A shock absorption balance control method and system for a quadruped robot

Abstract

This invention relates to the field of robot control technology, specifically to a vibration damping and balance control method and system for a quadruped robot. The method includes: acquiring the linear velocity and angular velocity of the quadruped robot during a monitoring period; determining the vertical velocity difference value and vertical constraint force index during the monitoring period; obtaining a deviation suppression weight based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index, combined with the balance degree of the vertical constraint force index; and using the deviation suppression weight to weight the solver cost function of the model predictive control algorithm to output the optimal control torque for controlling the quadruped robot. This invention achieves foot-adaptive control of the quadruped robot, maintaining high-precision trajectory tracking on hard surfaces and adapting to ground subsidence on soft surfaces through compliant torque output, fundamentally blocking the positive feedback loop of oscillations while ensuring robot posture balance and significantly improving all-terrain adaptability.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a shock absorption and balance control method and system for a quadruped robot. Background Technology

[0002] With the continuous upgrading of operational demands in fields such as field exploration, material transportation, and disaster relief, quadruped robots, with their flexible movement characteristics, have become an important vehicle for performing tasks in complex environments. In operational scenarios with structured hard ground, existing model predictive control (MPC)-based control algorithms, based on the core assumption of zero velocity constraint where the legs remain stationary when in contact with the ground, perform state estimation and torque calculation, enabling high-precision trajectory tracking and attitude balance control of the robot, thus ensuring the stability of the operation process.

[0003] In actual operations on unstructured soft ground such as sand, snow, and mud, the ground surface undergoes unavoidable plastic deformation under the pressure of the robot's feet, resulting in unexpected vertical settlement displacement of the feet. In this situation, traditional control algorithms relying on zero-velocity constraints misjudge this physical settlement as a system state error and calculate a large reverse support torque to attempt to restore the feet to a preset resting position. This type of control logic not only fails to effectively suppress the plastic deformation of the ground surface but also triggers a significant force-velocity linkage oscillation phenomenon: the greater the constraint force output by the controller, the faster the feet press deeper into the ground, thus triggering the controller to generate an even larger reverse torque demand. This positive feedback mechanism causes severe shaking of the robot's posture, and in severe cases, it can cause joint motor overload, even leading to robot instability and tipping, directly affecting the smooth progress of the task. Therefore, existing MPC control algorithms, based on structured hard ground surfaces for controlling quadruped robots, are prone to ground stiffness mismatch, resulting in poor rationality of the quadruped robot's control strategy. Summary of the Invention

[0004] To address the technical problem of poor control strategy rationality for quadruped robots using existing model predictive control-based control algorithms, the present invention aims to provide a vibration damping and balance control method and system for quadruped robots. The specific technical solution adopted is as follows: In a first aspect, the present invention provides a shock absorption and balance control method for a quadruped robot, comprising: The linear velocity and angular velocity of the quadruped robot during the monitoring period are obtained. Based on the linear velocity and angular velocity, the difference between the actual vertical velocity and the theoretical vertical velocity of the quadruped robot are analyzed to determine the vertical velocity difference value during the monitoring period. The parameter variables in the vertical velocity constraint of model predictive control based on quadruped robots are used to determine the vertical constraint force index during the monitoring period. Based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index during the monitoring period, and combined with the balance of the vertical constraint force index, the deviation suppression weight during the monitoring period is obtained. By using deviation suppression weights to weight the solver cost function of the model predictive control algorithm, the optimal control torque is output for controlling the quadruped robot.

[0005] Preferably, the step of obtaining the deviation suppression weight within the monitoring period based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index during the monitoring period, combined with the balance degree of the vertical constraint force index, specifically includes: Based on the consistency between the vertical velocity difference value at each moment within the monitoring period and the fluctuation trend of the vertical constraint force index at each moment, the positive feedback correlation coefficient of the monitoring period is obtained. Based on the balanced distribution of vertical velocity differences within the monitoring period, and combined with the positive feedback correlation coefficient, a linear mapping is performed to obtain the deviation suppression weight within the monitoring period.

[0006] Preferably, the step of obtaining the positive feedback correlation coefficient for the monitoring period based on the consistency between the vertical velocity difference value at each moment within the monitoring period and the fluctuation trend of the vertical constraint force index at each moment specifically includes: Based on the temporal characteristics and corresponding vertical velocity difference values ​​at each moment within the monitoring period, the weighted average velocity difference within the monitoring period is determined; based on the temporal characteristics and corresponding vertical constraint force indices at each moment within the monitoring period, the weighted average constraint force within the monitoring period is determined. Based on the vertical velocity difference and weighted average velocity difference at each moment within the monitoring period, as well as the vertical constraint force index and weighted average constraint force at each moment, the covariance is calculated to determine the positive feedback correlation coefficient for the monitoring period.

[0007] Preferably, the step of obtaining the deviation suppression weight within the monitoring period by performing a linear mapping based on the balanced distribution of vertical velocity differences within the monitoring time period and the positive feedback correlation coefficient specifically includes: The product of the normalized coefficient of the weighted mean velocity difference and the normalized coefficient of the positive feedback correlation coefficient during the monitoring period is negatively correlated to determine the deviation suppression weight during the monitoring period.

