimposed on the initial power-generating timing to obtain the updated power-generating timing. The updated power-generating timing is compared with the feasible range of the power-generating timing. If it is lower than the lower bound of the feasible range of the power-generating timing, it is replaced by the lower bound of the feasible range of the power-generating timing. If it is higher than the upper bound of the feasible range of the power-generating timing, it is replaced by the upper bound of the feasible range of the power-generating timing, thus obtaining the corrected power-generating timing.

[0008] Furthermore, the dynamic adjustment scheme for the timing of force application is assessed for propulsion force and slippage risk. After confirming that the requirements are met, the verified timing of force application scheme is output, including: Based on the adjustment direction and time offset range in the dynamic adjustment scheme for the timing of force exertion, and under the current step speed change range and friction resistance level, apply ankle plantar flexion torque at the optimal push-off moment; The tangential force between the sole of the foot and the ground at the moment of push-off is collected as the propulsive force value, the normal force is collected, and the actual friction utilization ratio is calculated. The actual friction utilization ratio is compared with the static friction upper limit value corresponding to the friction resistance level to determine the slippage risk mark; Compare the propulsion force with the preset propulsion efficiency threshold; If the propulsion force value is not lower than the propulsion efficiency threshold and the slippage risk is marked as low, then the dynamic adjustment scheme for the propulsion timing is used as the verified propulsion timing scheme output.

[0009] Furthermore, based on the comparison between the optimal push-off moment and the peak moment of plantar flexion torque, a dynamic adjustment scheme for the timing of force application is generated, including: Compare the optimal push-off moment with the peak moment of plantar flexion torque; If the optimal push-off moment is earlier than the peak moment of plantar flexion torque, then the adjustment direction is determined to be advanced. If the optimal push-off moment is later than the peak moment of plantar flexion torque, the adjustment direction is determined to be lagging. The absolute value of the difference between the optimal push-off moment and the peak moment of plantar flexion torque is used as the time offset amplitude; A dynamic adjustment scheme for the timing of force application is generated based on the adjustment direction and the time offset amplitude. The dynamic adjustment scheme for the timing of force application includes an adjustment direction marker and a corresponding time offset amplitude.

[0010] This invention provides an intelligent evaluation and training device for the naturalness of gait in a humanoid robot, mainly comprising: The data acquisition and working condition classification module is used to collect plantar flexion torque, plantar pressure and forward acceleration of the swing leg, determine the step speed change range and friction resistance level, obtain working condition labels through support vector machine classification based on the step speed change range and friction resistance level, and determine the initial force application timing based on the working condition labels and forward acceleration of the swing leg. The feasible range determination module is used to obtain the corresponding reference exertion timing record from the historical exertion timing database for the working condition label, and determine the feasible range of the exertion timing. The delay adjustment module is used to adjust the delay of the force exertion timing when the change in step speed exceeds a preset safety threshold and the friction resistance level is lower than a preset stable level. The iterative correction module is used to update the initial force exertion timing with the adjusted force exertion timing delay, and iteratively compare and correct it with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time. The dynamic adjustment scheme generation module is used to generate a dynamic adjustment scheme for the timing of force exertion based on the comparison between the optimal push-off force exertion time and the peak time of plantar flexion torque. The scheme evaluation and output module is used to evaluate the propulsion force and slippage risk of the dynamic adjustment scheme for the force application timing, and output the verified force application timing scheme after confirming that the requirements are met.

[0011] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses an intelligent evaluation and training method and device for the naturalness of humanoid robot gait. It captures the peak moment of plantar flexion torque, heel lift-off moment, and forward acceleration of the swing leg in real time, extracts the gait speed variation amplitude and friction resistance level as the main features, and inputs them into a support vector machine to achieve intelligent classification of current complex terrain or dynamic disturbance conditions and assign condition labels. Simultaneously, based on a historical force application timing database, it retrieves reference force application ranges under matching conditions. In high-risk scenarios with drastic gait speed changes and insufficient friction, which are prone to slippage, it uses the heel lift-off moment and forward acceleration of the swing leg as independent variables, employs the least squares method to fit the real-time signal and identify the delay amplitude, accurately calculating the adjusted force application timing delay. Subsequently, it iteratively updates the initial force application timing and verifies whether it falls within a feasible range, ultimately approximating the optimal push-off moment and generating a dynamic adjustment scheme accordingly. Through simulation evaluation of the adjustment scheme's propulsion force and slippage risk, after confirming that it meets the dual requirements of propulsion efficiency and stability, it outputs the verified optimal force application timing scheme. This invention solves the core problem of stiff gait and easy slipping caused by the difficulty in accurately matching the timing of the push-off force when humanoid robots are moving on variable speed and unstable friction surfaces. It significantly improves the naturalness of gait, propulsion efficiency and walking stability under complex working conditions. Attached Figure Description

[0012] Figure 1 This is a flowchart of an intelligent evaluation and training method for the naturalness of gait in a humanoid robot according to the present invention.

