Humanoid robot ground adaptive gait method and system based on foot end perception

Abstract

The application provides a kind of based on foot end perception humanoid robot ground adaptability gait method and system, it is related to robot technical field, this method includes: when robot foot end is grounded and meets preset trigger condition, by excitation device, ground is applied mechanical excitation, and the vibration response signal of ground is collected;Vibration response signal is carried out spectrum analysis, and ground state type is identified based on feature vector and the ground safety factor is calculated;According to ground safety factor, the boundary of allowable domain and the height of mass center are revised;The ground safety factor of each step of robot walking is marked to local ground safety map, and the next landing point is selected based on cost function, and gait parameter is adjusted on line based on ground safety factor;According to revised ZMP allowable domain, mass center height and gait parameter, determine the whole body joint torque, drive robot to execute step action.This scheme can make robot adapt to ground state and execute walking strategy, improve the walking safety of complex terrain.

Description

Technical Field

[0001] This invention relates to the field of robotics, and in particular to a ground-based adaptive gait method and system for humanoid robots based on foot perception. Background Technology

[0002] Although humanoid robots have been gradually deployed on a large scale in fields such as home service, industrial inspection, public security, and special operations, their ability to maintain autonomous balance and walk in complex terrain remains a core bottleneck restricting their transition from the laboratory to practical applications. Regarding the autonomous balance and walking capabilities of humanoid robots, current technologies still face the following challenges:

[0003] (1) There are blind spots in advance planning based on vision: Existing solutions mostly use vision or lidar to scan the terrain in advance and plan the landing point. However, in environments with dust, smoke, water reflection or insufficient light, vision sensors fail. Even if vision is normal, the ground that looks flat and solid may have holes inside (such as thin ice, hollow bricks, rotten wood planks), and vision cannot penetrate the surface to judge its load-bearing capacity.

[0004] (2) Inertial balance control lags behind actual instability: Existing balance control relies on IMU (Inertial Measurement Unit) to measure the body's attitude angles and angular velocities, and combines this with joint encoders to calculate the ZMP (Zero Moment Point) position. However, ZMP reflects the dynamic state that has "already occurred." When the ZMP is detected to deviate from the support domain, the robot is often already on the verge of instability. Especially on soft ground, the IMU cannot detect the slow sinking of the feet until the ground suddenly collapses and the attitude angle changes drastically, but by then it is too late.

[0005] (3) Lack of active perception capability of foot-ground interaction: Traditional humanoid robots only have six-dimensional force / torque sensors installed on the soles of their feet to measure ground reaction force. This is a kind of "passive force" perception, which can only know "how much support force the ground provides", but cannot know "why the ground cannot provide more force", that is, it cannot determine whether the ground bearing capacity is insufficient due to soft material or hollow cavity underneath.

[0006] Therefore, in existing control schemes for the autonomous balance and walking ability of humanoid robots, it is difficult for the robot to adapt to different ground conditions. That is, the lack or insufficient consideration of ground conditions makes it difficult for the robot to respond in a timely manner to different ground conditions, which in turn leads to the robot frequently stepping into gaps, falling down, etc., and the robot's walking safety in complex terrain is poor. Summary of the Invention

[0007] In view of this, and to address the above shortcomings, it is necessary to provide a ground-adaptive gait method and system for humanoid robots based on foot perception, so that the robot can adapt to the ground conditions to execute walking strategies and improve walking safety in complex terrain.

[0008] In a first aspect, the present invention provides a ground-based adaptive gait method for humanoid robots based on foot perception, comprising:

[0009] When the robot's foot touches the ground and the preset triggering conditions are met, the robot applies active mechanical excitation to the ground through the excitation device integrated in the foot and collects the vibration response signal of the ground through the accelerometer.

[0010] The vibration response signal is subjected to spectral analysis, and the foundation condition type is identified and the foundation safety factor is calculated based on the feature vector extracted from the spectral analysis.

[0011] The allowable boundary of the zero moment point ZMP and the height of the center of mass of the inverted pendulum model LIPM are corrected based on the foundation safety factor.

[0012] During the robot's walking process, the foundation safety factor of each step is marked on the local foundation safety map, and the next foothold is selected based on the cost function containing the local foundation safety map. The gait parameters are also adjusted online based on the foundation safety factor of the current supporting foot.

[0013] Based on the corrected ZMP allowable domain, center of mass height, and gait parameters, the whole-body joint torques are determined to drive the robot to perform stepping movements.

[0014] Preferably, the preset triggering condition is a perception-triggered gating mechanism, which must simultaneously satisfy:

[0015] The flexible tactile array on the sole of the foot detected that the contact area exceeded a first preset threshold of the sole area;

[0016] The ankle joint six-dimensional force sensor detected a vertical reaction force greater than the second preset threshold.

[0017] The angular velocities of the foot in all three axes are less than the third preset threshold.

[0018] Preferably, the active mechanical stimulation applied to the ground via the foot-integrated stimulation device includes:

[0019] The excitation force applied by the excitation device is expressed as:

[0020]

[0021] in, The excitation force applied to the excitation device, For the amplitude of the excitation force, The frequency sweep at time t;

[0022] The excitation force amplitude Based on the current positive pressure of the foot Dynamic adjustment is represented as follows:

[0023]

[0024] in, For maximum excitation force, For reference positive pressure.

[0025] Preferably, the step of performing spectral analysis on the vibration response signal and identifying the foundation condition type and calculating the foundation safety factor based on the feature vector extracted from the spectral analysis includes:

[0026] The vibration response signal is subjected to a fast Fourier transform to obtain a frequency domain representation, and the power spectral density is calculated based on this frequency domain representation.

[0027] Extract frequency domain feature vectors from the power spectral density; wherein, the frequency domain feature vectors include peak frequency, stiffness estimate, damping ratio estimate, and low-frequency energy proportion;

[0028] After the excitation stops, the residual vibration signal is collected by an accelerometer, the amplitude envelope of the residual vibration signal is extracted by Hilbert transform, and the decay time constant is calculated as a time domain feature.

