Humanoid robot stair climbing and descending limb coordination control method and device

Abstract

The application provides a humanoid robot upper and lower limb coordination control method and device, comprising: obtaining synchronization data of a robot pelvis twist phase and a stride cycle, combining foot end trajectory and pressure distribution data to form an initial posture data set; extracting a pelvis twist phase feature according to the filtered posture image, evaluating a roll margin divergence trend, generating a twist phase adjustment scheme corresponding to an upper body load distribution, and calibrating a critical posture recognition threshold; performing consistency verification on the foot end trajectory data through the critical posture recognition threshold, iteratively generating a pelvis twist phase control sequence according to the trajectory deviation obtained by verification, analyzing the roll margin divergence aggravation trend corresponding to the upper body load distribution, combining the pelvis twist phase control sequence and the roll margin divergence aggravation trend, and determining a stride cycle synchronization scheme; and generating a limb coordination stable control output sequence according to the stride cycle synchronization scheme.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method and device for controlling the coordination of limbs of a humanoid robot going up and down stairs. Background Technology

[0002] In the field of humanoid robot research, limb coordination control for climbing stairs is a crucial technology, directly related to the robot's adaptability and safety in complex environments. This not only affects the practical application of robots in home and industrial settings but also significantly contributes to improving the naturalness of robot-human interaction. With technological advancements, enabling robots to climb stairs smoothly and efficiently, much like humans, has become a core challenge in this field. Currently, while many solutions attempt to optimize robot stair climbing by mimicking human gait, these methods often overlook the interactions between different body parts in dynamic environments, especially when facing external disturbances, making control strategies inflexible. For example, when a robot carries an unbalanced load, relying solely on preset gait patterns struggles to handle sudden shifts in the center of gravity, leading to decreased stability and a risk of tilting or even falling. While pelvic rotation can optimize foot placement, bringing the feet closer to the body's midline for better stability on narrow stair treads, increasing the rotation angle when the upper body load shifts can exacerbate lateral swaying of the body's center of gravity, potentially causing a greater risk of imbalance instead of a stability-oriented movement. Specifically, in a real-world scenario, a humanoid robot is carrying heavy objects up a narrow staircase. Because the load is biased to one side, the robot's pelvic torsion angle fails to adapt in time when taking a step, causing its center of gravity to shift continuously to the side. This can even lead to a fall because the landing point is too close to the midline and the robot cannot effectively support its body. This imbalance between pelvic torsion and center of gravity stability under dynamic loads highlights the shortcomings of existing control methods. Therefore, how to coordinate the relationship between the pelvic torsion phase and the step cycle in real time under changing load conditions to ensure stable robot movement on narrow staircases has become a critical problem that urgently needs to be solved. Summary of the Invention

[0003] This invention provides a method for coordinating the movement of a humanoid robot up and down stairs, mainly including: Acquire synchronization data of the robot's pelvic torsion phase and step cycle, and combine them with foot trajectory and pressure distribution data to form an initial posture dataset; The initial posture dataset is subjected to noise filtering and smoothing to identify the instability features of pelvic torsion phase and step cycle synchronization and to define the instability state interval. The variation law of the tilt margin is analyzed within the instability state interval, and the filtered posture profile is determined based on the variation law. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. Based on the filtered posture profile, pelvic torsional phase features are extracted, the lateral tilt margin divergence is evaluated, a torsional phase adjustment scheme corresponding to the upper body load distribution is generated, and the critical posture recognition threshold is calibrated. The consistency of foot trajectory data is verified by the critical posture recognition threshold. Based on the trajectory deviation obtained from the verification, a pelvic torsion phase control sequence is generated iteratively. The trend of lateral tilt margin divergence corresponding to the upper body load distribution is analyzed. Combining the pelvic torsion phase control sequence and the trend of lateral tilt margin divergence, a step cycle synchronization scheme is determined. The limb coordination and stability control output sequence is generated based on the step-cycle synchronization scheme.

[0004] Furthermore, synchronization data of the robot's pelvic torsion phase and step cycle are acquired, and combined with foot trajectory and pressure distribution data to form an initial posture dataset, including: The angular velocity change trajectory of the torsional phase is obtained by an inertial sensor in the pelvis, and the gait switching beat within the step cycle is read by a hip joint encoder. The angular velocity change trajectory and the gait switching beat are aligned to form the synchronization data of the robot's pelvic torsional phase and step cycle. The lateral displacement curve of the swing leg foot retraction motion is extracted, and the narrowing of the distance between the two foot landing points is superimposed to obtain the original posture sampling sequence. Based on the original posture sampling sequence, the partition pressure readings of the plantar pressure array are called, and the lateral drift path of the pressure center on the tread is extracted according to the role switching of the swing leg and the supporting leg. The bias direction of the pressure distribution in each stride cycle is calibrated, the trunk acceleration is separated according to the lateral channel, and the peak frequency of the trunk lateral acceleration is extracted in the sliding time window through fast Fourier transform to obtain the oscillation frequency sequence. After aligning the oscillation frequency sequence with the synchronization data of the robot's pelvic torsion phase and step cycle, the initial posture dataset is established by fusing the lateral drift path and lateral displacement curve of the pressure center on the tread surface.

[0005] Furthermore, the initial attitude dataset is subjected to noise filtering and smoothing to identify instability features where the pelvic torsional phase and step cycle are synchronized and to define the instability state interval. Within the instability state interval, the variation law of the roll margin is analyzed, and based on this variation law, the filtered attitude profile is determined, including: For the multi-channel data of the initial attitude dataset, the Kalman filter algorithm is used to smooth the multi-channel data to form a smoothed multi-channel attitude timing sequence. The time when the pelvic torsional angular velocity crosses zero is compared with the time when the supporting leg switches during the stepping cycle. The time deviation is calculated. If the unidirectional accumulation of the time deviation exceeds a preset threshold, the period of instability is marked. Obtain the oscillation amplitude during the instability period, construct the tilt margin time series curve, track the turning point when the tilt margin changes from a converging trend of amplitude decaying step by step to a diverging trend of amplitude amplifying step by step, bind the multi-channel attitude time series before and after the turning point with the instability period label, and determine the filtered attitude profile.

[0006] Furthermore, pelvic torsional phase features are extracted from the filtered posture profile to assess the tilt margin divergence, including: The peak time, peak amplitude and zero-crossing interval of pelvic torsional angular velocity are extracted from the filtered posture image. Combined with the load bias direction and the lateral drift path of the pressure center, the pelvic torsional phase features are spliced ​​together. Based on the pelvic torsion phase characteristics, the values ​​of the tilt margin time series curve after the turning point are extracted along the time axis. The second difference of the tilt margin is calculated step by step to characterize the divergent acceleration. The divergent acceleration and the tilt margin amplitude amplification rate are concatenated to form a divergent evolution index sequence to generate a sample set. The input sample set is fed into the support vector machine classifier, which outputs a torsional phase adjustment scheme for different load distributions to determine the critical attitude recognition threshold.