[0008] Preferably, the method for obtaining the weighted average velocity difference is as follows: using the time sequence corresponding to each moment within the monitoring period as the weight, the vertical velocity difference value at each moment is weighted and the average is calculated to obtain the weighted average velocity difference within the monitoring period; The method for obtaining the weighted average constraint force is as follows: using the time sequence corresponding to each moment within the monitoring period as the weight, the vertical constraint force index at each moment is weighted and the average is calculated to obtain the weighted average constraint force for the monitoring period.

[0009] Preferably, the step of analyzing the difference between the actual vertical velocity and the theoretical vertical velocity of the quadruped robot based on linear velocity and angular velocity, respectively, and determining the vertical velocity difference value within the monitoring period, specifically includes: Extract the vertical velocity component of the linear velocity of the quadruped robot at each moment during the monitoring period to obtain the actual vertical velocity during the monitoring period. Obtain the angle data of each leg of the quadruped robot at each time step and the rotation matrix of the quadruped robot relative to the world coordinate system; Based on the angle, the Jacobian matrix is ​​solved, and the angular velocity of the quadruped robot during the monitoring period is mapped using the Jacobian matrix and the rotation matrix to obtain the theoretical vertical velocity during the monitoring period. The absolute value of the difference between the actual vertical velocity and the theoretical vertical velocity at the same moment within the monitoring period is taken as the vertical velocity difference value at each moment within the monitoring period.

[0010] Preferably, the step of solving the Jacobian matrix based on the angle, and using the Jacobian matrix and the rotation matrix to map the angular velocity of the quadruped robot during the monitoring period to obtain the theoretical vertical velocity during the monitoring period, specifically includes: For any given moment within the monitoring period; Obtain the angle of each leg of the quadruped robot at each joint point to construct the angle vector of each leg; obtain the angular velocity of each leg of the quadruped robot at each joint point to construct the angular velocity vector of each leg. The Jacobian matrix corresponding to each leg is solved based on the angle vector of each leg. The theoretical velocity vector of each leg is determined based on the dot product between the rotation matrix, the Jacobian matrix of each leg, and the angular velocity vector. The vertical component of the theoretical velocity vector of each leg is extracted to obtain the theoretical vertical velocity of each leg of the quadruped robot at any given time.

[0011] Preferably, the parameter variables in the vertical velocity constraint for model predictive control based on a quadruped robot, determining the vertical constraint force index within the monitoring period, specifically include: During the monitoring period, the absolute values ​​of the Lagrange multipliers in the model predictive control solution process at the previous time point are obtained as the vertical constraint force index at the current time point.

[0012] Preferably, the method further includes: when the weighted average velocity difference is less than a preset silent velocity threshold, setting the value of the positive feedback correlation coefficient to 0.

[0013] In a second aspect, the present invention provides a shock absorption and balance control system for a quadruped robot, including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the computer program, when executed by the processor, implements the steps of a shock absorption and balance control method for a quadruped robot.

[0014] The embodiments of the present invention have at least the following beneficial effects: This invention first collects core motion data such as linear velocity and angular velocity of a quadruped robot during a monitoring period, distinguishing and extracting the theoretical vertical velocity reflecting the actual motion state, as well as the theoretical vertical velocity derived based on the rigid ground assumption. It captures key characteristics of unexpected foot settlement, establishing an observational basis for ground stiffness mismatch. Then, based on the relevant parameter variables of the vertical velocity constraint in model predictive control, it quantifies the vertical constraint force index. This index directly reflects the force exerted by the controller to maintain foot stillness, clearly characterizing the force application intention of the control strategy. Furthermore, by analyzing the correlation between vertical velocity differences and fluctuations in the vertical constraint force index (capturing oscillation signs), and combining this with the balanced distribution of the vertical constraint force index (avoiding noise interference), the correlation characteristics are transformed into dynamically changing deviation suppression weights. This achieves adaptive adjustment of control stiffness, accurately matching different ground conditions. Finally, the dynamically generated deviation suppression weights are incorporated into the cost function of the model predictive control solver, realizing foot-adaptive control of the quadruped robot. High-precision trajectory tracking is maintained on hard ground, and the robot adapts to ground subsidence through compliant torque output on soft ground, fundamentally blocking the positive feedback loop of oscillation, while ensuring the robot's posture balance and significantly improving its all-terrain adaptability. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the steps of a shock absorption and balance control method for a quadruped robot provided by the present invention; Figure 2 This is a flowchart of the steps in the method for obtaining the vertical velocity difference value provided by the present invention; Figure 3 This is a flowchart of the steps for obtaining the deviation suppression weight within the monitoring time period provided by the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, structure, features, and effects of a shock absorption and balance control method and system for a quadruped robot proposed according to the present invention. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of the shock absorption and balance control method and system for a quadruped robot provided by this invention.