[0029] The frequency domain feature vector and the time domain feature vector are used as inputs to a pre-trained classifier, which outputs the ground type.

[0030] The foundation safety factor is calculated based on the stiffness estimate and the time-domain characteristics.

[0031] Preferably, the power spectral density is calculated as follows:

[0032]

[0033] in, For power spectral density, Let T be the frequency domain representation, and T be the sampling time window. The sampling frequency;

[0034] The extraction methods for the frequency domain feature vector include:

[0035]

[0036] in, Peak frequency, This is an estimate of the stiffness. The equivalent mass of the foot participating in vibration, and These are the half-power point frequencies, The proportion of low-frequency energy;

[0037] The attenuation time constant satisfies ;in, The amplitude envelope of the residual oscillation signal. The decay time constant, It is a natural constant;

[0038] The calculation of the foundation safety factor based on the stiffness estimate and the time-domain characteristics includes determining it in the following manner:

[0039]

[0040] Where S is the foundation safety factor. Minimum stiffness required for safe standing This is the minimum stiffness value. The preset stiffness threshold, The preset decay time threshold, This is the preset threshold for the proportion of low-frequency energy.

[0041] Preferably, the step of correcting the allowable boundary of the zero moment point ZMP based on the foundation safety factor includes:

[0042] Calculate boundary shrinkage ;in, The preset maximum shrinkage amount;

[0043] For the boundary shrinkage amount First-order low-pass filtering is performed for smoothing to obtain the smoothed boundary shrinkage. ;in, , To control the cycle, The smoothing time constant;

[0044] The allowable domain of the zero-moment point ZMP is shrunk from the original supporting polygon, and the shrunk domain is represented as:

[0045] in, The boundary of the allowable domain of the corrected zero moment point ZMP. This is the smoothed boundary shrinkage amount. and Let x be the lower and upper limits in the x-direction. and These are the lower and upper limits in the y-direction.

[0046] Preferably, the step of correcting the height of the center of mass of the inverted pendulum model LIPM based on the foundation safety factor includes:

[0047] The correction term for the centroid height calculated based on the foundation safety factor is expressed as follows:

[0048]

[0049] in, To correct the height of the center of mass, Standard center of gravity height For the maximum settlement, For the total mass of the robot, It is the acceleration due to gravity. Minimum stiffness value, This refers to the foundation settlement.

[0050] The corrected centroid height Substituting the dynamic equations of the inverted pendulum model LIPM This is to make the center of mass trajectory adapt to the actual stiffness of the ground.

[0051] Preferably, the foundation safety map is represented as follows:

[0052]

[0053] in, This is a map showing the probability of foundation safety determined at time t. The foundation safety factor is determined at time t;

[0054] The cost function is expressed as follows:

[0055]

[0056] in, This represents the total cost function value. The weighting coefficients represent the distance cost. Indicates the coordinates of the candidate landing point. Indicates the coordinates of the expected landing point. The weighting coefficients representing the security costs, These are the weighting coefficients for the location region cost. This indicates the punitive value of unexplored areas;

[0057] The online adjustment of gait parameters based on the current foundation safety factor of the supporting foot includes:

[0058]

[0059] Where S is the foundation safety factor.

[0060] Preferably, when S=0, the method further includes a three-level emergency response; the three-level emergency response includes:

[0061] Level 1: Stop stepping with the ankle joint, contract the knee joint to lift the foot, and tilt the torso to the opposite side at a preset angle;

[0062] Level 2: Contralateral ankle joint torque compensation, maintaining balance through a hybrid control of ankle and hip strategies; the control formula is: , This indicates the ankle joint compensating torque. Represents the proportional gain coefficient. Indicates the desired joint angle. Indicates the actual joint angle. Indicates the gain coefficient of micronized powder. This is the compensation torque after the side that was stepped on fails;

[0063] Level 3: Mark the point of misstep as a permanent danger zone, replan your landing point and detour.

[0064] Secondly, the present invention provides a humanoid robot ground-based adaptive gait system based on foot perception, comprising: a foot perception module, a ground recognition module, a correction module, a gait planning module, and an execution control module;

[0065] The foot-end sensing module is used to apply active mechanical excitation to the ground through an excitation device integrated in the sole of the foot when the robot's foot touches the ground and a preset trigger condition is met, and to collect the vibration response signal of the ground through an accelerometer.

[0066] The foundation identification module is used to perform spectral analysis on the vibration response signal, and identify the foundation condition type and calculate the foundation safety factor based on the feature vector extracted from the spectral analysis.

[0067] The correction module is used to correct the allowable domain boundary of the zero moment point ZMP and the centroid height of the inverted pendulum model LIPM according to the foundation safety factor.

[0068] The gait planning module is used to mark the foundation safety factor of each step on the local foundation safety map during the robot's walking process, select the next footing point based on the cost function containing the local foundation safety map, and adjust the gait parameters online based on the foundation safety factor of the current supporting foot.

[0069] The execution control module is used to calculate the whole-body joint torque based on the corrected ZMP allowable domain, center of mass height and gait parameters, and drive the robot to perform stepping movements.

[0070] As can be seen from the above technical solution, the humanoid robot ground-adaptive gait scheme based on foot perception provided in this solution firstly applies active mechanical excitation to the ground through an excitation device integrated in the foot when the robot's foot touches the ground and meets the preset triggering conditions, and collects the ground vibration response signal through an accelerometer, then performs spectral analysis on the vibration response signal, and identifies the ground condition type and calculates the ground safety factor based on the feature vector extracted from the spectral analysis. Further, the allowable domain boundary of the zero-moment point (ZMP) and the center-of-mass height of the inverted pendulum model (LIPM) are corrected according to the ground safety factor. Then, during the robot's walking process, the ground safety factor of each step is marked on the local ground safety map, and the next foothold is selected based on a cost function containing this local ground safety map, and the gait parameters are adjusted online based on the ground safety factor of the current supporting foot. Finally, based on the corrected ZMP allowable domain, center-of-mass height, and gait parameters, the whole-body joint torque is determined, driving the robot to perform the stepping action. Therefore, this scheme applies active mechanical excitation to the ground and collects vibration responses at the moment the foot touches the ground. The foundation type (e.g., solid hard ground / compacted soil / loose sand / cavity) is identified through a combination of spectral analysis and time-domain attenuation criteria, and the foundation safety factor S is calculated. S is then incorporated into the ZMP stability criterion; the smaller S is, the greater the contraction of the ZMP allowable domain. When S=0, a three-level slippage emergency response is triggered. Simultaneously, the LIPM centroid height parameter is corrected based on S, and a foundation safety map is constructed to drive foot placement and online adjustment of gait parameters. This allows the bipedal robot to predict ground bearing capacity solely through foot-based physical perception, even without vision or with unreliable vision, and to actively adjust its balance strategy on soft or potentially collapsing ground, significantly improving walking safety in complex terrain. Attached Figure Description