[0013] Figure 2 This is a schematic diagram of the structure of an intelligent evaluation and training device for the naturalness of gait of a humanoid robot according to the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0015] like Figures 1-2 This embodiment of a humanoid robot gait naturalness intelligent assessment training method and device may specifically include: Step S101: Collect plantar flexion torque, plantar pressure, and forward acceleration of the swing leg to determine the stride speed variation range and friction resistance level. Based on the stride speed variation range and friction resistance level, obtain the working condition label through support vector machine classification, and determine the initial force application timing based on the working condition label and the forward acceleration of the swing leg. Plantar flexion torque signals are collected from ankle joint sensors during the robot's walking process. The amplitude changes in these signals are scanned frame-by-frame along the time axis to pinpoint the inflection point where the amplitude changes from rising to falling, and the peak moment corresponding to this inflection point is recorded. Simultaneously, the pressure distribution in the heel region is read from plantar pressure sensors. When the pressure value in the heel region drops below a preset release threshold, this time point is marked as the heel lift-off moment. Forward linear acceleration is extracted from the inertial measurement units of the swing leg's knee and hip joints. This forward linear acceleration is then averaged and filtered according to the gait cycle to obtain the swing forward acceleration curve. Based on the difference in cycle duration between the peak moment and the heel lift-off moment in adjacent gait cycles, the stride speed change amplitude is calculated. c represents the gait speed variation amplitude, d represents the difference in cycle duration between the peak moment and the heel lift-off moment in adjacent gait cycles, p represents the gait cycle parameter, and b represents the baseline gait speed. The tangential and normal forces during the foot-ground contact phase are collected, their ratio is calculated, and the current friction resistance level is determined according to the levels corresponding to each ratio range in a pre-established friction resistance level table. The gait speed variation amplitude and the friction resistance level are combined to form a feature vector, which is input into a pre-trained support vector machine. The support vector machine classifies the working condition categories based on the projection position of the feature vector on the classification hyperplane and outputs the corresponding working condition labels. For the aforementioned working condition label, the timestamp of the peak acceleration occurrence is read from the forward acceleration curve of the swing. The difference between this timestamp and the heel-off moment is calculated to obtain the push-off response delay. If the push-off response delay is greater than a preset delay threshold, half of the push-off response delay is added as compensation based on the heel-off moment to obtain the initial force application timing under the current gait condition. If the push-off response delay is not greater than the delay threshold, the heel-off moment is directly used as the initial force application timing.

[0016] In one embodiment, a torque sensor is embedded in the robot's ankle joint. This sensor continuously records the torque value in the plantar flexion direction along the time axis at a fixed sampling frequency. The plantar flexion torque signal exhibits periodic fluctuation characteristics. Within each gait cycle, the torque value gradually increases from heel strike, reaches its peak during the push-off phase, and then rapidly decreases. The amplitude of adjacent sampling points is compared frame by frame for the plantar flexion torque signal. When the amplitude of the previous sampling point is higher than that of the next sampling point, and the amplitude of the previous sampling point is higher than that of the sampling point before that, the previous sampling point is determined to be the inflection point where the amplitude changes from increasing to decreasing. The time coordinate of this inflection point is recorded as the peak moment.

[0017] Specifically, plantar pressure sensors are deployed in the heel and forefoot areas, outputting normal pressure values ​​for the corresponding areas. As the robot reaches the end of its single-leg support phase, the heel gradually lifts, and the pressure value in the heel area continuously decreases. The release threshold corresponds to the critical pressure level at which the heel area is about to lose contact with the ground. When the pressure value in the heel area drops below this release threshold, this moment is marked as the moment the heel leaves the ground.

[0018] For example, during the forward swing of the swing leg after it leaves the ground, the inertial measurement units at the knee and hip joints output triaxial acceleration data, from which the forward linear acceleration component along the walking direction is extracted. Since the original forward linear acceleration is mixed with high-frequency noise caused by joint vibration, the forward linear acceleration is mean-filtered according to the gait cycle. That is, the arithmetic mean of all forward linear acceleration samples in each gait cycle is taken, and this average value is used to replace the original value of each sampling point in the cycle to obtain the smoothed swing forward acceleration curve.

[0019] In one possible implementation, the magnitude of the gait rate change is reflected by the duration difference between two adjacent gait cycles.

[0020] Specifically, the duration from heel strike to the next heel strike on the same side is recorded for each of two adjacent gait cycles, and the absolute value of the difference between the two is calculated. A larger absolute value indicates a greater variation in gait speed. The determination of the friction resistance level relies on force sensing data during the foot-ground contact phase. Tangential and normal forces acting on the foot are collected during the support phase, and the ratio of tangential force to normal force is calculated. The friction resistance level table is pre-divided into several ratio intervals based on the friction characteristics of different ground materials. For example, a ratio below the upper limit of a first preset interval corresponds to a low friction resistance level; a ratio between the upper limit of the first and second preset intervals corresponds to a medium friction resistance level; and a ratio above the upper limit of the second preset interval corresponds to a high friction resistance level. The ratio is compared one by one with the boundaries of each interval, and the interval falling into is determined as the corresponding friction resistance level. The support vector machine is a classification method based on statistical learning theory. Its core is to find a hyperplane that separates samples of different categories as much as possible on both sides of the hyperplane, and maximizes the distance from the nearest sample point to the hyperplane. During the training phase, the step speed variation and friction resistance level, already labeled with work condition categories in historical walking data, are used as training samples. The support vector machine determines the position and orientation of the hyperplane by solving a quadratic programming problem. During the classification phase, the feature vector composed of the current step speed variation and friction resistance level is projected onto the space containing the hyperplane. Based on which side of the hyperplane the feature vector falls on, the system is classified into the corresponding work condition category, and a work condition label for that category is output. The work condition label represents the comprehensive type of the current walking conditions, including the combined state information of step speed and ground friction.