[0007] Furthermore, the input sample set is fed into a support vector machine classifier, and the output is a torsional phase adjustment scheme for different load distributions. The critical pose recognition threshold is determined by: dividing the sample set into a training subset and a validation subset; using a support vector machine to train the classifier with the load bias direction label as the class label; outputting torsional phase adjustment schemes for three load distributions: left-biased, right-biased, and center-biased; and taking the boundary value of the decision function corresponding to the highest classification accuracy of the support vector machine on the validation subset to determine the critical pose recognition threshold.

[0008] Furthermore, the analysis of the increased divergence of lateral tilt margin corresponding to the upper body load distribution, combined with the pelvic torsion phase control sequence and the increased divergence of lateral tilt margin, determines the step cycle synchronization scheme, including: The associated trend curve is obtained by concatenating the oscillation frequency sequence along the timestamp. The phase is iteratively adjusted step by step with the goal of reducing the amplitude of the associated trend curve. After each iteration, the critical posture recognition threshold is brought back to re-verify the proportion of the maximum foot contraction that synchronously enters the preset narrowing interval, until the amplitude of the associated trend curve falls into the preset convergence interval and the foot lateral displacement curve contraction trend is synchronously matched, thus obtaining the pelvic torsion phase control sequence.

[0009] Furthermore, the step-by-step periodic synchronization scheme is determined, including: The waveform of the phase characteristic channel is reassembled according to the pelvic torsion phase control sequence. The tilt margin timing corresponding to the upper body load distribution is extracted along the time axis, the tilt margin is analyzed, and the step cycle synchronization scheme is generated.

[0010] Furthermore, after generating the limb coordination and stability control output sequence according to the step-cycle synchronization scheme, it includes: The pelvic torsion angle command, hip joint swing angle command, and lateral coordinate command of the foot landing point of the swinging leg in the limb coordination and stability control output sequence are distributed to the corresponding joint links. The target position, target velocity, and target torque of the actuator are calculated by inverse kinematics. The data is synchronously sent to the corresponding joint actuators along the timestamp to track the target position, target velocity, and target torque, thereby driving the robot body to complete the up and down stairs movement.

[0011] This invention provides a limb coordination control device for a humanoid robot going up and down stairs, the device comprising: The first module is used to acquire the synchronization data of the robot's pelvic torsion phase and step cycle, and combine it with foot trajectory and pressure distribution data to form an initial posture dataset. The second module is used to perform noise filtering and smoothing on the initial posture dataset, identify the instability characteristics of the pelvic torsion phase and the step cycle synchronization and delineate the instability state interval, analyze the change law of the tilt margin in the instability state interval, and determine the filtered posture profile based on the change law. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. The third module is used to extract pelvic torsional phase features based on the filtered posture profile, evaluate the lateral tilt margin divergence, generate a torsional phase adjustment scheme corresponding to the upper body load distribution, and calibrate the critical posture recognition threshold. The fourth module is used to verify the consistency of foot trajectory data through the critical posture recognition threshold, iteratively generate a pelvic torsion phase control sequence based on the trajectory deviation obtained from the verification, analyze the trend of lateral tilt margin divergence aggravation corresponding to the upper body load distribution, and determine the step cycle synchronization scheme by combining the pelvic torsion phase control sequence and the trend of lateral tilt margin divergence aggravation. The fifth module is used to generate a limb coordination and stability control output sequence based on the step-cycle synchronization scheme.

[0012] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects: This invention discloses a method and device for limb coordination control of a humanoid robot when climbing stairs. Addressing the problems of pelvic torsion phase and stride cycle synchronization instability, lateral tilt margin divergence, and insufficient consistency in foot trajectory convergence during stair climbing in operational scenarios, this invention collects an initial posture dataset and uses Kalman filtering for noise filtering and smoothing estimation to identify instability phenomena and determine the filtered posture profile. Then, it uses support vector machine to classify upper body load distribution features and determine the critical posture recognition threshold. This invention combines foot trajectory data consistency verification to iteratively generate a pelvic torsion phase control sequence, evaluates the exacerbation of lateral tilt margin divergence, and finally forms a stride cycle synchronization scheme and a limb coordination stability control output sequence to drive the robot to perform stair climbing actions. This invention significantly improves the stability and coordination of humanoid robots in complex terrain, effectively reduces the risk of instability, and achieves safe and efficient stair climbing capabilities. Attached Figure Description

[0013] Figure 1 This is a flowchart of the humanoid robot's limb coordination control method for going up and down stairs according to the present invention.

[0014] Figure 2 This is a schematic diagram of the structure of the humanoid robot's limb coordination control device for going up and down stairs according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0016] like Figure 1 The method for coordinating the limbs of a humanoid robot going up and down stairs in this embodiment may specifically include: S101. Acquire the synchronization data of the robot's pelvic torsion phase and step cycle, and combine them with the foot trajectory and pressure distribution data to form an initial posture dataset.

[0017] The angular velocity change trajectory of the robot's pelvic torsional phase is obtained by an inertial sensor deployed at the robot's pelvis. Combined with the gait switching beat within the stride cycle read by the hip joint encoder, the angular velocity change trajectory and the beat signal are aligned on the time axis to form a synchronous comparison record of the torsional phase and stride cycle. For the foot tuck-up action of the swing leg in the air phase, the end-effector lateral displacement curve is extracted. Specifically, x-axis coordinate data is collected from the foot position sensor. The difference between adjacent sampling points is calculated at a sampling rate of 100Hz to obtain the displacement increment. Kalman filtering is applied to smooth the curve to obtain the end-effector lateral displacement curve. The narrowing amplitude of the foot landing point distance is superimposed. The landing point distance is calculated from the coordinates of the foot landing points read by the hip joint encoder. The narrowing amplitude is the current distance minus the distance of the previous cycle. The superposition uses a weighted average fusion with a weight of 0.6 for the displacement curve and 0.4 for the narrowing amplitude. The resulting sequence is the original posture sampling sequence. Based on the original posture sampling sequence, the pressure readings of the plantar pressure array are called, and the lateral drift path of the pressure center on the tread is extracted according to the role switching of the swing leg and the supporting leg. The bias direction of the pressure distribution in each step cycle is calibrated by referring to the end lateral displacement curve. The trunk acceleration collected by the upper body inertial unit is separated into lateral channels. The peak frequency of the trunk lateral acceleration is extracted in a sliding time window by fast Fourier transform. The time window size is 1 second and the step size is 0.5 seconds, thus obtaining the oscillation frequency sequence caused by the load bias at the center of gravity. Align the oscillation frequency sequence with the synchronous control record by timestamp, fuse the lateral drift path of the pressure center, the narrowing of the landing point spacing, and the lateral displacement curve at the end, and reconstruct multi-channel sampling segments according to the single-step action of going up and down stairs. The specific fusion steps are as follows: first, align all sequences by timestamp, fill missing values ​​using linear interpolation, and then use the step cycle as the slicing standard, with each cycle starting at the moment the supporting leg lands, reconstructing into segments, and labeling the bias direction accompanying the upper body load distribution. Principal component analysis is used to extract principal features, and the weights are based on the variance contribution rate, such as path 0.3, narrowing amount 0.3, and displacement curve 0.4, to establish an initial attitude dataset covering phase features, cycle switching, pressure distribution, load bias, and lateral oscillation multi-dimensional channels.