[0071] Figure 1 A flowchart of a ground-based adaptive gait method for humanoid robots based on foot perception, provided as an embodiment of the present invention. Detailed Implementation

[0072] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0073] like Figure 1 As shown, this invention provides a ground-based adaptive gait method for humanoid robots based on foot perception, which may include the following steps:

[0074] Step 101: When the robot's foot touches the ground and the preset triggering conditions are met, the robot applies active mechanical excitation to the ground through the excitation device integrated in the foot and collects the vibration response signal of the ground through the accelerometer.

[0075] Step 102: Perform spectral analysis on the vibration response signal, and identify the foundation condition type and calculate the foundation safety factor based on the feature vector extracted from the spectral analysis;

[0076] Step 103: Correct the allowable domain boundary of the zero moment point ZMP and the centroid height of the inverted pendulum model LIPM according to the foundation safety factor;

[0077] Step 104: During the robot's walking process, mark the foundation safety factor of each step on the local foundation safety map, select the next footing point based on the cost function containing the local foundation safety map, and adjust the gait parameters online based on the foundation safety factor of the current supporting foot.

[0078] Step 105: Determine the whole-body joint torques based on the corrected ZMP allowable domain, center of mass height, and gait parameters, and drive the robot to perform a stepping motion.

[0079] In this embodiment, the foot is transformed from a "passive force-bearing terminal" into an "active sensing probe." Vibration excitation is applied to the ground at the moment of foot contact, the foundation dynamics response is analyzed, its stiffness and damping properties are inverted, and the foundation safety factor S is calculated. S is then used as a constraint to correct the ZMP stability domain, centroid trajectory, and gait parameters in real time. Specifically, at the moment of foot contact, an active mechanical excitation is applied to the ground via an excitation device, and the vibration response is collected. The foundation type (e.g., solid hard ground / compacted soil / loose sand / cavity) is identified using a spectral analysis and time-domain attenuation criterion, and the foundation safety factor S is calculated. S is then incorporated into the ZMP stability criterion; a smaller S indicates a larger contraction of the ZMP allowable domain. When S=0, a level three misstep emergency response is triggered. Simultaneously, the LIPM centroid height parameter is corrected based on S, and a foundation safety map is constructed to drive the selection of the footing point and the online adjustment of gait parameters. This allows bipedal robots to predict ground bearing capacity solely through physical perception at their feet, even in the absence of vision or under unreliable vision conditions. They can then proactively adjust their balance strategy on soft or potentially collapsing ground, significantly improving walking safety in complex terrain.

[0080] The following provides a more detailed explanation of each of the above steps.

[0081] For step 101, when the robot's foot touches the ground and the preset triggering conditions are met, an active mechanical excitation is applied to the ground through the excitation device integrated on the foot, and the vibration response signal of the ground is collected by the accelerometer.

[0082] In this step, we consider integrating the following sensing and actuation elements into the soles of each foot of the humanoid robot:

[0083] (1) Electromagnetic excitation device: Specifically, it can be a miniature electromagnetic vibrator, embedded in the center area of ​​the sole of the foot, which can generate a sweep frequency excitation force of 10Hz~2kHz, with a maximum amplitude of ±0.5mm and a peak force of 50N.

[0084] (2) Triaxial MEMS accelerometer: installed near the exciter, with a sampling rate of not less than 4kHz and a range of ±16g, used to collect foot vibration response.

[0085] (3) Flexible tactile array: covering the pressure area of ​​the sole (forefoot and heel), composed of piezoresistive or capacitive sensing units, with a sampling rate of 1kHz, used to monitor the pressure distribution and contact status of the sole.

[0086] (4) Six-dimensional force / torque sensor: installed at the ankle joint, measuring range F z ≥2000N, used for measuring ground reaction force and ZMP calculation.

[0087] The activation of the exciter in the foot sensing module requires the simultaneous fulfillment of the following three conditions to constitute a sensing gating: Condition 1: The flexible tactile array on the sole detects a contact area exceeding 30% of the sole area. This condition ensures sufficient physical contact has been established between the foot and the ground, allowing the vibration excitation to be effectively coupled to the foundation, rather than exciting the foot structure itself in a suspended or point-contact state. Condition 2: The six-dimensional force sensor of the ankle joint detects the vertical reaction force F. z > 0.1mg (m is the robot's total mass, g is the acceleration due to gravity). This condition ensures that the foot has already borne part of the robot's weight, and there is sufficient pre-pressure between the foot and the ground, allowing the excitation signal to be effectively transmitted to the ground. Condition 3: The angular velocity of the foot in all three axes is less than a preset threshold. This condition is used to eliminate foot slippage or swaying, ensuring that the foot is in a relatively static and stable contact state at the time of excitation. If all the above conditions are not met, the exciter remains silent, and the system performs gait control based on the previous effective sensing result or default conservative parameters. This mechanism ensures that the vibration response signal truly reflects the foundation characteristics, rather than the foot structure's own vibration or air resistance interference.