[0021] Understandably, the push-off response delay reflects the lag between the peak time of the swing leg's forward acceleration and the heel lift-off time. The push-off response delay is obtained by locating the peak acceleration time from the forward acceleration curve and subtracting the heel lift-off time from that time. If the push-off response delay exceeds a preset delay threshold, it indicates a significant time lag in the push-off action. In this case, half of the push-off response delay is added backwards from the heel lift-off time as a compensation to obtain the initial power generation timing under the current gait condition. If the push-off response delay is not greater than the delay threshold, the heel lift-off time itself serves as the initial power generation timing, requiring no additional compensation.

[0022] Step S102: For the working condition label, obtain the corresponding reference power generation time record from the historical power generation time database to determine the feasible range of power generation time.

[0023] For the aforementioned working condition label, all reference force application timing records corresponding to the working condition label are retrieved from the historical force application timing database. The historical force application timing database uses a combination of step speed change amplitude and friction resistance level as the search index. The force application time of each record is extracted from all retrieved reference force application timing records and arranged in ascending order of the force application time to obtain a reference timing sequence. The minimum force application time at the beginning and the maximum force application time at the end of the reference timing sequence are read from the sequence. Using the minimum force application time as the lower bound and the maximum force application time as the upper bound, the feasible range of force application timings is determined.

[0024] The historical force application timing database pre-stores records of force application timing collected under different walking conditions. Each record includes the corresponding step speed change range, friction resistance level, and the actual moment of force application under that condition. The database uses the combination of step speed change range and friction resistance level as a search index, with several historical records corresponding to the same combination.

[0025] Specifically, once the working condition label is determined, based on the step speed change range and friction resistance level contained in the working condition label, all reference force application time records that match the combination are located in the historical force application time database.

[0026] For example, when the working condition label indicates a moderate pace variation and a low frictional resistance level, the search results are all the moments of force exertion accumulated during each walk under this combination of conditions.

[0027] In one possible implementation, the power-up timing is read one by one from all retrieved reference power-up timing records and arranged in ascending order of value to form a reference timing sequence. This reference timing sequence reflects the distribution range of historical power-up timings under the same operating conditions. Based on this reference timing sequence, the minimum power-up timing at the beginning and the maximum power-up timing at the end are read. Using the minimum power-up timing as the lower bound and the maximum power-up timing as the upper bound, the feasible range of power-up timings is defined. This feasible range indicates that, under the current operating conditions, the value of the power-up timing is neither lower than the earliest time in the historical records nor higher than the latest time in the historical records.

[0028] Step S103: When the change in step speed exceeds a preset safety threshold and the frictional resistance level is lower than a preset stable level, adjust the delay in the timing of force application.

[0029] The stride speed variation is compared with a preset safety threshold, and the friction resistance level is compared with a preset stability level. If the stride speed variation exceeds the safety threshold and the friction resistance level is lower than the stability level, the heel-off time and the corresponding forward acceleration peak of the swing leg are extracted from several consecutive gait cycles to form a heel-off time sequence and a forward acceleration peak sequence, respectively. Using the heel-off time sequence as the independent variable and the forward acceleration peak sequence as the dependent variable, the correspondence between the two is fitted using the least squares method, that is, a set of fitting coefficients is obtained to minimize the sum of squares of the deviations between each measured forward acceleration peak and the fitted value, thus obtaining a fitting curve. The measured forward acceleration peak corresponding to the heel-off time of the current gait cycle is collected, and the measured forward acceleration peak is substituted into the fitting curve to obtain the corresponding predicted forward acceleration peak. The difference between the measured forward acceleration peak and the predicted forward acceleration peak is calculated to obtain the acceleration deviation. The slope value at the current heel lift-off moment is obtained by taking the derivative of the fitted curve. The adjusted force exertion timing delay is calculated by dividing the acceleration deviation by the slope value.

[0030] The safety threshold is a critical value for the range of gait changes determined in advance based on robot walking stability tests. When the range of gait changes exceeds this critical value, it indicates that the robot is in a walking phase with rapidly changing gait speed. The stability level is a boundary line for friction resistance levels pre-defined based on the friction characteristics of different ground materials. When the friction resistance level is lower than this boundary line, it indicates that the friction force provided by the current ground is too low. When both the range of gait changes exceeds the safety threshold and the friction resistance level is lower than the stability level, an adjustment process for the delay in force application is triggered.