[0018] When a humanoid robot walks on a narrow stair tread, the coupling between the pelvic torsional phase and the stride cycle directly affects the lateral distance between the two foot landing points.

[0019] In one embodiment, the data acquisition process is jointly undertaken by an inertial sensor located at the center of the pelvis, a rotary encoder at the hip joint, a pressure array on the sole of the foot, and an inertial unit at the torso position, which respectively correspond to the sensing channels for the torsional phase, step cycle, foot landing point, and upper body load.

[0020] Specifically, the inertial sensor is attached to the geometric center of the pelvis and outputs triaxial angular velocity and triaxial acceleration. The rotational component of the pelvis relative to the torso base about the vertical axis is extracted as the trajectory of the angular velocity change of the torsional phase. This trajectory presents a cyclic waveform of first positive acceleration, zero crossing, and reverse acceleration within a complete step cycle.

[0021] Preferably, the sampling frequency is 200 Hz to avoid phase aliasing. The hip joint encoder records the relative rotation angle of the left and right hip joints. A step cycle is defined with the moment the swing leg leaves the ground as the start of the cycle and the moment the same leg leaves the ground again as the end of the cycle. Timestamps are simultaneously added to the trajectory of angular velocity change and the mark of this cycle to form a synchronous comparison record of torsional phase and step cycle.

[0022] It should be noted that the plantar pressure array consists of pressure-sensitive units distributed in three zones: the forefoot, arch, and heel, with each zone outputting an independent pressure reading.

[0023] In one possible implementation, the pressure readings of both feet at the same instant are weighted and averaged to obtain the instantaneous lateral coordinates of the pressure center. When the robot switches from left-foot support to right-foot support, the pressure center transitions from the left-foot region to the right-foot region, and the transition trajectory is recorded as the lateral drift path of the pressure center on the tread surface. During the airborne phase, the lateral displacement curve of the swing leg is calculated by the ankle joint encoder and the lower leg inertial unit. The starting point of this curve corresponds to the instant of leg lifting, and the ending point corresponds to the instant of landing. The maximum convergence of the curve is compared with the offset of the drift path to determine the bias direction of the pressure distribution within each stride cycle, i.e., left-leaning, right-leaning, or centered.

[0024] Understandably, the upper body inertial unit is installed in the back of the chest cavity, outputting the acceleration of the torso in the lateral, longitudinal, and vertical directions. The acceleration signal of the lateral channel of the torso is isolated, and a sliding time window with a length equal to two step cycles and a step size equal to half a step cycle is constructed along the time axis. Within each time window, a Fast Fourier Transform is performed on the lateral acceleration to obtain an amplitude spectrum ranging from 0 Hz to 20 Hz. The frequency corresponding to the maximum amplitude is the peak frequency within that time window. The peak frequencies of adjacent time windows are concatenated in chronological order to form an oscillation frequency sequence caused by the load biased at the center of gravity. When the upper body load is biased to one side, the values ​​of the oscillation frequency sequence will drift to a higher frequency range. Furthermore, the oscillation frequency sequence is aligned with the synchronous comparison record by timestamp.

[0025] Specifically, the single-step motion from the swing leg leaving the ground to landing is used as the slice unit. Segments of the same time interval are extracted from four types of channels: synchronous control record, lateral drift path of the pressure center, narrowing of the landing point spacing, and lateral displacement curve at the end point, and then recombined into multi-channel sampling segments.

[0026] For example, in a single-step motion slice, the moment when the peak angular velocity of the torsional phase occurs, the moment when the pressure center transitions from the supporting leg to the swinging leg, the moment when the foot retracts laterally, and the peak frequency of the lateral oscillation of the torso can be observed simultaneously. The relative positions of these four moments on the time axis reflect the bias direction accompanying the upper body load distribution.

[0027] Preferably, each single-step slice is labeled with three categories: left-skewed, right-skewed, and centered, serving as discrete annotations for the bias direction. Thus, the five data streams—phase feature channel, cycle switching channel, pressure distribution channel, load bias channel, and lateral oscillation channel—are organized into a unified time coordinate system, forming the initial attitude dataset. This dataset can simultaneously reflect the coupling state of the pelvic torsional phase and the step cycle, as well as the lateral oscillation of the center of gravity caused by upper body load offset.

[0028] S102. The initial posture dataset is subjected to noise filtering and smoothing, the instability characteristics of the pelvic torsion phase and step cycle synchronization are identified and the instability state interval is defined. The change law of the tilt margin is analyzed within the instability state interval, and the filtered posture profile is determined based on the change law. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread.

[0029] For the phase feature channel, cycle switching channel, pressure distribution channel, load bias channel, and lateral oscillation channel in the initial attitude dataset, Kalman filtering is used to establish state transition equations and observation equations respectively. The pelvic torsional angular velocity and step cycle switching beats are set as state variables, and the inertial sensor output and pressure array partition pressure readings are set as observations. The state estimate and covariance matrix are iteratively updated beat by beat along the time axis to remove high-frequency jitter and sudden spikes, resulting in a smoothed multi-channel attitude timing sequence. Based on the smoothed multi-channel attitude timing sequence, the zero-crossing point when the pelvic torsional angular velocity changes from positive to negative is regarded as the torsional phase switching moment. The time deviation between the two is compared beat by beat along the time axis with the support leg switching moment of the step cycle switching. If the time deviation accumulates unidirectionally and exceeds a preset threshold within several consecutive step cycles, it is determined that the pelvic torsional phase and step cycle are synchronously unstable. The corresponding unstable period is marked, and the oscillation amplitude of that period is extracted from the lateral oscillation channel. Based on the oscillation amplitude of the unstable period and the oscillation amplitude of the adjacent non-unstable period, a time series curve of the tilt margin is constructed. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. The turning point of the tilt margin from a convergent trend of decreasing amplitude step by step to a divergent trend of increasing amplitude step by step is tracked along the time axis. The multi-channel attitude time series before and after the turning point is bound to the unstable period label to determine the filtered attitude profile.