[0088] In one embodiment, the sensing action is completed within a brief window (approximately 50-80 ms) during the period when the foot contacts the ground but the body weight has not yet been fully transferred. This timing ensures that foundation information is obtained before the weight is fully loaded, allowing sufficient time for possible emergency responses. Specifically, the timing can be as follows: t=0 ms: The foot contact sensor detects initial contact, and the gating conditions begin to be verified; t=2 ms: All gating conditions are met, and the exciter's push rod extends; t=5 ms: The exciter starts frequency sweeping (lasting 30 ms); t=5~35 ms: The accelerometer synchronously acquires vibration response signals; t=35~40 ms: Excitation stops, and residual vibration is acquired (for calculating the decay time constant τ); t=40~48 ms: FFT processing, feature extraction, and classifier output; t=48~50 ms: The foundation safety factor S value is output to the control layer; after t=50 ms: The body weight begins to transfer to the foot.

[0089] In one embodiment, for applying active mechanical excitation to the ground via a foot-integrated excitation device, the contact dynamics system between the foot and the ground is first modeled as a Kelvin-Voigt viscoelastic model. This model consists of a parallel connection of an elastic element (stiffness k) and a damping element (damping c), and its differential equation of motion is:

[0090]

[0091] Where m is the equivalent mass (kg) of the foot participating in the vibration, which is a known design parameter; c is the equivalent ground damping coefficient (N·s / m); k is the equivalent ground stiffness coefficient (N / m); F(t) is the excitation force (N) applied by the exciter; and x(t) is the foot displacement response (m), which is obtained by the second integration of the accelerometer signal.

[0092] Furthermore, the excitation force F(t) applied by the exciter is expressed as:

[0093]

[0094] in, The excitation force applied to the excitation device, For the amplitude of the excitation force, Let t be the frequency sweep at time t, which linearly sweeps from 10Hz to 2Hz.

[0095] In one embodiment, the excitation force amplitude is considered to be adaptively adjusted, i.e., the excitation force amplitude... Based on the current positive pressure of the foot Dynamic adjustment is used to avoid excessive excitation force on soft ground causing localized damage, or insufficient excitation force on hard ground leading to inadequate signal-to-noise ratio. Specifically, adaptive adjustment is expressed as follows:

[0096]

[0097] in, The maximum excitation force is 50 N. For reference normal force, it is generally taken as 30% of the robot's own weight. The physical basis for this adaptive adjustment is: normal force F z This reflects the degree of preload between the foot and the ground; the greater the preload, the greater the excitation force the ground can withstand. On soft ground, F z The growth is slow, and the excitation force is automatically reduced to avoid localized damage caused by the excitation itself.

[0098] Furthermore, while the exciter applies excitation, a triaxial MEMS accelerometer installed near the exciter acquires the vibration response signal of the foot at a sampling rate of no less than 4kHz. The accelerometer and the exciter are mounted on the same aluminum alloy base to ensure direct transmission of vibration. The exciter's push rod extends upon triggering, directly contacting the ground to transmit vibration, ensuring effective coupling of the excitation signal to the foundation. During normal walking, it retracts into the buffer layer, which alone bears the impact load, protecting the exciter from overload damage.

[0099] For step 102, the vibration response signal is subjected to spectral analysis, and the foundation condition type is identified and the foundation safety factor is calculated based on the feature vector extracted from the spectral analysis.

[0100] In this step, when extracting feature vectors to identify the foundation type and calculating the foundation safety factor, the following approach is considered:

[0101] S21: Perform a fast Fourier transform on the vibration response signal to obtain a frequency domain representation, and calculate the power spectral density based on the frequency domain representation;

[0102] In this step, the time-domain response signal acquired by the accelerometer is considered. Perform a windowed Fast Fourier Transform (FFT) to obtain the frequency domain representation X(f), and then calculate the one-sided power spectral density, as shown below:

[0103]

[0104] in, For power spectral density, Let T be the frequency domain representation, and T be the sampling time window. The sampling frequency.

[0105] S22: Extract frequency domain feature vectors from the power spectral density; wherein the frequency domain feature vectors include peak frequency, stiffness estimate, damping ratio estimate, and low-frequency energy proportion.

[0106] In this step, we consider extracting the following four features from the power spectral density (PSD) as the feature vector f for ground-based semantic classification. PSD They are respectively:

[0107]

[0108] in, Peak frequency, This is an estimate of the stiffness. The equivalent mass of the foot participating in vibration, i.e., the reduced mass of a single robot leg participating in vibration at the moment of contact with the ground, is determined based on the modal mass distribution coefficient and is approximately 1 / 6 to 1 / 4 of the robot's total mass. and These are the half-power point frequencies, i.e., the frequencies corresponding to a 3dB drop in the PSD peak value, used for damping ratio calculation. This represents the proportion of low-frequency energy.

[0109] S23: After the excitation stops, the residual vibration signal is collected by the accelerometer, the amplitude envelope of the residual vibration signal is extracted by Hilbert transform, and the decay time constant is calculated as the time domain feature.

[0110] In this step, considering the ambiguity in distinguishing between "loose sand" and "cavity" solely based on frequency domain PSD features, a time-domain energy decay rate is added as a supplementary criterion. After the exciter stops excitation, the accelerometer continues to collect residual vibration signals. A Hilbert transform is performed on the residual vibration signal to extract the amplitude envelope A(t). The decay time constant is defined. satisfy ;in, The amplitude envelope of the residual oscillation signal. The decay time constant, It is a natural constant.

[0111] The continuous reverberation caused by reflections inside the cavity leads to structural problems in the cavity. The energy is significantly greater than that of solid structures, while the energy of loose soil decays extremely rapidly. Therefore, the joint criterion can be obtained as follows: Based on this joint criterion, the false alarm rate of the cavity can be significantly reduced.

[0112] S24: Use the frequency domain feature vector and the time domain feature as input to a pre-trained classifier and output the ground type.