[0031] For example, when a robot switches from slow to fast walking on a smooth surface, the change in walking speed will increase significantly, while the level of ground friction resistance is in a low range. At this time, the above two conditions are met, that is, it enters the working condition range of delay adjustment.

[0032] Specifically, the heel-off-ground time sequence is obtained by: tracing back several consecutive gait cycles from the current moment, reading the time points in each gait cycle when the heel pressure value drops below a pre-set release threshold, and arranging them chronologically. The release threshold is typically set to 0.1 Newtons to ensure complete heel liftoff. Correspondingly, the forward acceleration peak sequence is obtained by: locating the peak value of acceleration on the swing forward acceleration curve in each gait cycle, recording the acceleration value at that peak point, and arranging them in the same periodic order as the heel-off-ground time sequence to form paired data sets. The swing forward acceleration curve is derived from raw data collected by an acceleration sensor installed on the swing leg, processed by low-pass filtering. The least squares method is used to minimize the sum of squares of the deviations between the predicted and observed values ​​by adjusting the fitting coefficients. In this embodiment, the independent variable is the value at each moment in the heel-off-ground time sequence, and the dependent variable is the corresponding acceleration peak value in the forward acceleration peak sequence. The fitting process is as follows: A linear regression model is used, assuming a linear correspondence between the heel-off time and the peak forward acceleration. Linear regression is preferred due to its simplicity and high fitting accuracy. Initial values ​​for the fitting coefficients are set. For each time value in the heel-off time sequence, the corresponding predicted acceleration peak is calculated by substituting it into the fitting function. This predicted peak is then subtracted from the corresponding measured acceleration peak in the forward acceleration peak sequence to obtain the deviation at each point. The squares of all deviations are summed to obtain the sum of squared deviations. The fitting coefficients are gradually adjusted to continuously reduce this sum of squared deviations until convergence. The determined fitting coefficients constitute the fitting curve. Historical data is used to construct the curve through the above sequence. The input is the time sequence, and the output is the peak prediction and slope. This fitting curve describes the mapping relationship between the heel-off time and the peak forward acceleration of the swing leg under the current operating conditions.

[0033] In one possible implementation, a first-order linear fitting is preferably used, where the fitted curve is a straight line and the fitting coefficients include both slope and intercept values. When there is a strong linear trend between the heel lift-off time and the peak forward acceleration, the first-order fitting can sufficiently reflect the correspondence between the two. Further, within the current gait cycle, the measured peak forward acceleration corresponding to the current heel lift-off time is read, and this current heel lift-off time is substituted into the fitted curve to obtain the corresponding predicted peak forward acceleration on the fitted curve. The measured peak forward acceleration comes from real-time sensor data, and the predicted peak forward acceleration comes from the mapped output of the fitted curve. The difference between the two is the acceleration deviation, which reflects the degree of deviation of the current gait cycle from the historical fitted pattern.

[0034] It is understood that the acceleration deviation is in the acceleration dimension, while the force release timing delay is in the time dimension. The conversion between the two is achieved using the slope value of the fitted curve at the current heel-off moment. The slope value is obtained by taking the first derivative of the function expression of the fitted curve with respect to the independent variable, and then substituting the current heel-off moment into the derivative expression. The physical meaning of the slope value is: the magnitude of change in the peak forward acceleration for each unit of time change in the heel-off moment. Dividing the acceleration deviation by the slope value converts the deviation in the acceleration dimension into an offset in the time dimension, and this offset is the adjusted force release timing delay.

[0035] For example, when the slope value is large, it indicates that a small change in the heel lift-off time causes a large change in the peak forward acceleration, and the same acceleration deviation corresponds to a small time offset; when the slope value is small, the same acceleration deviation corresponds to a large time offset. Through this conversion method, the force application timing delay can adaptively reflect the degree to which the real-time signal deviates from the fitted pattern under the current working condition.

[0036] Step S104: Update the initial force exertion timing using the adjusted force exertion timing delay, and iteratively compare and correct it with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time.

[0037] The adjusted power-generating timing delay is superimposed on the initial power-generating timing, i.e., the value of the initial power-generating timing is added to the value of the power-generating timing delay to obtain the updated power-generating timing. The updated power-generating timing is compared with the feasible range of the power-generating timing. If the updated power-generating timing is lower than the lower bound of the feasible range, the lower bound is used instead. If the updated power-generating timing is higher than the upper bound of the feasible range, the upper bound is used instead to obtain the corrected power-generating timing. Substituting the corrected acceleration timing into the fitted curve (using a linear regression model, with inputs being the historical acceleration timing t and the corresponding acceleration peak a, and output being the predicted a=f(t)) yields the corresponding predicted forward acceleration peak. The difference between this predicted peak and the measured forward acceleration peak is calculated to obtain a new acceleration deviation. This new acceleration deviation is divided by the slope value (i.e., the slope of the linear regression model) to obtain a new acceleration timing delay. This new acceleration timing delay is then superimposed on the corrected acceleration timing, and an over-limit correction is performed again. This superposition and correction process is repeated until the absolute value of the difference between two adjacent corrected acceleration timings is less than a preset convergence threshold, at which point iterative convergence is determined. During iterative convergence, the last corrected acceleration timing is taken as the optimal push-off acceleration moment and output.