[0030] When a humanoid robot walks on a narrow stair tread, its initial posture dataset comes from multiple samples from inertial sensors, hip joint encoders, foot pressure arrays, and upper body inertial units, resulting in jitter spikes and step peaks. In one implementation, the phase feature channel, cycle switching channel, pressure distribution channel, load bias channel, and lateral oscillation channel in the initial posture dataset are respectively fed into a Kalman filter for noise filtering and smoothing estimation.

[0031] Specifically, Kalman filtering consists of two parts: a state transition equation and an observation equation. The state variables are represented by two components: the pelvic torsional angular velocity ω and the step cycle switching beat τ. The state transition equation extrapolates the state variables of the current beat to the next beat under the assumption of uniform change, providing a prior estimate. The observation equation linearly maps the triaxial angular velocities output from the inertial sensor and the pressure readings of the pressure array in the forefoot, arch, and heel regions onto the state variables, providing an observation prediction. The residual between the prior estimate and the observation prediction is multiplied by the Kalman gain K and written back into the state variables to obtain the posterior estimate for the current beat.

[0032] Preferably, the process noise covariance Q is on the order of 0.01, and the observation noise covariance R is calibrated based on the measured variance of each sensor when it is stationary.

[0033] It should be noted that the covariance matrix P evolves alternately with each prediction and update cycle, and the value of K adaptively adjusts with P. When a sudden spike occurs in the pressure array, the value of K automatically decreases, making the posterior estimate more inclined to extrapolate the prior; when the sensor readings stabilize, the value of K increases, making the posterior estimate fit the observation. The posterior estimates of the five channels are concatenated along the time axis to form a smoothed multi-channel attitude timing sequence.

[0034] Specifically, the timing waveform of the pelvic torsional angular velocity ω is read from the phase feature channel of the smoothed multi-channel attitude timing, and the instant when ω crosses the zero line from a positive value to a negative value is marked as the torsional phase switching moment t1. Simultaneously, the instant when the supporting leg switches from left to right or from right to left is read from the cycle switching channel. Each step cycle generates a pair of t1 and t2 values.

[0035] In one embodiment, the time deviation curve is obtained by arranging the difference Δt between t1 and t2 within the same step cycle along the time axis. When the robot is normally going up and down stairs, Δt oscillates slightly around zero; when the upper body load is biased to one side, causing the pelvic torsion rhythm to decouple from the supporting leg switching rhythm, Δt will accumulate step by step in the same direction.

[0036] For example, if Δt cumulatively exceeds a preset threshold θ1 within three consecutive step cycles (θ1 is 40 milliseconds), it is determined that the pelvic torsion phase and the step cycle are synchronously unstable. The interval from the start of the first cumulative beat to the cumulative fall is marked as the unstable period, and the oscillation amplitude A of each beat within this period is read from the lateral oscillation channel. Further, the oscillation amplitude A of the unstable period and the oscillation amplitude A0 of the immediately preceding and following unstable periods are aligned by timestamps to construct a time series curve of the tilt margin M. The tilt margin M is the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. The larger the margin, the farther the robot is from the edge of the tread and the stronger its anti-tilt capability.

[0037] It is understandable that the first-order difference ΔM is taken for the time series curve of M along the time axis for each frame. If ΔM remains negative for several consecutive frames and its absolute value decreases for each frame, then M is in a converging trend with decreasing amplitude; if ΔM turns from negative to positive for several consecutive frames and its absolute value increases for each frame, then M changes from a converging trend to a diverging trend with increasing amplitude. The instantaneous position of the sign reversal is the turning point tc.

[0038] Preferably, five step cycles are taken before and after tc, and corresponding multi-channel attitude time segments are extracted and bound to the instability period label. Thus, the five smooth channels of phase feature, cycle switching, pressure distribution, load bias, and lateral oscillation, together with the instability period label and the turning point tc, form a filtered attitude profile, which can depict the occurrence time of synchronous instability of pelvic torsion phase and step cycle, as well as the evolution process of the tilt margin from convergence to divergence.

[0039] S103. Extract pelvic torsional phase features based on the filtered posture image, evaluate the lateral tilt margin divergence, generate a torsional phase adjustment scheme corresponding to the upper body load distribution, and calibrate the critical posture recognition threshold.

[0040] Based on the filtered attitude profile, the peak moment, peak amplitude, and zero-crossing interval of the pelvic torsional angular velocity within each step cycle are extracted from the phase feature channel and the cycle switching channel. Combined with the offset direction marked by the load offset channel and the lateral drift path of the pressure center in the pressure distribution channel, the pelvic torsional phase acquisition feature vector is obtained by splicing in each unstable period slice. The pelvic torsional phase acquisition feature vector includes the phase peak amplitude, phase zero-crossing interval, offset direction label, the cumulative time deviation between the torsional switching moment and the support switching moment, and the lateral offset of the drift path. Based on the pelvic torsion phase acquisition feature vector, the values ​​of the tilt margin time series curve after the turning point are extracted along the time axis. The second difference of the tilt margin is calculated step by step to characterize the divergent acceleration. The divergent acceleration and the tilt margin amplitude amplification rate are concatenated to form a divergent evolution index sequence. The divergent evolution index sequence is bound to the trunk lateral acceleration amplitude acquired by the lateral oscillation channel to obtain a sample set reflecting the real-time distribution of upper body load accompanied by the pelvic torsion phase adjustment scheme. Each sample carries phase acquisition features, divergent evolution index and load bias direction label. The sample set is divided into a training subset and a validation subset. A support vector machine (SVM) is used to train a classifier with the load bias orientation label as the class label. The SVM uses a Gaussian kernel function to map the phase acquisition features and the divergence evolution index to a high-dimensional space to solve for the maximum margin hyperplane, and outputs torsional phase adjustment schemes for three load distributions: left-biased, right-biased, and center-biased. The torsional phase adjustment schemes give the increase or decrease in pelvic torsion amplitude and the torsional initiation phase offset corresponding to each load distribution. The critical pose recognition threshold is determined by taking the boundary value of the decision function corresponding to the highest classification accuracy of the SVM on the validation subset.

[0041] When a humanoid robot carries an unbalanced load on a narrow stair tread, there is a strong coupling relationship between the acquired features of the pelvic torsional phase and the real-time load distribution on the upper body. In one embodiment, feature engineering is performed on the phase feature channel, cycle switching channel, pressure distribution channel, load bias channel, and lateral oscillation channel stored in the filtered posture profile to extract multi-dimensional vectors that can be used for classification.