[0020] Please see Figure 1 The diagram illustrates a flowchart of a shock absorption and balance control method for a quadruped robot according to an embodiment of the present invention. The method includes the following steps: Step S100: Obtain the linear velocity and angular velocity of the quadruped robot during the monitoring period. Based on the linear velocity and angular velocity, analyze the difference between the actual vertical velocity and the theoretical vertical velocity of the quadruped robot, and determine the vertical velocity difference value during the monitoring period.

[0021] The main purpose of this step is to capture the vertical motion deviation characteristics of the quadruped robot. By distinguishing the difference between the absolute vertical velocity performance under actual motion and the vertical velocity performance under theoretical motion, the unexpected foot settlement effect is quantified, providing core observational data for subsequent identification of oscillations caused by ground stiffness mismatch. On unstructured soft ground, ground deformation can cause unexpected vertical displacement of the robot's feet, and this displacement can only be accurately characterized by comparing motion states.

[0022] Therefore, this step first collects the core motion data of the robot during the monitoring period, and then, based on these raw data, extracts the actual motion speed in the vertical dimension and derives the theoretical vertical motion speed that conforms to the rigid ground assumption. Finally, by comparing the two types of vertical speeds at the same moment, the difference in vertical speed that can reflect the degree of foot settlement is obtained, laying a data foundation for subsequent analysis of the correlation between control behavior and settlement response.

[0023] As a concrete example, such as Figure 2 As shown, the method for obtaining the vertical velocity difference value can be implemented by steps S101 to S105.

[0024] Step S101: Obtain the linear velocity, angle, and angular velocity of the quadruped robot at each moment during the monitoring period.

[0025] It should be understood that since the inertial measurement unit (IMU) installed on the robot body can independently sense the motion state of the body relative to inertial space and is not affected by the contact between the feet and the ground, its calculation results can be used as the physical reference for the actual motion of the body.

[0026] Specifically, the linear velocity of the quadruped robot at each moment during the monitoring period is directly read from the API interface of the airborne state estimator. The linear velocity is represented as a 3×1 column vector, reflecting the movement speed of the robot's center of mass along the three axes (x, y, z) of the world coordinate system.

[0027] More specifically, the angles of each leg of the quadruped robot at each joint point, as well as the angular velocity of each leg at each joint point, are obtained from the inertial measurement unit (IMU) mounted on the robot body. A typical leg of a quadruped robot contains three joints: a lateral hip joint, a forward hip joint, and a knee joint.

[0028] In this embodiment, the monitoring time period consists of the current moment and N-1 consecutive historical moments before it. N represents the total number of moments included in the monitoring time period, and its value can be an integer between 20 and 100. The time interval between adjacent moments can be 1 millisecond. For example, if N is 50, the corresponding monitoring time period is 50 milliseconds.

[0029] Step S102: Extract the velocity component of the linear velocity of the quadruped robot in the vertical dimension at each moment during the monitoring period to obtain the actual vertical velocity during the monitoring period.

[0030] For each moment within the monitoring period, the actual vertical velocity at each moment within the monitoring period is obtained by performing a dot product operation between the linear velocity vector and the unit vector in the range [0,0,1]. This reflects the actual physical motion feedback of the robot body in the vertical direction.

[0031] Step S103: Obtain the angle data of each leg of the quadruped robot at each time step and the rotation matrix of the quadruped robot relative to the world coordinate system.

[0032] The rotation matrix represents the coordinate transformation matrix from the robot's body coordinate system to the world coordinate system at each moment during the monitoring period. Its core function is to describe the robot's attitude (pitch, roll, yaw) in three-dimensional space.

[0033] Specifically, first, the robot's body pose angle vector at each moment is obtained. .in, This is the roll angle. The pitch angle, This refers to the heading angle. These values ​​are directly output by the airborne inertial navigation system (INS). Then, by reading the real-time attitude angles output by the INS, a 3×3 direction cosine matrix is ​​synthesized using trigonometric functions, which is used to transform the vectors in the fuselage coordinate system to the world coordinate system.

[0034] Specifically, the rotation matrix is ​​the fuselage attitude angle vector. The matrix mapping function. Using the standard ZYX rotation order, its calculation logic is as follows: ,in, This represents the rotation matrix at the k-th time point within the monitoring period. These are unit rotation matrices for rotations about the three axes of the world coordinate system. For example... Use of angle The rotation matrix is ​​obtained by filling a 3×3 matrix with the cosine and sine values, which is a well-known technique to those skilled in the art and will not be elaborated further here. It should be understood that each moment in the monitoring period corresponds to a rotation matrix.

[0035] Step S104: Solve the Jacobian matrix based on the angle, and use the Jacobian matrix and the rotation matrix to map the angular velocity of the quadruped robot during the monitoring period to obtain the theoretical vertical velocity during the monitoring period.

[0036] It should be understood that each moment in the monitoring period corresponds to a Jacobian matrix and a theoretical vertical velocity. Here, we will take any moment as an example for explanation.

[0037] The first step is to obtain the angle of each leg of the quadruped robot at each joint point and construct the angle vector of each leg. ,in, These correspond to the angles of the hip joint, forward hip joint, and knee joint, respectively. The angular velocity of each leg of the quadruped robot at each joint is obtained to construct the angular velocity vector for each leg.