[0038] When the adjusted power-generating timing delay is superimposed on the initial power-generating timing, the time value of the power-generating timing delay is directly added to the time value of the initial power-generating timing, and the sum is the updated power-generating timing.

[0039] For example, if the initial force exertion occurs at a certain point in the gait cycle after the heel leaves the ground, after adding the delay amount, this point in time will shift backward along the time axis of the gait cycle, and the magnitude of the shift depends on the size of the delay amount.

[0040] Specifically, the feasible range consists of a lower bound and an upper bound. The lower bound corresponds to the minimum exertion time under the same working condition in the historical exertion time record, and the upper bound corresponds to the maximum exertion time. The updated exertion time is compared with the lower bound and the upper bound one by one: if the value of the updated exertion time is less than the lower bound, it means that the time is earlier than the earliest time in the historical record, and in this case, the value of the lower bound is directly replaced by the updated exertion time; if the value of the updated exertion time is greater than the upper bound, it means that the time is later than the latest time in the historical record, and in this case, the value of the upper bound is directly replaced by the upper bound. After the above comparison and replacement, the corrected exertion time is obtained, and the corrected exertion time always falls within the feasible range.

[0041] It should be noted that the specific execution method of the iterative process is as follows: Substitute the corrected force application timing into the fitted curve, and obtain a predicted forward acceleration peak value along the functional relationship of the fitted curve. There is a difference between this predicted value and the measured forward acceleration peak value. Subtract the predicted value from the measured value to obtain a new acceleration deviation. Then, divide the new acceleration deviation by the slope of the fitted curve at the current heel-off moment to calculate a new force application timing delay. Add this new delay to the corrected force application timing to obtain a new round of updated force application timing. Perform the out-of-bounds comparison and substitution correction again to obtain a new round of corrected force application timing. This process is repeated, with each iteration generating a corrected force application timing.

[0042] In one possible implementation, the convergence threshold is a pre-set minimum time interval. When the absolute value of the difference between the corrected propulsion timing obtained from two adjacent iterations is less than this convergence threshold, the iteration process is considered to have converged, meaning the change in the corrected propulsion timing has stabilized. Upon iteration convergence, the value of the last corrected propulsion timing is taken as the optimal push-off propulsion moment. This optimal push-off propulsion moment is within the feasible range and has undergone multiple rounds of deviation correction based on the fitted curve, reflecting the balance between propulsion timing and walking stability under the current operating conditions.

[0043] Step S105: Based on the comparison between the optimal push-off moment and the peak moment of plantar flexion torque, a dynamic adjustment scheme for the timing of force exertion is generated.

[0044] The optimal push-off moment is compared with the peak plantar flexion moment. If the optimal push-off moment is earlier than the peak plantar flexion moment, the adjustment direction is determined to be advanced; if the optimal push-off moment is later than the peak plantar flexion moment, the adjustment direction is determined to be delayed. The absolute value of the difference between the two is used as the time offset. Based on the adjustment direction and the time offset, a dynamic adjustment scheme for the push-off timing is generated. The dynamic adjustment scheme for the push-off timing includes an adjustment direction marker and a corresponding time offset value.

[0045] The optimal push-off moment is the ankle joint push-off time point determined after iterative convergence, and the peak moment of plantar flexion torque is the time coordinate at which the amplitude of the ankle joint plantar flexion torque signal reaches an inflection point. By comparing the two values ​​on the same time axis, the sequential position of the optimal push-off moment relative to the peak moment of plantar flexion torque is determined.

[0046] Specifically, if the value of the optimal push-off moment is less than the value of the peak plantar flexion moment, it indicates that the push-off action occurs before the peak moment, and the adjustment direction is marked as advanced; if the value of the optimal push-off moment is greater than the value of the peak plantar flexion moment, it indicates that the push-off action occurs after the peak moment, and the adjustment direction is marked as delayed. Simultaneously, the absolute value of the difference between the optimal push-off moment and the peak plantar flexion moment is used to obtain the time offset.

[0047] In one embodiment, when the robot walks quickly on a low-friction surface, the adjustment direction is mostly delayed, with a relatively large time offset. When the robot walks at a constant speed on a normal surface, the adjustment direction is either advanced or delayed, with smaller offsets. Based on the adjustment direction marker and the time offset, a dynamic adjustment scheme for the timing of force application is formed. This dynamic adjustment scheme uses the adjustment direction marker to indicate whether the push-off moment should be shifted in an advanced or delayed direction, and the time offset magnitude indicates the specific order of magnitude of the shift.

[0048] Step S106: Assess the propulsion force and slippage risk of the dynamic adjustment scheme for the power application timing, and output the verified power application timing scheme after confirming that the requirements are met.