[0113] In this step, the foundation type classifier is obtained through pre-training. In one embodiment, the classifier can also be updated online adaptively, and the specific update process may include:

[0114] Step A1: During the robot's walking process, collect the vibration response feature vector f for each step. PSDThe temporal attenuation characteristic τ, and the corresponding classifier output foundation type and safety factor S;

[0115] Step A2: During subsequent walking, continuously monitor actual walking performance data, including: whether abnormal settlement of the foot exceeds the preset threshold, whether the foot slips, and the posture angle recovery rate after gait execution;

[0116] Step A3: When the actual walking performance is inconsistent with the safety level predicted by the classifier. Specifically: the classifier determines it as "solid hard ground" (S≥0.7) but abnormal settlement or slippage actually occurs, or the classifier determines it as "loose sand" (S≤0.4) but the actual walking performance is good. In this case, the feature vector corresponding to this step is marked as a "misclassified sample".

[0117] Step A4: When the accumulated number of "misclassified samples" reaches a preset threshold N (N≥10), during robot idle or charging periods, the model parameters of the classifier are incrementally updated using the misclassified samples. The update method is as follows: for a support vector machine-based classifier, the misclassified samples are added to the support vector set with decreasing weights; for a neural network-based classifier, the misclassified samples are fine-tuned several times using a reduced learning rate.

[0118] Step A5: After the incremental update is completed, the updated classifier is used for ground identification in subsequent walks, so that the classifier gradually adapts to the specific ground characteristics of the current working environment and continuously improves the classification accuracy.

[0119] S25: Calculate the foundation safety factor based on the stiffness estimate and the time-domain characteristics.

[0120] In this step, the foundation types can be divided into four categories, each corresponding to different physical characteristics and safety levels, as shown in Table 1 below:

[0121] Table 1

[0122] The foundation safety factor S can be determined as follows:

[0123]

[0124] Where S is the foundation safety factor. Minimum stiffness required for safe standing This is the minimum stiffness value. The preset stiffness threshold, The preset decay time threshold, which has been experimentally calibrated, typically ranges from 15 to 25 ms. This is the preset threshold for the proportion of low-frequency energy.

[0125] In this embodiment, a Level 3 emergency response can be further executed when S=0. Specifically, the Level 3 emergency response may include:

[0126] Level 1 (0-20ms, sensor trigger): The ankle joint stops stepping, the knee joint contracts and lifts the foot, and the torso tilts to the opposite side at a preset angle. This preset angle can be 5°.

[0127] Level 2 (20-100ms, balance protection): Contralateral ankle joint torque compensation, maintaining balance through a hybrid control of ankle and hip strategies; the control formula is as follows: , This indicates the ankle joint compensating torque. Represents the proportional gain coefficient. Indicates the desired joint angle. Indicates the actual joint angle. Indicates the gain coefficient of micronized powder. The compensating torque for the failure of the side that has lost its footing is entirely supported by the opposite leg.

[0128] Level 3 (100-500ms): Mark the point of misstep as a permanent danger zone, replan the landing point and detour.

[0129] For step 103, the allowable domain boundary of the zero moment point ZMP and the centroid height of the inverted pendulum model LIPM are corrected according to the foundation safety factor.

[0130] In this step, the foundation safety factor S is considered as a correction factor for the ZMP stability domain, and the LIPM model is incorporated into the S constraint to achieve a closed loop of "perception-model correction-constraint-control".

[0131] Specifically, in one embodiment, when correcting the allowable domain boundary of the zero moment point ZMP based on the foundation safety factor, it may include:

[0132] S31: Calculate boundary shrinkage ;in, This is the preset maximum shrinkage amount.

[0133] The physical significance of this step is that the softer the ground, the smaller the allowable range of center of gravity shift, forcing the robot to maintain a more conservative posture.

[0134] S32: The boundary contraction amount First-order low-pass filtering is performed for smoothing to obtain the smoothed boundary shrinkage. ;in, , To control the cycle, This is the smoothing time constant.

[0135] In this step, to avoid a step change at the boundary of the ZMP allowable domain caused by changes in the S value, which could lead to a sudden change in ankle joint torque command, the boundary contraction amount was considered. A first-order low-pass filter is applied for smoothing. This smoothing process ensures the continuity of the supporting torque during the switching of the plantar pressure center, avoiding joint torque shocks caused by sudden changes in the S value.

[0136] S33: Shrink the allowable domain of the zero moment point ZMP from the original supporting polygon. The shrunken domain is represented as:

[0137]

[0138] in, The boundary of the allowable domain of the corrected zero moment point ZMP. This is the smoothed boundary shrinkage amount. and Let x be the lower and upper limits in the x-direction. and These are the lower and upper limits in the y-direction; when S=0, the foot is invalid, and the center of gravity shifts to the lateral foot.

[0139] In this embodiment, the traditional ZMP stability condition requires that the calculated ZMP lies within the support polygon, i.e. , The ZMP position is calculated by a six-dimensional force / torque sensor. , h is the vertical distance from the sensor to the ground. and These are the ankle joint moments in the x and y directions, respectively. and This represents the ground reaction force component.

[0140] Furthermore, we consider modifying the ZMP allowable domain by combining traditional geometric constraints with the foundation safety factor S. This involves shrinking the allowable domain of the zero moment point ZMP from the original support polygon to obtain the aforementioned shrunken representation.

[0141] In one embodiment, when calculating the centroid height correction term based on the foundation safety factor, it may include: the traditional linear inverted pendulum model (LIPM) assumes that the support surface has infinite rigidity and the centroid height z c Constant, which is expressed as: When the foundation is soft ground (S<1), the support legs experience slight settlement during loading. This leads to a time-varying error in the equivalent centroid height. Therefore, this invention considers introducing a centroid height correction term based on S, expressed as follows:

[0142]

[0143] in, To correct the height of the center of mass, Standard center of gravity height The maximum settlement is estimated using the static relationship between the foot normal force and the foundation stiffness. For the total mass of the robot, It is the acceleration due to gravity. Minimum stiffness value, This refers to the foundation settlement.

[0144] Furthermore, the corrected centroid height Substituting the dynamic equations of the inverted pendulum model LIPM This yields a corrected acceleration command. This correction adapts the center-of-mass trajectory to the actual ground stiffness, preventing overshoot or oscillations caused by model mismatch. When S=1, It degenerates into the standard LIPM.