[0038] The second step is to solve for the Jacobian matrix corresponding to each leg based on the angle vector of each leg.

[0039] Specifically, the position coordinate vector of the quadruped robot's foot at the k-th time step is constructed. The position coordinate vector refers to the three-dimensional coordinates of the robot's foot in a coordinate system with its center of mass as the origin.

[0040] Construct a nonlinear function between the position coordinate vector and the angle vector of each leg. Then, partial derivative operations are performed to obtain the Jacobian matrix. Defined as position coordinate vector For angle vector Partial derivatives: Here, k represents the k-th time point within the monitoring period. At each time point, the corresponding Jacobian matrix can be obtained using the same method described above. It should be understood that the process of solving the nonlinear function and the Jacobian matrix is ​​derived through the robot's forward kinematics. It can be solved based on the DH parameter method or the forward kinematic equations derived from the combination of link geometry length and joint angle trigonometric functions. This is a well-known technique in the field and will not be elaborated further here.

[0041] The third step involves determining the theoretical velocity vector of each leg based on the dot product between the rotation matrix, the Jacobian matrix of each leg, and the angular velocity vector. The vertical component of the theoretical velocity vector of each leg is then extracted to obtain the theoretical vertical velocity of each leg of the quadruped robot at any given moment.

[0042] As a concrete example, taking the data of any one leg of a quadruped robot at any time, the method for obtaining the theoretical velocity vector of the i-th leg of the quadruped robot at the k-th time can be expressed by the formula: in, This represents the theoretical velocity vector of the i-th leg of the quadruped robot at time k. This represents the rotation matrix at time k. This represents the Jacobian matrix of the i-th leg of a quadruped robot at time k. This represents the angle vector of the i-th leg of the quadruped robot at the k-th time. Let represent the angular velocity vector of the i-th leg of the quadruped robot at the k-th moment.

[0043] The negative sign in the formula indicates the relative relationship between the direction of motion of the quadruped robot's body and the direction of the thrust at the foot. The component of the theoretical velocity vector in the Z-axis direction is extracted as the corresponding theoretical vertical velocity. This value represents the vertical component generated by the joint motion of the quadruped robot's body under ideal rigid ground conditions.

[0044] Step S105: The absolute value of the difference between the actual vertical velocity and the theoretical vertical velocity at the same moment within the monitoring period is taken as the vertical velocity difference value at each moment within the monitoring period.

[0045] On soft ground, the absolute vertical speed of the robot is actually the result of the combination of the lifting caused by joint drive and the settlement caused by ground deformation. The actual vertical speed reflects the vertical speed data of the quadruped robot during actual movement, while the theoretical vertical speed reflects the theoretical vertical speed estimated by the quadruped robot through a kinematic model based on the angle data of various dimensions. Therefore, the difference between the two speeds can directly reflect the intensity of unexpected settlement movement.

[0046] More specifically, for each moment within the monitoring period, the absolute value of the difference between the actual vertical velocity and the theoretical vertical velocity at the same moment is calculated for each leg as the vertical velocity difference value of the quadruped robot at each moment. It should be noted that in this embodiment, the calculation method for the feature data corresponding to each leg is the same, so this embodiment only uses the data of one leg of the quadruped robot for introduction, and will not be elaborated further.

[0047] Step S200: Determine the vertical constraint force index within the monitoring period based on the parameter variables in the vertical velocity constraint of the quadruped robot for model predictive control.

[0048] The main purpose of this step is to evaluate the force intensity exerted by the quadruped robot at each moment during the monitoring period, based on the parameter variables of the vertical velocity constraint of the model predictive control (MPC), reflecting the control strength of the quadruped robot in trying to maintain the blocking height.

[0049] Specifically, during the monitoring period, the absolute values ​​of the Lagrange multipliers obtained in the model predictive control solution process at the previous moment are used as the vertical constraint force index at the current moment.

[0050] As a concrete example, taking any leg of a quadruped robot at the k-th moment within the monitoring period, we obtain the absolute value of the Lagrange multiplier corresponding to the vertical velocity constraint output by the MPC solver at the (k-1)-th moment as the vertical constraint force index at the k-th moment. In the optimization problem of MPC, each supporting leg corresponds to a vertical velocity constraint condition, that is, the vertical velocity is equal to 0. When the solver satisfies the vertical velocity constraint, it generates the corresponding Lagrange multiplier. That is, the Lagrange multiplier is the core dual variable of MPC in solving constrained quadratic programming (QP) problems, reflecting the increment of the MPC cost function caused by each violation of the constraint condition per unit length.

[0051] In the hard ground scenario, there is no actual settlement at the foot, and the zero velocity constraint is naturally satisfied. At this time, the MPC does not need to exert additional force to maintain the constraint, the Lagrange multiplier approaches 0, and the vertical constraint force index approaches 0, indicating that the controller has no intention to apply additional force.