[0049] Based on the adjustment direction marker and time offset amplitude in the dynamic adjustment scheme for the force exertion timing, under the current step speed variation amplitude and friction resistance level, the ankle plantar flexion torque is applied at the optimal push-off moment. The tangential force between the sole of the foot and the ground at the moment of push-off is collected as the propulsive force value, and the normal force between the sole of the foot and the ground at the moment of push-off is also collected. The actual friction utilization ratio is obtained by dividing the propulsive force value by the normal force. The static friction upper limit value is obtained from the preset friction characteristic data corresponding to the current friction resistance level. The actual friction utilization ratio is compared with the static friction upper limit value. If the actual friction utilization ratio exceeds the static friction upper limit value, the slippage risk is marked as high; if the actual friction utilization ratio does not exceed the static friction upper limit value, the slippage risk is marked as low. For the slippage risk marker, the propulsion force value is compared with a preset propulsion efficiency threshold. If the propulsion force value is not lower than the propulsion efficiency threshold and the slippage risk marker is low, the propulsion efficiency and stability requirements are met. If the propulsion force value is lower than the propulsion efficiency threshold or the slippage risk marker is high, the requirements are not met. Based on the determination result, if the propulsion efficiency and stability requirements are met, the dynamic adjustment scheme for the propulsion timing is output as the verified propulsion timing scheme. If the requirements are not met, the current time offset is reduced by half, and the propulsion force acquisition and slippage risk determination are re-executed based on the reduced time offset until the requirements are met, resulting in the verified propulsion timing scheme.

[0050] Before the dynamic adjustment scheme for the timing of force application is confirmed, plantar flexion torque is applied to the ankle joint according to the optimal push-off moment. At the instant of the push-off action, the plantar pressure sensor simultaneously records the tangential force and normal force on the contact surface between the sole and the ground. The tangential force is the horizontal thrust exerted by the sole on the ground along the walking direction, which is the propulsive force value; the normal force is the pressure of the sole perpendicular to the ground.

[0051] Specifically, the actual friction utilization ratio means the proportion of the horizontal thrust exerted by the foot on the ground at the moment of push-off to the vertical pressure. The higher this ratio, the greater the dependence of the foot on the ground friction force during push-off. When this ratio approaches or exceeds the friction limit provided by the ground, the static friction state between the foot and the ground is broken, and sliding occurs.

[0052] It should be noted that the static friction upper limit is obtained by querying preset friction characteristic data based on the current friction resistance level. The preset friction characteristic data is obtained by conducting friction tests on different ground materials in advance, and recording the maximum ratio of tangential force to normal force allowed for each type of ground material under normal pressure. This maximum ratio is the static friction upper limit for the corresponding friction resistance level.

[0053] For example, a lower friction resistance level corresponds to a smaller upper limit of static friction, indicating that the ground will slide under a smaller horizontal thrust; a higher friction resistance level corresponds to a larger upper limit of static friction, indicating that the ground can withstand a larger horizontal thrust without sliding.

[0054] In one possible implementation, the actual friction utilization ratio is compared one by one with the static friction upper limit. If the actual friction utilization ratio exceeds the static friction upper limit, it indicates that the horizontal thrust at the moment of push-off has exceeded the bearing capacity of the ground friction, and the foot has the potential to slip. In this case, the slippage risk is set to high. If the actual friction utilization ratio does not exceed the static friction upper limit, it indicates that the horizontal thrust at the moment of push-off is still within the bearing capacity of the ground friction, and the foot does not have the potential to slip. In this case, the slippage risk is set to low. Through the above comparison, continuous numerical judgments are transformed into discrete risk level labels, which facilitates joint judgment with propulsion efficiency conditions. Furthermore, the propulsion efficiency threshold is determined in advance based on the minimum horizontal thrust that the robot should achieve to complete a normal push-off action at the target walking speed. When the propulsion force value is not lower than the propulsion efficiency threshold, it indicates that the propulsion force generated by the push-off action has reached the lower limit requirement for maintaining the target walking speed.

[0055] For example, the joint determination process is as follows: simultaneously reading the comparison results of the slippage risk marker and the propulsion force value against the propulsion efficiency threshold. Only when both conditions are met—the propulsion force value being not lower than the propulsion efficiency threshold and the slippage risk marker being low—is the current dynamic adjustment scheme for the power delivery timing determined to meet the propulsion efficiency and stability requirements. If either condition is not met, i.e., insufficient propulsion force or a risk of slippage exists, the requirements are determined not to be met. This joint determination mechanism ensures that propulsion efficiency and walking stability are considered simultaneously in the same evaluation process.

[0056] Understandably, when the judgment result is that the requirements are not met, the time offset amplitude is reduced to half of the current value. After reduction, based on the adjustment direction mark and the reduced time offset amplitude, the push-off force moment is re-determined, and the ankle plantar flexion torque is applied again at that moment. The propulsion force value and normal force are re-acquired, the actual friction utilization ratio is recalculated, and the slippage risk judgment and propulsion efficiency judgment are re-executed. This process is repeated cyclically, with the time offset amplitude gradually reduced in each iteration until the joint judgment result meets the requirements. When the judgment result meets the propulsion efficiency and stability requirements, the current dynamic adjustment scheme for the force exertion timing is output as the verified force exertion timing scheme. The verified force exertion timing scheme includes the verified adjustment direction mark and the corresponding time offset amplitude value. This scheme can provide sufficient push-off propulsion force under the current working conditions without causing foot slippage.