[0145] For step 104, during the robot's walking process, the foundation safety factor of each step is marked on the local foundation safety map, and the next foothold is selected based on the cost function containing the local foundation safety map, and the gait parameters are adjusted online based on the foundation safety factor of the current supporting foot.

[0146] In this step, we first consider marking the S value of each step on an environmental grid map centered on the robot during its movement, forming a ground safety probability map M(x,y):

[0147]

[0148] in, This is a map showing the probability of foundation safety determined at time t. The foundation safety factor is determined at time t. The state of each grid includes: solid hard ground, compacted soil, loose sand, cavity, and unexplored.

[0149] Then, when generating candidate landing points, the walking planner combines the following cost function and selects the candidate point with the lowest cost as the landing point:

[0150]

[0151] in, This represents the total cost function value. The weighting coefficients represent the distance cost. Indicates the coordinates of the candidate landing point. Indicates the coordinates of the expected landing point. The weighting coefficients representing the security costs, These are the weighting coefficients for the location region cost. This represents the penalty cost for unexplored areas. The first term in the cost function is the distance cost, which is the penalty for deviating from the desired landing point; the second term is the safety cost, with a higher penalty for a lower foundation safety factor; the third term is the area cost. This indicates that the area is unexplored, and choosing a known safe area is encouraged.

[0152] Furthermore, consider adjusting gait parameters in real time based on the current foundation safety factor S of the supporting foot. The adjustment logic is as follows: the softer the ground, the smaller the stride, the slower the speed, the higher the foot lifts (to prevent dragging and sinking), and the lower the center of gravity (to reduce the overturning moment). Specifically, the adjustment formula is as follows:

[0153]

[0154] In one embodiment, the adjustment results of the gait parameters can be shown in Table 2 below:

[0155] Table 2 stride length 0.4 m 0.25 m 0.15 m Gait cycle 0.8 s 1.0 s 1.5 s foot height 0.03 m 0.05 m 0.08 m Center of gravity height 0.85 m 0.80 m 0.75 m ZMP allows domain shrinkage Δ 0 15% 35%

[0156] Step 105: Determine the whole-body joint torques based on the corrected ZMP allowable domain, center of mass height, and gait parameters, and drive the robot to perform a stepping motion.

[0157] In this step, we first consider the modified ZMP allowable domain. The desired ZMP target position is calculated. When S>0, the ZMP target position is within the corrected allowable domain (preferably near the domain center); when S=0, the ZMP target position immediately shifts to the center of the opposite supporting foot. Then, based on the corrected LIPM dynamic equations and the desired ZMP trajectory, the robot's center of mass trajectory is generated. Further, based on the generated center of mass trajectory and gait parameters (stride length, gait period, foot lift height, center of gravity height), the desired angle trajectory of each joint is calculated using inverse kinematics. Then, a control method based on calculated torque is used to calculate the driving torque of each joint. Finally, the calculated joint torque commands are sent to the servo drivers of each joint to drive the robot to perform a stepping action and complete a single-step walking motion. After performing the stepping action, the system returns to the initial step, waiting for the next foot to land, and then re-triggers the perception-control closed loop. The entire process is completed within a single control cycle (<5ms), ensuring real-time performance.

[0158] The following is a practical verification scenario for different ground types, and the results are shown below:

[0159] Solid hard surfaces: laboratory tile flooring, concrete flooring. S=1.0, ZMP allows fully supported polygons, LIPM model requires no modification, gait parameters are at default values.

[0160] Compacted soil layer: outdoor compacted dirt roads and grass. S≈0.7, ZMP allowable shrinkage 15%, centroid height correction δz≈0.3·δz max The stride length was reduced to 62%.

[0161] Loose sandy land: dry sand pits, gravelly ground. S≈0.4, ZMP allowable domain shrinkage 35%, centroid height correction δz≈0.6·δz max The gait changes to "small steps and slow walking", the center of gravity is lowered, and the feet are lifted higher to prevent dragging and getting stuck.

[0162] Cavity / Thin Shell: Hidden pits or hollow bricks covered with thin plywood. A joint criterion (frequency domain + time domain attenuation rate) triggers S=0.0, immediately initiating a Level 3 emergency response for slippage; the robot retracts the foot before weight loading.

[0163] Furthermore, it should be noted that when ambient lighting is good and the vision system is available, the terrain type pre-identified by vision can serve as a prior estimate of S, which is then verified and corrected by foot perception. When the two contradict each other (e.g., vision determines a "flat road surface" but foot perception detects a cavity), the foot perception result takes precedence to ensure safety. When the vision system fails, foot perception operates independently, without relying on vision.

[0164] Furthermore, when the foot sensing module malfunctions in the following ways, the system automatically degrades its operation to ensure the robot can still walk safely even when the sensor fails:

[0165] Exciter malfunction: Failure to start or abnormal current indicates hardware failure;

[0166] Accelerometer saturation or noise exceeding limits: peak-to-peak signal exceeds 80% of the range or signal-to-noise ratio is below 10dB;

[0167] Low confidence in multiple consecutive classifications: The maximum probability output by the classifier in three consecutive steps is less than 0.6.

[0168] Specifically, the degradation mode may include: forcibly setting the S-value to a conservative value of 0.3; shrinking the ZMP allowable domain to 50% of the baseline; switching gait parameters to a preset "conservative parameter set" (stride length 0.1m, cycle 2.0s, center of gravity lowered by 20%, foot lift height increased by 50%); if the vision system is available, prioritizing visual terrain classification results as a supplement. The degradation state continues until the perception module returns to normal (classification confidence > 0.8 for 3 consecutive steps), or is deactivated by the operator via a remote interface.

[0169] In one embodiment, the present invention also provides a humanoid robot ground-based adaptive gait system based on foot perception, comprising: a foot perception module, a ground recognition module, a correction module, a gait planning module, and an execution control module;

[0170] The foot-end sensing module is used to apply active mechanical excitation to the ground through an excitation device integrated in the sole of the foot when the robot's foot touches the ground and a preset trigger condition is met, and to collect the vibration response signal of the ground through an accelerometer.