[0052] In a soft ground scenario, the foot experiences vertical settlement due to ground subsidence, violating the zero-velocity constraint. To bring the foot back to rest, the MPC attempts to output a larger joint torque to pull the foot back. Violating the constraint at this time will cause the cost function to increase significantly, the Lagrange multiplier to increase sharply, and the vertical constraint force index to increase accordingly, indicating that the controller's intention to apply force is extremely strong.

[0053] It should be understood that the MPC solution process is essentially solving a constrained quadratic programming (QP) optimization problem at each time step. By transforming the internal control behavior of the MPC controller into quantifiable and analyzable time-series data, it is easier to perform correlation analysis with the physical response of foot settlement (vertical velocity difference value), thereby identifying force-velocity coupled oscillations. The vertical constraint force index characterizes the quantitative feature of the MPC controller's force application intention to maintain the vertical zero velocity constraint of the foot; its magnitude reflects the strength of the controller's desire to correct foot settlement.

[0054] It should be noted that, in order to avoid high-frequency noise affecting the data acquisition results, the absolute values ​​of the Lagrange multipliers in the model predictive control solution process at the previous time step are low-pass filtered before obtaining the vertical constraint force index. This data processing method is a well-known technique and will not be described in detail here.

[0055] Step S300: Based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index during the monitoring period, and combined with the balance of the vertical constraint force index, the deviation suppression weight during the monitoring period is obtained.

[0056] The main purpose of this step is to generate dynamic deviation suppression weights adapted to the ground conditions. By correlating vertical motion deviation with the controller's force application characteristics and combining the distribution characteristics of motion deviation, it provides precise parameter support for subsequent compliant control, enabling adaptive switching between hard ground rigidity and soft ground compliance. In complex terrain, the closeness of the correlation between the controller's force application intention and the foot settlement response, as well as the uniformity of the settlement deviation distribution, directly reflects the ground stiffness matching state. When the fluctuation trends of both are consistent but the deviation distribution is uneven, it often indicates a risk of oscillation.

[0057] Therefore, this step captures the degree of correlation between the vertical velocity difference value and the vertical constraint force index based on the fluctuation synergy characteristics within the monitoring period; then, combined with the balanced distribution of the vertical velocity difference value, the correlation characteristics are adapted and transformed to obtain the deviation suppression weight that can dynamically match the ground condition, ensuring that the weight adjustment not only fits the actual working conditions, but also accurately avoids oscillations or maintains control accuracy.

[0058] As a concrete example, such as Figure 3 As shown, the method for obtaining the deviation suppression weight within the monitoring period can be implemented by steps S301 and S302.

[0059] Step S301: Based on the consistency between the vertical velocity difference value at each moment during the monitoring period and the fluctuation trend of the vertical constraint force index at each moment, obtain the positive feedback correlation coefficient of the monitoring period.

[0060] This invention identifies the oscillation behavior of a quadruped robot by utilizing the time-weighted correlation between force and velocity. Compared to oscillation judgment methods that rely solely on amplitude or fixed thresholds, this method effectively distinguishes between systematic oscillations and random disturbances, improving the accuracy of oscillation identification. However, considering that conventional Pearson correlation coefficient calculations assign equal weight to all historical data, the sensitivity of the correlation index to current state changes is easily underestimated, leading to identification lag. To enable the control system to quickly perceive and respond to newly emerging oscillation trends, this embodiment of the invention introduces a time-weighted mechanism in the correlation calculation process.

[0061] The first step is to determine the weighted average velocity difference within the monitoring period based on the temporal characteristics of each moment and the corresponding vertical velocity difference value.

[0062] Specifically, the weighted average of the vertical velocity difference values ​​at each moment is calculated by taking the corresponding time sequence within the monitoring period as the weight. This yields the weighted average velocity difference within the monitoring period.

[0063] The second step is to determine the weighted average constraint force for the monitoring period based on the temporal characteristics of each moment within the monitoring period and the corresponding vertical constraint force index.

[0064] Specifically, the weighted average of the vertical constraint force index at each moment is calculated by taking the time sequence corresponding to each moment within the monitoring period as the weight, thus obtaining the weighted average constraint force for the monitoring period.

[0065] More specifically, the weight of each moment can be obtained by performing a nonlinear mapping on the time series corresponding to each moment within the monitoring period, and then performing weighted averaging. It should be noted that the methods for obtaining the weights corresponding to the weighted average velocity difference and the weighted average constraint force are the same, that is, one weight corresponds to one moment, which is used to perform weighted averaging on different data separately.

[0066] As a concrete example, taking any moment within the monitoring period as an illustration, the method for obtaining the weight corresponding to the k-th moment can be expressed as: in, This represents the weight corresponding to the k-th time point within the monitoring period. The presupposed forgetting factor is stated. This represents the time sequence corresponding to the k-th moment within the monitoring period, which is also the time sequence number. Let α represent an exponential function with the natural constant e as its base. In conventional quadruped robot control, the empirical range of α is [0.05, 0.2]. In this embodiment, the forgetting factor can be 0.1, which is an empirical value obtained through experiments with a large amount of data.