[0042] Specifically, the continuous waveform of the pelvic torsional angular velocity ω within each step cycle is read from the phase characteristic channel, and its extreme points are scanned along the time axis. The instant when ω reaches its maximum positive value is recorded as the peak time tp, and the corresponding amplitude is recorded as the phase peak amplitude Ap. The time interval between two consecutive zero crossings of ω is recorded as the phase zero crossing interval Tz.

[0043] In one possible implementation, the instantaneous time ts of the support leg switching is read from the periodic switching channel, and the difference between ts and tp within the same step is accumulated along the time axis to obtain the cumulative time deviation Δs between the torsion switching time and the support switching time. The offset direction label L corresponding to the step is read from the load offset channel, and L is taken from three discrete values: left offset, right offset, and center.

[0044] It should be noted that the lateral drift path of the pressure center is read from the pressure distribution channel, and the maximum lateral displacement of the pressure center from the supporting leg region to the swing leg region within each stride is extracted and denoted as the lateral offset d of the drift path. The components are assembled in a fixed order to form a pelvic torsion phase feature vector x, denoted as... The vector length is 5.

[0045] Preferably, within each slice of the instability period, the x values ​​of adjacent steps are smoothed by mean, so that abnormal fluctuations in individual steps do not dominate the classification results. The feature vector x obtained from the pelvic torsion phase serves as the basic data for subsequent classifier input. Further, the extraction of divergent evolution indices relies on the already labeled turning point tc in the filtered posture profile. The roll margin M is read frame by frame after tc, and a first-order difference of M is constructed along the time axis. Then construct the second-order difference Let d2M be the divergent acceleration, which reflects the inflection point strength of the roll margin decay rate.

[0046] Understandably, the ratio of the maximum amplitude of M within N consecutive steps after tc to the value of M at tc is denoted as the roll margin amplitude amplification rate r, where N ranges from 5 to 8. The divergence acceleration d²M and the amplitude amplification rate r are then concatenated along the time axis to form a divergence evolution index sequence. .

[0047] Specifically, the divergent evolution index sequence e, the pelvic torsional phase acquisition feature vector x, and the trunk lateral acceleration amplitude Ah acquired by the lateral oscillation channel are bound by the same timestamp to form a sample Si=x,e,Ah,L, where L serves as the sample category label. Several Si samples acquired over multiple steps are aggregated into a sample set, reflecting the correspondence between the real-time distribution of upper body load and the pelvic torsional phase adjustment scheme.

[0048] In one embodiment, the sample set is divided into a training subset and a validation subset in a 7:3 ratio, and a support vector machine (SVM) is used for multi-class classification training. The core idea of ​​the SVM is to project the samples into a high-dimensional space and then solve for a hyperplane that maximizes the distance between the two classes of samples.

[0049] Specifically, the support vector machine uses a Gaussian kernel function. The input space spanned by the phase acquisition feature vector and the divergent evolution index is mapped to an infinite-dimensional feature space. γ is the kernel width parameter, which controls the sample similarity decay rate. ‖x1-x2‖ represents the Euclidean distance between two sample vectors.

[0050] Preferably, γ takes a value from 0.5 to 2.0, and the penalty coefficient C takes a value from 1 to 10. The maximum margin hyperplane is solved on the training subset to obtain decision functions for the three types of load distributions: left-skewed, right-skewed, and center-skewed. Furthermore, from Read the support vectors and their weights corresponding to each type of load distribution, and deduce the torsional phase adjustment scheme for that type of load. The torsional phase adjustment scheme provides the increase / decrease in pelvic torsion amplitude Δθ and the torsional initiation phase offset Δφ corresponding to each type of load distribution. For example, for left-biased loads, Δθ is positive (3 to 5 degrees) and Δφ is advanced (2 to 4 degrees), while for right-biased loads, Δθ is negative. The validation subsets are then fed into... Perform category decision-making, statistically analyze the change curve of classification accuracy as a function boundary value takes, and record the value corresponding to the highest classification accuracy. The value is denoted as the critical attitude recognition threshold θc, which serves as the criterion for determining whether the robot has entered a critical state of tilt instability.

[0051] In another approach, a one-to-one multi-class strategy is adopted to construct three pairs of binary classifiers: left-biased to right-biased, left-biased to center, and right-biased to center. The decision function f(x) of each pair of classifiers has a boundary of 0. The value of f(x) is calculated for the validation set samples, and the average absolute deviation of the three pairs of boundaries is used as the threshold θc. For example, when the deviations are 0.5, 0.6, and 0.7, θc = 0.6, which is used to identify the critical pose.

[0052] S104. The consistency of foot trajectory data is verified by the critical posture recognition threshold. Based on the trajectory deviation obtained from the verification, a pelvic torsion phase control sequence is generated iteratively. The trend of lateral tilt margin divergence corresponding to the upper body load distribution is analyzed. The stride cycle synchronization scheme is determined by combining the pelvic torsion phase control sequence and the trend of lateral tilt margin divergence.

[0053] The critical posture recognition threshold and the maximum convergence of the foot lateral displacement curve in each step cycle are compared step by step along the timestamp. The percentage of the maximum convergence of the foot synchronously enters the preset narrowing interval in the instantaneous interval when the value of the decision function exceeds the critical posture recognition threshold. If the percentage exceeds the preset verification threshold, the consistency verification conclusion is determined to be consistent; otherwise, it is determined to be a deviation and the difference step cycle is marked. The phase acquisition feature vector and the lateral offset of the drift path are extracted from the posture image slice corresponding to the difference step cycle to obtain the consistency verification conclusion and the deviation slice record. A correlation index is constructed using the lateral offset of the drift path and the narrowing of the landing point distance in the deviation slice record. The frequency sequence of the lateral oscillation of the center of gravity is concatenated along the timestamp to obtain the correlation trend curve. With the goal of reducing the amplitude of the correlation trend curve, the torsional initiation phase offset and the increase or decrease of the torsional amplitude in the phase acquisition feature vector are iteratively adjusted step by step. After each iteration, the critical posture recognition threshold is brought back to re-verify the proportion until the amplitude of the correlation trend curve falls into the preset convergence range and the foot lateral displacement curve converges synchronously, thus obtaining the pelvic torsion phase control sequence. The waveform of the phase feature channel in the posture image is reassembled according to the pelvic torsion phase control sequence. The tilt margin time sequence corresponding to the upper body load distribution is extracted along the time axis. The second-order difference evolution of the tilt margin is calculated step by step to obtain the divergence acceleration growth rate. If the divergence acceleration growth rate exceeds the preset aggravation threshold within a certain number of consecutive step cycles, it is determined that the upper body load distribution is accompanied by aggravation of tilt margin divergence. The divergence aggravation period is mapped to the step cycle switching beat. The pelvic torsion phase control sequence and the step cycle switching beat are aligned according to the timestamp to determine the step cycle synchronization scheme.