[0171] The foundation identification module is used to perform spectral analysis on the vibration response signal, and identify the foundation condition type and calculate the foundation safety factor based on the feature vector extracted from the spectral analysis.

[0172] The correction module is used to correct the allowable domain boundary of the zero moment point ZMP and the centroid height of the inverted pendulum model LIPM according to the foundation safety factor.

[0173] The gait planning module is used to mark the foundation safety factor of each step on the local foundation safety map during the robot's walking process, select the next footing point based on the cost function containing the local foundation safety map, and adjust the gait parameters online based on the foundation safety factor of the current supporting foot.

[0174] The execution control module is used to calculate the whole-body joint torque based on the corrected ZMP allowable domain, center of mass height and gait parameters, and drive the robot to perform stepping movements.

[0175] In summary, the humanoid robot ground-based adaptive gait method and system based on foot perception provided by this invention have at least the following beneficial effects:

[0176] (1) Breaking through visual dependence to achieve active perception of the physical properties of the foundation. Traditional humanoid robots rely on vision or lidar for terrain prediction, and cannot work in environments where vision fails, such as dust, smoke, water reflection, or insufficient lighting. This invention integrates a micro electromagnetic vibrator and accelerometer at the foot end to apply active vibration excitation to the ground at the moment of foot landing. By using power spectral density (PSD) feature extraction and Kelvin-Voigt model parameter identification, the stiffness coefficient and damping ratio of the foundation can be directly inverted, so that the robot can perceive the physical properties of the ground without relying on any optical sensors, fundamentally solving the problem of visual blind spots.

[0177] (2) From "passive response" to "active early warning," breaking through the lag bottleneck of traditional ZMP control. Traditional ZMP stability criteria rely on IMU measurements of already occurring attitude changes. When ZMP deviation from the support domain is detected, the robot is often already on the verge of instability. This invention incorporates the foundation safety factor S as a first-level constraint into the ZMP stability domain. When S < 1, the ZMP allowable boundary is actively contracted, forcing the robot to maintain a more conservative attitude. When S = 0 (cavity is detected), a three-level emergency response is triggered within 0-20ms, including ankle joint emergency stop, knee joint contraction and foot lifting, and trunk tilting. This achieves a leap from post-event compensation to pre-event avoidance.

[0178] (3) Make the ZMP stability criterion applicable to non-rigid support surfaces. The traditional ZMP theory implicitly assumes that the support surface is infinitely rigid, and the criterion fails when the ground itself may collapse. This invention dynamically shrinks the ZMP allowable domain by using the foundation safety factor S, and introduces a first-order low-pass filter to smooth the surface and avoid the moment impact caused by boundary step jumps, so that the ZMP stability criterion has the adaptive capability to foundation bearing safety for the first time.

[0179] (4) Correcting the ground settlement error of the LIPM model to improve walking stability. The traditional linear inverted pendulum model (LIPM) assumes a constant center of mass height and ignores the slight settlement of the supporting foot on soft ground. This invention introduces a center of mass height correction term based on S, which enables the center of mass trajectory to adapt to the actual ground stiffness in real time, avoiding overshoot or oscillation caused by model mismatch, and significantly improving walking stability on soft ground.

[0180] (5) Constructing a foundation safety map to achieve online gait planning with safety as the priority. This invention marks the foundation safety coefficient S identified for each step in a local grid map to form a foundation safety probability map. The gait planner selects the landing point through a cost function and adjusts the stride, cycle, foot lift height and center of gravity height in real time according to the S value, realizing the continuous learning ability of exploring, memorizing and optimizing as soon as it is walked.

[0181] (6) The sensing-control closed-loop response speed meets real-time requirements. The entire six-step closed loop of sensing-identification-model correction-constraint-planning-control is completed within a single control cycle. The exciter of the foot sensing module must simultaneously meet three interlocking conditions to ensure that the vibration response signal truly reflects the foundation characteristics rather than the natural vibration of the foot structure or air resistance interference. The excitation force amplitude is dynamically and adaptively adjusted according to the normal pressure of the foot, automatically reducing the excitation force on soft ground to avoid local damage. Cavity detection adopts a joint criterion of frequency domain PSD + time domain attenuation rate, which significantly reduces the false alarm rate.

[0182] This specification also provides a computer-readable storage medium having a computer program stored thereon, which, when executed in a computer, causes the computer to perform the methods in any of the embodiments of the specification.

[0183] This specification also provides a computing device, including a memory and a processor, wherein the memory stores executable code, and when the processor executes the executable code, it implements the method in any of the embodiments of the specification.

[0184] Since the system embodiments and method embodiments of the present invention are based on the same inventive concept, detailed descriptions can be found in the method embodiments, and will not be repeated here.

[0185] The modules or units in the device of this invention can be merged, divided, and deleted according to actual needs. The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the processes of the above embodiments and making equivalent changes according to the claims of this invention still fall within the scope of the invention.

Claims

1. A humanoid robot ground adaptive gait method based on foot end sensing, characterized by, include: When the robot's foot touches the ground and the preset triggering conditions are met, the robot applies active mechanical excitation to the ground through the excitation device integrated in the foot and collects the vibration response signal of the ground through the accelerometer. The vibration response signal is subjected to spectral analysis, and the foundation condition type is identified and the foundation safety factor is calculated based on the feature vector extracted from the spectral analysis. The allowable boundary of the zero moment point ZMP and the height of the center of mass of the inverted pendulum model LIPM are corrected based on the foundation safety factor. During the robot's walking process, the foundation safety factor of each step is marked on the local foundation safety map, and the next foothold is selected based on the cost function containing the local foundation safety map. The gait parameters are also adjusted online based on the foundation safety factor of the current supporting foot. Based on the corrected ZMP allowable domain, center of mass height, and gait parameters, the whole-body joint torques are determined to drive the robot to perform stepping movements.