[0067] The exponential form of time series data is used for nonlinear mapping to amplify recent data and suppress older data. Control systems require extremely high real-time response to ground subsidence; the e-exponential model generates a nonlinear "nearer-larger, farther-smaller" effect, ensuring that abrupt changes in recent periods dominate the covariance and preventing dilution by historical stationary data. The forgetting factor α determines the steepness of the weight decay curve, i.e., the rate of forgetting historical data.

[0068] The third step involves calculating the covariance based on the vertical velocity difference and weighted average velocity difference at each moment within the monitoring period, as well as the vertical constraint force index and weighted average constraint force at each moment, and determining the positive feedback correlation coefficient for the monitoring period.

[0069] It should be noted that the positive feedback correlation coefficient is essentially the covariance between two sets of data: the vertical velocity difference value and the vertical constraint force index, within the monitoring period. The calculation method is a well-known technique. The difference from the existing calculation method is that the mean value involved in the calculation is replaced with the weighted mean value of the velocity difference and the weighted mean value of the constraint force obtained by weighted calculation in this embodiment. Therefore, the calculation method is a well-known technique and will not be described in detail here.

[0070] It should be further explained that the sign of the covariance determines the nature of the feedback. The positive feedback correlation coefficient only has physical meaning when the covariance is positive, indicating positive feedback, that is, the force applied leads to an increase in the deviation. When the covariance calculation result is less than 0, the corresponding positive feedback correlation coefficient is directly set to 0 to truncate the negative calculation result and discard the case where the increased force leads to a decrease in the deviation.

[0071] The larger the positive feedback correlation coefficient during the monitoring period, the more significant the positive feedback characteristic of "the stronger the force applied, the faster the settlement" is in the historical data, indicating that the system is in a state of stiffness mismatch oscillation; the smaller the value, the more unrelated the two are, and the system is in a state of normal control or random disturbance.

[0072] It should be noted that when the robot is stationary or making minor adjustments on a hard surface, its feet do not actually settle. At this time, the vertical velocity difference value is mainly composed of measurement noise from the IMU and encoder. Although the amplitude of this noise is small, it can still mathematically result in a high correlation coefficient. This means that the noise from the two sets of data—the vertical velocity difference value and the vertical constraint force index—may accidentally synchronize, leading to unnecessary fluctuations in control stiffness. To filter out this spurious signal, the positive feedback correlation coefficient is set to 0 when the weighted average velocity difference is less than a preset silent velocity threshold. It should be understood that when the weighted average velocity difference is greater than or equal to the preset silent velocity threshold, there is a significant movement or settling trend; the positive feedback correlation coefficient for the monitoring period can then be obtained using the calculation method described above.

[0073] The silent velocity threshold is set to 0.02 m / s. This value is an empirical value obtained through a large number of experiments. When the weighted velocity difference is less than 0.02 m / s and the direction of the cooperative and actual vertical velocities is upward, it indicates that the robot is currently stationary or in a state of slight movement on a hard surface, or that there is no downward settlement phenomenon. In this case, the correlation coefficient calculated by the above method is unreliable. Therefore, the positive feedback correlation coefficient is set to zero. This ensures that some potential correlation features are ignored within the low dynamic range, and high stiffness control is maintained at all times to ensure the stability of the robot's standing position.

[0074] Step S302: Based on the balanced distribution of vertical velocity differences within the monitoring period, and combined with the positive feedback correlation coefficient, a linear mapping is performed to obtain the deviation suppression weight within the monitoring period.

[0075] Specifically, the product of the normalized coefficient of the weighted average velocity difference and the normalized coefficient of the positive feedback correlation coefficient during the monitoring period is negatively correlated to determine the bias suppression weight during the monitoring period.

[0076] As a concrete example, the method for obtaining the bias suppression weight can be expressed by the formula: in, This indicates the deviation suppression weight for the monitoring time period corresponding to the current moment. This represents the positive feedback correlation coefficient for the monitoring time period corresponding to the current moment. This represents the weighted average speed difference over the monitoring period corresponding to the current moment. This indicates the preset benchmark rigidity weight. This represents the preset sensitivity gain factor. This represents the preset normalized velocity constant. The preset constant coefficient represents the empirical peak value of the covariance calculated under normal robot walking conditions, used to correlate the correlation coefficient. Normalize.

[0077] It should be noted that the baseline rigidity weight reflects the cost of hard ground baseline rigidity. When the denominator is 1, the quadruped robot has no oscillation, and the deviation suppression weight is equal to... This ensures that the legs do not buckle when walking on hard surfaces. The value of this benchmark rigidity weight can be determined by jointly calibrating the motor's peak torque and the maximum tolerable steady-state position error. In this embodiment, an empirical value is taken as [value missing]. .

[0078] The sensitivity gain factor is used to adjust the system's sensitivity to oscillations, for example, a value of 50. The larger the value, the faster the weight decreases under the same correlation coefficient, and the more aggressively the system enters a compliant state.