[0054] When a humanoid robot carries an unbalanced load on a narrow stair tread, the critical attitude recognition threshold θc obtained in the preceding steps is only a criterion for judging tilt instability. It still needs to be cross-verified with the actual foot retraction action and a closed-loop iteration should be performed on the coupling law between the landing point distance and the center of gravity oscillation. In one implementation, the overall closed loop includes three process nodes: consistency verification, correlation iteration, and step cycle alignment.

[0055] Specifically, the values ​​of the support vector machine decision function are... Compare frame-by-frame with θc along the time axis and mark the results. The continuous beat interval that continuously exceeds θc is recorded as the over-limit interval. Simultaneously, the maximum contraction amount u within each step cycle is read from the lateral displacement curve of the foot, and the range of values ​​where u is less than the lateral half-width of the tread minus the preset safety margin is recorded as the preset narrowing interval.

[0056] It should be noted that the preset narrowing range is a dual constraint on both the geometric dimensions of the stair tread and the robot's foot support surface. For example, if the lateral half-width of the tread is 180 mm and the safety margin is 40 mm, the preset narrowing range is... Millimeters. Within each out-of-limit interval, check whether u falls within the preset narrowing interval for each frame. Divide the number of frames that simultaneously meet the requirement by the total number of frames in the out-of-limit interval to obtain the consistency ratio p.

[0057] Preferably, the preset verification threshold is 0.85. If the result is consistent, the consistency verification conclusion is determined to be consistent; otherwise, it is determined to be a deviation, and the corresponding step cycle is marked as a difference step cycle.

[0058] Understandably, a match indicates that the determination of the pelvic torsion phase aligns with the actual foot retraction movement, while a deviation indicates a disconnect between the two rhythms. Further, the phase feature vector x and the lateral offset d of the pressure center's lateral drift path are read from the posture profile slice corresponding to the differential step cycle, and together with the narrowing of the landing point spacing ds within that slice, are aggregated into a deviation slice record R. R is used throughout subsequent iterative closed-loop iterations.

[0059] In one embodiment, the correlation index is constructed with R as input.

[0060] Specifically, the correlation index is obtained by multiplying d and ds for each time step within R by the same timestamp and taking a moving average. c reflects the degree to which the drift amplitude and the narrowing amplitude of the landing point are amplified in the same direction.

[0061] Specifically, the frequency sequence h of the horizontal oscillation of the center of gravity is read synchronously from R, and c and h are concatenated along the timestamp to form a related trend curve. α is a weighting coefficient, ranging from 0.5 to 1.5 to balance the magnitudes of the two. When the upper body load is severely offset, d and ds expand in the same direction and h drifts towards higher frequencies, pushing the amplitude of g(t) upwards in the same direction. Iterative adjustments are made to reduce the peak amplitude of g(t). The torsional initiation phase offset Δφ and the torsional amplitude increase / decrease Δθ in x are used as adjustable quantities, and the values ​​are tested step by step with a small step size of ±0.5 degrees along their value axis. After each iteration, the phase characteristic channel waveform is reassembled with new Δφ and Δθ, and f(x,e) is solved again, p is statistically analyzed, and g(t) is refreshed.

[0062] Preferably, the iteration termination condition is that the peak amplitude of g(t) is less than 0.1 of the preset convergence interval and Simultaneously, the Δφ and Δθ adjustments accumulated over the differential step cycle are connected in chronological order to form the pelvic torsion phase control sequence S, which runs through the entire process of going up and down stairs.

[0063] In one possible implementation, the waveform of S is reassembled into the phase feature channel of the attitude profile and fed into the roll margin calculation path along with the original cycle switching, pressure distribution, load bias, and lateral oscillation channels to obtain the adjusted roll margin timing M'. The first-order difference dM' and the second-order difference d2M' are taken along the time axis of M' to calculate the divergence acceleration growth rate β = d2M'(k) / d2M'(k-1). If β exceeds the preset aggravation threshold of 1.2 for three consecutive step cycles, it is determined that the upper body load distribution is accompanied by an aggravation of the roll margin divergence, and this period is recorded as the divergence aggravation period.

[0064] Specifically, the start and end beats of the divergence intensification period are mapped to the supporting leg switching beats of the cycle switching channel, and the two closest switching beats tA and tB before and after it are identified. The torsional peak moment in S is shifted by timestamp to strictly align it with the midpoint of tA and tB; this alignment process is called same-to-same alignment. Thus, the pelvic torsion phase control sequence and the stepping cycle switching beat maintain a stable phase relationship on the time axis, forming a stepping cycle synchronization scheme, ensuring that the pelvic torsion movement and the swing leg stepping rhythm tighten synchronously in each divergence intensification period.

[0065] S105. Generate a limb coordination and stability control output sequence according to the step cycle synchronization scheme.

[0066] The peak torsion moment, support switching beat, and corresponding torsion amplitude increase / decrease and torsion start phase offset in the stride cycle synchronization scheme are expanded beat by beat along the timestamp. Based on the upper limit of the lateral distance between the two feet in the given landing point reduction requirement of the half-width of the stair tread and the half-width of the robot's foot support, the upper limit is superimposed to tighten the constraint on the lateral coordinate of the swing leg's foot landing point. The pelvic torsion angle command, hip joint swing angle command, and swing leg's foot landing point lateral coordinate command are calculated beat by beat to obtain the limb coordination and stability control output sequence. The pelvic torsion angle command, hip joint swing angle command, and swing leg's foot landing point lateral coordinate command in the limb coordination and stability control output sequence are distributed to the trunk waist, hip joint, and ankle joint links. The inverse kinematics of each joint is used to convert them into the target position, target velocity, and target torque of the actuator, which are synchronously sent to the corresponding joint actuator along the timestamp. The actuator tracks the target position, target velocity, and target torque in a closed loop to drive the humanoid robot body to perform the up and down stairs movement.

[0067] When a humanoid robot carries an unbalanced load on a narrow stair tread, the stride cycle synchronization scheme determines the phase relationship between the pelvic twisting motion and the swing leg's stepping rhythm, while the requirement to reduce the landing point distance constrains the lateral landing range of the feet on the tread. In one implementation, the overall control process includes two stages: motion command calculation and joint actuation, which are sequentially connected along the timestamp.

[0068] Specifically, the torsion peak time tp, support switching beat ts, torsion amplitude increase / decrease Δθ and torsion start phase offset Δφ within each step cycle are read from the step cycle synchronization scheme and expanded into a continuous phase command time base along the timestamp, with each beat corresponding to a 10-millisecond cycle.