2. The foot-based perception based humanoid robot gait adaptation method according to claim 1, wherein, The preset triggering condition is a perception-triggered gating mechanism, which must simultaneously meet the following conditions: The flexible tactile array on the sole of the foot detected that the contact area exceeded a first preset threshold of the sole area; The ankle joint six-dimensional force sensor detected a vertical reaction force greater than the second preset threshold. The angular velocities of the foot in all three axes are less than the third preset threshold. 3.The foot-end sensory based humanoid robot gait adaptation method according to claim 1, wherein, The active mechanical stimulation applied to the ground via the foot-integrated stimulation device includes: The excitation force applied by the excitation device is expressed as: ； in, The excitation force applied to the excitation device, For the amplitude of the excitation force, The frequency sweep at time t; The excitation force amplitude Based on the current positive pressure of the foot Dynamic adjustment is represented as follows: ； in, For maximum excitation force, For reference positive pressure.

4. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 1, characterized in that, The step of performing spectral analysis on the vibration response signal, identifying the foundation condition type based on the feature vector extracted from the spectral analysis, and calculating the foundation safety factor includes: The vibration response signal is subjected to a fast Fourier transform to obtain a frequency domain representation, and the power spectral density is calculated based on this frequency domain representation. Extract frequency domain feature vectors from the power spectral density; wherein, the frequency domain feature vectors include peak frequency, stiffness estimate, damping ratio estimate, and low-frequency energy proportion; After the excitation stops, the residual vibration signal is collected by an accelerometer, the amplitude envelope of the residual vibration signal is extracted by Hilbert transform, and the decay time constant is calculated as a time domain feature. The frequency domain feature vector and the time domain feature vector are used as inputs to a pre-trained classifier, which outputs the ground type. The foundation safety factor is calculated based on the stiffness estimate and the time-domain characteristics.

5. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 4, characterized in that, The power spectral density is calculated as follows: ； in, For power spectral density, Let T be the frequency domain representation, and T be the sampling time window. The sampling frequency; The extraction methods for the frequency domain feature vector include: ； in, Peak frequency, This is an estimate of the stiffness. The equivalent mass of the foot participating in vibration, and These are the half-power point frequencies, The proportion of low-frequency energy; The attenuation time constant satisfies ;in, The amplitude envelope of the residual oscillation signal. The decay time constant, It is a natural constant; The calculation of the foundation safety factor based on the stiffness estimate and the time-domain characteristics includes determining it in the following manner: ； Where S is the foundation safety factor. Minimum stiffness required for safe standing This is the minimum stiffness value. The preset stiffness threshold, The preset decay time threshold, This is the preset threshold for the proportion of low-frequency energy.

6. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 5, characterized in that, The permissible domain boundary for correcting the zero moment point ZMP based on the foundation safety factor includes: Calculate boundary shrinkage ;in, The preset maximum shrinkage amount; For the boundary shrinkage amount First-order low-pass filtering is performed for smoothing to obtain the smoothed boundary shrinkage. ;in, , To control the cycle, The smoothing time constant; The allowable domain of the zero-moment point ZMP is shrunk from the original supporting polygon, and the shrunk domain is represented as: ； in, The boundary of the allowable domain of the corrected zero moment point ZMP. This is the smoothed boundary shrinkage amount. and Let x be the lower and upper limits in the x-direction. and These are the lower and upper limits in the y-direction.

7. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 6, characterized in that, The step of correcting the centroid height of the inverted pendulum model LIPM based on the foundation safety factor includes: The correction term for the centroid height calculated based on the foundation safety factor is expressed as follows: ； in, To correct the height of the center of mass, Standard center of gravity height For the maximum settlement, For the total mass of the robot, It is the acceleration due to gravity. Minimum stiffness value, This refers to the foundation settlement. The corrected centroid height Substituting the dynamic equations of the inverted pendulum model LIPM This is to make the center of mass trajectory adapt to the actual stiffness of the ground.

8. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 5, characterized in that, The foundation safety map is shown below: ； in, This is a map showing the probability of foundation safety determined at time t. The foundation safety factor is determined at time t; The cost function is expressed as follows: ； in, This represents the total cost function value. The weighting coefficients represent the distance cost. Indicates the coordinates of the candidate landing point. Indicates the coordinates of the expected landing point. The weighting coefficients representing the security costs, These are the weighting coefficients for the location region cost. This indicates the punitive value of unexplored areas; The online adjustment of gait parameters based on the current foundation safety factor of the supporting foot includes: ； Where S is the foundation safety factor.

9. The humanoid robot ground-based adaptive gait method based on foot perception according to claim 5, characterized in that, When S=0, the method also includes a three-level emergency response; the three-level emergency response includes: Level 1: Stop stepping with the ankle joint, contract the knee joint to lift the foot, and tilt the torso to the opposite side at a preset angle; Level 2: Contralateral ankle joint torque compensation, maintaining balance through a hybrid control of ankle and hip strategies; the control formula is: , This indicates the ankle joint compensating torque. Represents the proportional gain coefficient. Indicates the desired joint angle. Indicates the actual joint angle. Indicates the gain coefficient of micronized powder. This is the compensation torque after the side that was stepped on fails; Level 3: Mark the point of misstep as a permanent danger zone, replan your landing point and detour.

10. A ground-based adaptive gait system for humanoid robots based on foot perception, characterized in that, include: Foot perception module, ground recognition module, correction module, gait planning module, and execution control module; The foot-end sensing module is used to apply active mechanical excitation to the ground through an excitation device integrated in the sole of the foot when the robot's foot touches the ground and a preset trigger condition is met, and to collect the vibration response signal of the ground through an accelerometer. The foundation identification module is used to perform spectral analysis on the vibration response signal, and identify the foundation condition type and calculate the foundation safety factor based on the feature vector extracted from the spectral analysis. The correction module is used to correct the allowable domain boundary of the zero moment point ZMP and the centroid height of the inverted pendulum model LIPM according to the foundation safety factor. The gait planning module is used to mark the foundation safety factor of each step on the local foundation safety map during the robot's walking process, select the next footing point based on the cost function containing the local foundation safety map, and adjust the gait parameters online based on the foundation safety factor of the current supporting foot. The execution control module is used to calculate the whole-body joint torque based on the corrected ZMP allowable domain, center of mass height and gait parameters, and drive the robot to perform stepping movements.

Images