[0079] The normalized velocity constant is used to map settlement velocities of different magnitudes to a dimensionless range, avoiding the impact of differences in the physical magnitudes of velocity amplitudes on the consistency of weight adjustment; it is calibrated according to the maximum safe settlement velocity allowed by the robot design. For a conventional medium-sized quadruped robot, the value ranges from 0.1 m / s to 0.3 m / s, and in this embodiment, it can be taken as 0.1 m / s.

[0080] It should be understood that the aforementioned preset parameters are all empirical values ​​derived from numerous experiments. In this embodiment, the product of the two is negatively correlated by taking their reciprocals to correct the logical relationship between the deviation suppression weight and the positive feedback correlation coefficient and the average weighted velocity difference.

[0081] Furthermore, the positive feedback correlation coefficient only characterizes the degree of correlation between data and cannot reflect the amplitude of data fluctuations. The velocity mean is superimposed during weighted feature analysis to ensure that the weights are only significantly reduced when the features are highly correlated and the actual settling is rapid, preventing minor correlation noise from causing the robot to suddenly kneel.

[0082] The magnitude of the deviation suppression weight directly determines the tolerance of the MPC controller to zero constraint on the vertical velocity of the foot, i.e. the stiffness and compliance of the controller. Furthermore, the continuous change of the deviation suppression weight enables a smooth switching of control stiffness, avoiding body shaking caused by sudden changes in stiffness.

[0083] When there is no ground oscillation, the positive feedback correlation coefficient is 0, and the oscillation comprehensive quantification term... When the denominator approaches 1, the bias suppression weight equals 0. The MPC cost function imposes a strong penalty on slack variables, and the solver will force them to approach 0. The vertical velocity of the foot is strictly limited to 0, and the controller is in a high-stiffness mode to ensure high-precision trajectory tracking and attitude balance on hard ground.

[0084] The smaller the values ​​of the weighted velocity difference mean and the positive feedback correlation coefficient, the smaller the oscillation comprehensive quantitative term. The smaller the value, the larger the bias suppression weight, which means that the MPC's penalty for slack variables is extremely weak, allowing the foot to sink significantly. The controller is in a highly compliant mode, actively giving up rigid resistance to the ground, fundamentally blocking the positive feedback loop of force application and sinking, and eliminating severe vibrations of the fuselage.

[0085] The larger the values ​​of the weighted velocity difference mean and the positive feedback correlation coefficient, the greater the oscillation comprehensive quantitative term. The larger the value, the smaller the bias suppression weight. At this time, the penalty of MPC on the slack variable is moderately reduced, allowing the foot to sink slightly. The controller is in a mildly compliant mode to alleviate weak oscillations and avoid attitude drift caused by excessive sinking.

[0086] The quantization process of deviation suppression weights involves quantifying the abstract oscillation state into specific weight parameters, which can be directly applied to the soft constraint penalty term of the MPC cost function. This eliminates oscillations on soft ground and avoids fuselage attitude jitter caused by sudden changes in stiffness, while also taking into account the control requirements of high stiffness on hard ground and high compliance on soft ground.

[0087] It should be noted that all calculation processes in this step are the same for each leg of the quadruped robot. Therefore, this embodiment uses the parameter data corresponding to any one leg as an example for explanation. That is, following the same calculation method as steps S100 and S300, the data of each leg of the quadruped robot in the monitoring time period corresponding to the current moment is substituted into the calculation process to obtain the characteristic parameters corresponding to each leg. In other words, an independent judgment and analysis process is performed for each leg of the quadruped robot, and then a centralized global optimization problem is constructed in step S400.

[0088] Step S400: The cost function of the solver of the model predictive control algorithm is weighted using the deviation suppression weight to output the optimal control torque for controlling the quadruped robot.

[0089] It should be noted that the core of model predictive control algorithms is solving a global optimization problem, that is, finding the control input that minimizes the cost function while satisfying constraints. Therefore, it is first necessary to define the solution vector, i.e., the optimization vector, in the solution process. ,in This represents the vector of decision variables to be optimized by the solver within the monitoring time period corresponding to the current moment, where This represents the joint torque control command data to be solved. This data is an n×1 column vector, where n is the total number of joints. It is the transpose of the vector; Let be the allowable velocity deviation to be solved, a non-negative variable representing the allowable deviation of the vertical velocity at the foot from 0.

[0090] As a concrete example, the solver cost function of the model predictive control algorithm can be expressed by the formula: in, This represents the cost function for the monitoring period corresponding to the current time t. The data represents the joint torque control commands to be solved. The dimension is The positive definite diagonal weight matrix represents the penalty weight for the control torque output (energy consumption). This is the standard quadratic energy cost term, used to encourage the solver to use as little torque as possible while satisfying constraints. This parameter corresponds to the known control input minimization term of MPC.

[0091] This is the actual state vector of the quadruped robot at the current moment, specifically a vector composed of the quadruped robot's position, attitude angles (pitch angle, roll angle, yaw angle), linear velocity, and angular velocity at the current moment. The desired reference state vector is generated by receiving the desired horizontal linear velocity and desired yaw rate from the external control handle, combined with the preset desired fuselage ground clearance (e.g., 0.45m) and desired horizontal attitude (e.g., roll and pitch angles are both 0), through kinematic integration, generating the desired six-degree-of-freedom attitude and desired velocity of the fuselage in the world coordinate system at the current moment, and combining them into the reference state vector. , as the given tracking target of the MPC controller.