[0069] Preferably, the total length of the phase command time base covers the complete step sequence of a single stair-climbing action.

[0070] It should be noted that the requirement to reduce the landing distance stems from the margin constraint between the geometric dimensions of the stair tread and the robot's foot support surface. Let W be the horizontal half-width of the stair tread, F be the horizontal half-width of the robot's foot support, and s be the horizontal safety margin. The upper limit of the horizontal distance between the two feet... .

[0071] In one embodiment, when W is 180 mm, F is 55 mm, and s is 15 mm, D is 220 mm; when the stair tread is narrower and W is 140 mm, D is reduced to 140 mm accordingly.

[0072] It is understood that the upper limit D adapts adaptively to the tread size and never exceeds the maximum possible distance between the two feet simultaneously resting on the tread. The upper limit D tightens the constraint on the lateral coordinates of the landing point of the swing leg, so that the range of the landing point coordinates for each beat is limited to a lateral band of width D with the torso centerline as the axis of symmetry.

[0073] Specifically, the beat-by-beat instruction calculation is executed in parallel along three branches. The pelvic torsion angle instruction θ(k) is generated by superimposing Δθ and Δφ in the phase instruction time base, θ(k) = θ0 + Δθ·sin(2π·(k-Δφ) / T), where θ0 is the initial torsion angle, T is the step cycle length, and k is the current beat number. The hip joint swing angle instruction φh(k) is derived according to the phase relationship of the support switching beat ts, linearly transitioning from the swing start phase to the swing end phase within each step. The lateral coordinate instruction yf(k) of the swing leg foot landing point is taken in a direction close to the midline of the torso under the constraint of the upper limit D, and recorded as... yf0(k) represents the initial lateral coordinate of the landing point given by kinematic preset, and L represents the load bias direction label. θ(k), φh(k), and yf(k) are concatenated along timestamps to form a limb coordination stabilization control output sequence. Further, the limb coordination stabilization control output sequence is distributed to three links according to joint affiliation: θ(k) is sent to the trunk / waist link, φh(k) to the hip joint link, and yf(k) to the ankle joint link. Each link corresponds to a set of actuators, and each link maintains its own feedback loop.

[0074] In one possible implementation, each link performs inverse kinematics calculations based on the link length, joint axis direction, and initial posture of each joint. For example, the ankle link inversely derives yf(k) into a combination of ankle pitch and roll angles, while the hip link inversely derives φh(k) into a combination of hip pitch, abduction, and internal rotation angles.

[0075] Preferably, the inverse kinematics solution employs a combination of analytical solutions and numerical iterations, switching to numerical iterations when singular postures occur to avoid solution failure. The solved joint angles, along with their first derivatives and joint torque requirements, are converted into the target position pi, target velocity vi, and target torque τi of the actuator, where i is the joint number.

[0076] Specifically, pi, vi, and τi are synchronously sent to the corresponding joint actuators along the timestamp, with the sending cycle consistent with the 10-millisecond beat of the phase command time base. Each actuator has a built-in three-layer closed loop: a position loop, a velocity loop, and a torque loop, adjusting the motor current according to the feedback of position deviation, velocity deviation, and torque deviation. As a result, the robot's torso, waist, hip joints, and ankle joints coordinate their movements according to the same timestamp, the pelvic twisting rhythm and the swing leg stepping rhythm remain synchronously tightened in each step cycle, and the landing points of both feet always fall within the upper limit D, enabling stable up and down stairs on narrow stair treads.

[0077] This invention also provides a limb coordination control device for a humanoid robot going up and down stairs, mainly comprising: The first module is used to acquire the synchronization data of the robot's pelvic torsion phase and step cycle, and combine it with foot trajectory and pressure distribution data to form an initial posture dataset. The second module is used to perform noise filtering and smoothing on the initial posture dataset, identify the instability characteristics of the pelvic torsion phase and the step cycle synchronization and delineate the instability state interval, analyze the change law of the tilt margin in the instability state interval, and determine the filtered posture profile based on the change law. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. The third module is used to extract pelvic torsional phase features based on the filtered posture profile, evaluate the lateral tilt margin divergence, generate a torsional phase adjustment scheme corresponding to the upper body load distribution, and calibrate the critical posture recognition threshold. The fourth module is used to verify the consistency of foot trajectory data through the critical posture recognition threshold, iteratively generate a pelvic torsion phase control sequence based on the trajectory deviation obtained from the verification, analyze the trend of lateral tilt margin divergence aggravation corresponding to the upper body load distribution, and determine the step cycle synchronization scheme by combining the pelvic torsion phase control sequence and the trend of lateral tilt margin divergence aggravation. The fifth module is used to generate a limb coordination and stability control output sequence based on the step-cycle synchronization scheme.

[0078] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for coordinating limb control when a humanoid robot goes up and down stairs, characterized in that, include: Acquire synchronization data of the robot's pelvic torsion phase and step cycle, and combine them with foot trajectory and pressure distribution data to form an initial posture dataset; The initial posture dataset is subjected to noise filtering and smoothing to identify the instability characteristics of pelvic torsion phase and step cycle synchronization and to define the instability state interval; the variation law of tilt margin is analyzed within the instability state interval, and the filtered posture profile is determined based on the variation law; wherein, the tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. Based on the filtered posture profile, pelvic torsional phase features are extracted, the lateral tilt margin divergence is evaluated, a torsional phase adjustment scheme corresponding to the upper body load distribution is generated, and the critical posture recognition threshold is calibrated. The consistency of foot trajectory data is verified by the critical posture recognition threshold. Based on the trajectory deviation obtained from the verification, a pelvic torsion phase control sequence is iteratively generated. The trend of lateral tilt margin divergence corresponding to the upper body load distribution is analyzed. The pelvic torsion phase control sequence and the trend of lateral tilt margin divergence are combined to determine the step cycle synchronization scheme. The limb coordination and stability control output sequence is generated based on the step-cycle synchronization scheme.

2. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 1, characterized in that, The process of acquiring synchronization data of the robot's pelvic torsion phase and step cycle, combined with foot trajectory and pressure distribution data to form an initial posture dataset, includes: The angular velocity change trajectory of the torsional phase is obtained by an inertial sensor in the pelvis, and the gait switching beat within the step cycle is read by a hip joint encoder. The angular velocity change trajectory and the gait switching beat are aligned to form the synchronization data of the robot's pelvic torsional phase and step cycle. The lateral displacement curve of the swing leg foot retraction motion is extracted, and the narrowing of the distance between the two foot landing points is superimposed to obtain the original posture sampling sequence. Based on the original posture sampling sequence, the partition pressure readings of the plantar pressure array are called, and the lateral drift path of the pressure center on the tread is extracted according to the role switching of the swing leg and the supporting leg. The bias direction of the pressure distribution in each stride cycle is calibrated, the trunk acceleration is separated according to the lateral channel, and the peak frequency of the trunk lateral acceleration is extracted in the sliding time window through fast Fourier transform to obtain the oscillation frequency sequence. After aligning the oscillation frequency sequence with the synchronization data of the robot's pelvic torsion phase and step cycle, the initial posture dataset is established by fusing the lateral drift path and lateral displacement curve of the pressure center on the tread surface.

3. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 2, characterized in that, The initial posture dataset is subjected to noise filtering and smoothing to identify the instability characteristics of pelvic torsion phase and step cycle synchronization and to define the instability state interval. Analyze the variation pattern of roll margin within the aforementioned instability state range, and determine the filtered attitude profile based on this variation pattern, including: For the multi-channel data of the initial attitude dataset, the Kalman filter algorithm is used to smooth the multi-channel data to form a smoothed multi-channel attitude timing sequence. The time when the pelvic torsional angular velocity crosses zero is compared with the time when the supporting leg switches during the stepping cycle. The time deviation is calculated. If the unidirectional accumulation of the time deviation exceeds a preset threshold, the period of instability is marked. Obtain the oscillation amplitude during the instability period, construct the tilt margin time series curve, track the turning point when the tilt margin changes from a converging trend of amplitude decaying step by step to a diverging trend of amplitude amplifying step by step, bind the multi-channel attitude time series before and after the turning point with the instability period label, and determine the filtered attitude profile.

4. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 2, characterized in that, The step of extracting pelvic torsional phase features from the filtered posture profile and assessing the tilt margin divergence includes: The peak time, peak amplitude and zero-crossing interval of pelvic torsional angular velocity are extracted from the filtered posture image. Combined with the load bias direction and the lateral drift path of the pressure center, the pelvic torsional phase features are spliced ​​together. Based on the pelvic torsion phase characteristics, the values ​​of the tilt margin time series curve after the turning point are extracted along the time axis. The second difference of the tilt margin is calculated step by step to characterize the divergent acceleration. The divergent acceleration and the tilt margin amplitude amplification rate are concatenated to form a divergent evolution index sequence to generate a sample set. The input sample set is fed into the support vector machine classifier, which outputs a torsional phase adjustment scheme for different load distributions to determine the critical attitude recognition threshold.

5. The method for coordinated limb control of a humanoid robot going up and down stairs as described in claim 4, characterized in that, The input sample set is fed into a support vector machine classifier, and the output is a torsional phase adjustment scheme for different load distributions. The critical pose recognition threshold is determined by: dividing the sample set into a training subset and a validation subset; using a support vector machine to train the classifier with the load bias direction label as the class label; outputting a torsional phase adjustment scheme for three load distributions: left-biased, right-biased, and center-centered; and taking the boundary value of the decision function corresponding to the highest classification accuracy of the support vector machine on the validation subset to determine the critical pose recognition threshold.

6. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 2, characterized in that, The analysis of the upper body load distribution corresponds to an exacerbation of the tilt margin divergence. Combined with the pelvic torsion phase control sequence and the exacerbation of the tilt margin divergence, a step-cycle synchronization scheme is determined, including: The associated trend curve is obtained by concatenating the oscillation frequency sequence along the timestamp. The phase is iteratively adjusted step by step with the goal of reducing the amplitude of the associated trend curve. After each iteration, the critical posture recognition threshold is brought back to re-verify the proportion of the maximum foot contraction that synchronously enters the preset narrowing interval, until the amplitude of the associated trend curve falls into the preset convergence interval and the foot lateral displacement curve contraction trend is synchronously matched, thus obtaining the pelvic torsion phase control sequence.

7. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 1, characterized in that, The method for determining the step-cycle synchronization scheme includes: The waveform of the phase characteristic channel is reassembled according to the pelvic torsion phase control sequence. The tilt margin timing corresponding to the upper body load distribution is extracted along the time axis, the tilt margin is analyzed, and the step cycle synchronization scheme is generated.

8. The method for coordinating limb control of a humanoid robot going up and down stairs as described in claim 1, characterized in that, After generating the limb coordination stability control output sequence according to the step-cycle synchronization scheme, the process includes: The pelvic torsion angle command, hip joint swing angle command, and lateral coordinate command of the foot landing point of the swinging leg in the limb coordination and stability control output sequence are distributed to the corresponding joint links. The target position, target velocity, and target torque of the actuator are calculated by inverse kinematics. The data is synchronously sent to the corresponding joint actuators along the timestamp to track the target position, target velocity, and target torque, thereby driving the robot body to complete the up and down stairs movement.

9. A humanoid robot's limb coordination control device for going up and down stairs, characterized in that, The device includes: The first module is used to acquire the synchronization data of the robot's pelvic torsion phase and step cycle, and combine it with foot trajectory and pressure distribution data to form an initial posture dataset. The second module is used to perform noise filtering and smoothing on the initial posture dataset, identify the instability characteristics of the pelvic torsion phase and the step cycle synchronization and delineate the instability state interval, analyze the change law of the tilt margin in the instability state interval, and determine the filtered posture profile based on the change law. The tilt margin is taken as the distance margin between the lateral drift path of the pressure center and the edge of the stair tread. The third module is used to extract pelvic torsional phase features based on the filtered posture profile, evaluate the lateral tilt margin divergence, generate a torsional phase adjustment scheme corresponding to the upper body load distribution, and calibrate the critical posture recognition threshold. The fourth module is used to verify the consistency of foot trajectory data through the critical posture recognition threshold, iteratively generate a pelvic torsion phase control sequence based on the trajectory deviation obtained from the verification, analyze the trend of lateral tilt margin divergence aggravation corresponding to the upper body load distribution, and determine the step cycle synchronization scheme by combining the pelvic torsion phase control sequence and the trend of lateral tilt margin divergence aggravation. The fifth module is used to generate a limb coordination and stability control output sequence based on the step-cycle synchronization scheme.

10. The humanoid robot stair-climbing coordination control device as described in claim 9, characterized in that, After generating the limb coordination stability control output sequence according to the step-cycle synchronization scheme, the process includes: The pelvic torsion angle command, hip joint swing angle command, and lateral coordinate command of the foot landing point of the swinging leg in the limb coordination and stability control output sequence are distributed to the corresponding joint links. The target position, target velocity, and target torque of the actuator are calculated by inverse kinematics. The data is synchronously sent to the corresponding joint actuators along the timestamp to track the target position, target velocity, and target torque, thereby driving the robot body to complete the up and down stairs movement.

Images