[0057] This invention provides an intelligent evaluation and training device for the naturalness of gait in a humanoid robot, mainly comprising: The data acquisition and working condition classification module is used to collect plantar flexion torque, plantar pressure and forward acceleration of the swing leg, determine the step speed change range and friction resistance level, obtain working condition labels through support vector machine classification based on the step speed change range and friction resistance level, and determine the initial force application timing based on the working condition labels and forward acceleration of the swing leg. The feasible range determination module is used to obtain the corresponding reference exertion timing record from the historical exertion timing database for the working condition label, and determine the feasible range of the exertion timing. The delay adjustment module is used to adjust the delay of the force exertion timing when the change in step speed exceeds a preset safety threshold and the friction resistance level is lower than a preset stable level. The iterative correction module is used to update the initial force exertion timing with the adjusted force exertion timing delay, and iteratively compare and correct it with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time. The dynamic adjustment scheme generation module is used to generate a dynamic adjustment scheme for the timing of force exertion based on the comparison between the optimal push-off force exertion time and the peak time of plantar flexion torque. The scheme evaluation and output module is used to evaluate the propulsion force and slippage risk of the dynamic adjustment scheme for the force application timing, and output the verified force application timing scheme after confirming that the requirements are met.

[0058] If the technical solution of this application involves personal information, the product using this solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If sensitive personal information is involved, the user's separate consent has been obtained before processing, and the "express consent" requirement is met. For example, a clear sign is placed at the collection device such as a camera to inform the user that they have entered the collection area, and the user's voluntary entry is considered as consent; or the processing device clearly indicates the processing rules and obtains authorization through pop-up windows or by asking the user to upload information themselves. The personal information processing rules include the processor, the purpose of processing, the processing method, and the types of personal information.

[0059] The above description is merely a preferred embodiment of this application and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in this application is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the concept of this application. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this application.

Claims

1. A method for intelligent evaluation and training of the naturalness of gait in a humanoid robot, characterized in that, The method includes: Collect plantar flexion torque, plantar pressure, and forward acceleration of the swinging leg to determine the range of gait speed variation and friction resistance level; Based on the step speed change range and the friction resistance level, a working condition label is obtained by support vector machine classification, and the initial force application timing is determined based on the working condition label and the forward acceleration of the swing leg. For the aforementioned working condition label, the corresponding reference power generation time record is obtained from the historical power generation time database to determine the feasible range of power generation time. When the change in step speed exceeds a preset safety threshold and the frictional resistance level is lower than a preset stable level, adjust the delay in the timing of force application. The initial force exertion timing is updated using the adjusted force exertion timing delay, and iteratively compared and corrected with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time; Based on the comparison between the optimal push-off moment and the peak moment of plantar flexion torque, a dynamic adjustment scheme for the timing of force exertion is generated; The dynamic adjustment scheme for the timing of force application is evaluated for propulsion force and slippage risk. Once the requirements are confirmed to be met, the verified timing of force application scheme is output.

2. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 1, characterized in that, The process of collecting plantar flexion torque, plantar pressure, and forward acceleration of the swing leg to determine the range of gait speed variation and friction resistance level includes: Scan the plantar flexion torque signal along the time axis, locate the inflection point where the amplitude changes from rising to falling, and record the peak time corresponding to the inflection point; Read the pressure of the plantar pressure sensor in the heel area, and mark the time when the heel leaves the ground when the pressure in the heel area drops below a preset release threshold. The gait speed change amplitude is calculated based on the difference in cycle duration between the peak moment and the heel-off moment in adjacent gait cycles. The tangential and normal forces during the contact phase between the foot and the ground are collected, and the friction resistance level is determined based on these forces.

3. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 2, characterized in that, The step of determining the initial force application timing based on the working condition label and the forward acceleration of the swing leg includes: Obtain the timestamp of the peak forward acceleration of the swing, and calculate the difference between the timestamp and the time when the heel leaves the ground to obtain the push-off response delay. If the push-off response delay is greater than a preset delay threshold, then half of the push-off response delay is added as compensation based on the heel lift-off time to obtain the initial power generation timing. If the push-off response delay is not greater than the preset delay threshold, then the moment when the heel leaves the ground is taken as the initial force application timing.

4. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 1, characterized in that, The step of retrieving corresponding reference power-generating timing records from the historical power-generating timing database for the aforementioned working condition label, and determining the feasible range of power-generating timings, includes: For the aforementioned operating condition label, retrieve all reference power generation time records corresponding to the aforementioned operating condition label from the historical power generation time database; Extract each power exertion moment from all retrieved reference power exertion moment records, and arrange them in ascending order to obtain the reference moment sequence; Read the minimum force exertion time at the very beginning and the maximum force exertion time at the very end of the reference timing sequence, and use the minimum force exertion time as the lower bound and the maximum force exertion time as the upper bound to determine the feasible range of force exertion timing.

5. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 1, characterized in that, The step of updating the initial power-generating timing using the adjusted power-generating timing delay includes: The adjusted power-generating timing delay is superimposed on the initial power-generating timing to obtain the updated power-generating timing. The updated power-generating timing is compared with the feasible range of the power-generating timing. If it is lower than the lower bound of the feasible range of the power-generating timing, it is replaced by the lower bound of the feasible range of the power-generating timing. If it is higher than the upper bound of the feasible range of the power-generating timing, it is replaced by the upper bound of the feasible range of the power-generating timing, thus obtaining the corrected power-generating timing.

6. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 1, characterized in that, The process of assessing the propulsion force and slippage risk of the dynamic adjustment scheme for the timing of force application, and outputting the verified timing of force application scheme after confirming that the requirements are met, includes: Based on the adjustment direction and time offset range in the dynamic adjustment scheme for the timing of force exertion, and under the current step speed change range and friction resistance level, apply ankle plantar flexion torque at the optimal push-off moment; The tangential force between the sole of the foot and the ground at the moment of push-off is collected as the propulsive force value, the normal force is collected, and the actual friction utilization ratio is calculated. The actual friction utilization ratio is compared with the static friction upper limit value corresponding to the friction resistance level to determine the slippage risk mark; Compare the propulsion force with the preset propulsion efficiency threshold; If the propulsion force value is not lower than the propulsion efficiency threshold and the slippage risk is marked as low, then the dynamic adjustment scheme for the propulsion timing is used as the verified propulsion timing scheme output.

7. The intelligent evaluation and training method for the naturalness of humanoid robot gait as described in claim 1, characterized in that, The step of generating a dynamic adjustment scheme for the timing of force exertion based on a comparison between the optimal push-off moment and the peak moment of plantar flexion torque includes: Compare the optimal push-off moment with the peak moment of plantar flexion torque; If the optimal push-off moment is earlier than the peak moment of plantar flexion torque, then the adjustment direction is determined to be advanced. If the optimal push-off moment is later than the peak moment of plantar flexion torque, the adjustment direction is determined to be lagging. The absolute value of the difference between the optimal push-off moment and the peak moment of plantar flexion torque is used as the time offset amplitude; A dynamic adjustment scheme for the timing of force application is generated based on the adjustment direction and the time offset amplitude. The dynamic adjustment scheme for the timing of force application includes an adjustment direction marker and a corresponding time offset amplitude.

8. A humanoid robot gait naturalness intelligent evaluation and training device, characterized in that, The device includes: The data acquisition and working condition classification module is used to collect plantar flexion torque, plantar pressure and forward acceleration of the swing leg, determine the step speed change range and friction resistance level, obtain working condition labels through support vector machine classification based on the step speed change range and friction resistance level, and determine the initial force application timing based on the working condition labels and forward acceleration of the swing leg. The feasible range determination module is used to obtain the corresponding reference exertion timing record from the historical exertion timing database for the working condition label, and determine the feasible range of the exertion timing. The delay adjustment module is used to adjust the delay of the force exertion timing when the change in step speed exceeds a preset safety threshold and the friction resistance level is lower than a preset stable level. The iterative correction module is used to update the initial force exertion timing with the adjusted force exertion timing delay, and iteratively compare and correct it with the feasible range of the force exertion timing to obtain the optimal push-off force exertion time. The dynamic adjustment scheme generation module is used to generate a dynamic adjustment scheme for the timing of force exertion based on the comparison between the optimal push-off force exertion time and the peak time of plantar flexion torque. The scheme evaluation and output module is used to evaluate the propulsion force and slippage risk of the dynamic adjustment scheme for the force application timing, and output the verified force application timing scheme after confirming that the requirements are met.

9. The intelligent evaluation and training device for the naturalness of humanoid robot gait as described in claim 8, characterized in that, The process of collecting plantar flexion torque, plantar pressure, and forward acceleration of the swing leg to determine the range of gait speed variation and friction resistance level includes: Scan the plantar flexion torque signal along the time axis, locate the inflection point where the amplitude changes from rising to falling, and record the peak time corresponding to the inflection point; Read the pressure of the plantar pressure sensor in the heel area, and mark the time when the heel leaves the ground when the pressure in the heel area drops below a preset release threshold. The gait speed change amplitude is calculated based on the difference in cycle duration between the peak moment and the heel-off moment in adjacent gait cycles. The tangential and normal forces during the contact phase between the foot and the ground are collected, and the friction resistance level is determined based on these forces.

10. The intelligent evaluation and training device for the naturalness of humanoid robot gait as described in claim 9, characterized in that, The step of determining the initial force application timing based on the working condition label and the forward acceleration of the swing leg includes: Obtain the timestamp of the peak forward acceleration of the swing, and calculate the difference between the timestamp and the time when the heel leaves the ground to obtain the push-off response delay. If the push-off response delay is greater than a preset delay threshold, then half of the push-off response delay is added as compensation based on the heel lift-off time to obtain the initial power generation timing. If the push-off response delay is not greater than the preset delay threshold, then the moment when the heel leaves the ground is taken as the initial force application timing.

Images