[0092] It reflects the tracking error vector between the actual state and the desired state. This is the positive definite diagonal weight matrix for state tracking error, representing the penalty weights for fuselage position and attitude deviations. This is the trajectory tracking cost term, used to ensure the fuselage maintains a stable attitude and desired altitude. It corresponds to the state error feedback term in the well-known MPC (Multi-Process Control) model.

[0093] This represents the deviation suppression weight for the monitoring time period corresponding to the i-th leg at the current moment. represents the vertical velocity relaxation of the i-th leg of the quadruped robot, which is a scalar and indicates the allowable slip or settlement of the leg end that is permitted to violate the "zero-velocity rigid contact" constraint; m represents the number of legs of the quadruped robot, reflecting the number of legs in the supported state at the current moment.

[0094] The compliance control cost term, which is the core component of this invention's embodiment using deviation suppression weights to weight the cost function, represents the weighted sum of penalties for all supporting legs violating the rigid contact assumption. When a leg vibrates on soft ground, its corresponding... Minimal, the solver will tolerate larger values In exchange for torque The smoothness eliminates stiffness mismatch oscillations in that leg; when a leg is on a hard surface, the corresponding... The maximum value will force the solver to suppress the maximum value. The value approaches 0 to restore the rigid support of the leg. That is, the third term of the cost function is a weighted sum of the penalty terms corresponding to all supporting legs, thereby realizing a differentiated compliant control strategy for each leg in a single optimization problem.

[0095] Furthermore, regarding the constraints of solving the local optimization problem, the original equality constraints are transformed into inequality constraints. That is, the equality constraint that the vertical velocity at the foot is equal to 0 is transformed into a range that allows fluctuations. This range is determined by the relaxation amount of the vertical velocity to be solved. Dynamically limited.

[0096] For multi-legged systems, constraints are established individually for each supporting leg. Therefore, the original equality constraints need to be broken down into independent inequalities for each leg.

[0097] For the i-th leg in a supported state, the inequality constraints are set as follows: , ,in This represents the actual vertical velocity of the i-th leg of the quadruped robot. Let represent the relaxation factor of the vertical velocity of the i-th leg of the quadruped robot. This represents the physical limit value of the maximum vertical settling speed allowed by the quadruped robot mechanism. This value is a preset output value based on the robot hardware specifications or experimental calibration.

[0098] This inequality constraint transforms the original MPC requirement that the vertical velocity at the foot of the supporting leg be strictly zero (rigid contact assumption) into a constraint centered at zero with a width of... The dynamically symmetric interval. When solving, It is also an unknown quantity that needs to be optimized. When the i-th leg experiences settlement oscillation, the bias suppression weight is extremely small, resulting in a very low penalty cost in the third term of the cost function. In this case, the QP solver will automatically select a larger value. This means that the foot is allowed to have a large vertical sliding speed.

[0099] This avoids the need to force the vertical velocity of each leg of the quadruped robot to be zero, which would otherwise generate a huge counteracting torque. This physically eliminates the violent rebound oscillations caused by force-velocity coupling. Conversely, if the first... With one leg on hard ground, a large deviation suppression weight will force the solver output... Approaching 0 converts the inequality into an equality, restoring rigid control.

[0100] Finally, the cost function and inequality constraints described above are input into an airborne QP solver (such as OSQP). The solver performs iterative calculations and outputs the torque control command results. This is the optimal control torque, which the system converts into current commands to drive the motor. If ground subsidence occurs at this point, the motor will output a smoother torque to accommodate the subsidence, thus eliminating vibrations.

[0101] It should be noted that after the solution is completed, the dual variable corresponding to the vertical velocity inequality constraint at the foot end, i.e., the Lagrange multiplier, is extracted from the solution result of the solver at the current time and can be used in the calculation process at the next time step.

[0102] In summary, this invention first analyzes the deviation between the actual and theoretical motion of a quadruped robot, obtaining data characteristics of foot settlement. Then, it quantifies the constraint force based on the dual variables output by the MCP solver. Furthermore, using time-weighted correlation analysis, it accurately identifies the positive feedback characteristics between force application and settlement in the time-series dimension, effectively distinguishing between systematic oscillations and random environmental noise. Further, it dynamically generates deviation suppression weights, achieving a nonlinear mapping from the identification results to control parameters. Finally, it reconstructs the optimization problem with slack variables within the MPC framework, sacrificing instantaneous velocity constraint accuracy for smooth torque output and possessing automatic recovery capability after oscillations.

[0103] This invention also provides a shock absorption and balance control system for a quadruped robot, including a memory, a processor, and a computer program stored in the memory and running on the processor. When the computer program is executed by the processor, it implements the steps of a shock absorption and balance control method for a quadruped robot. Since the method embodiments have already been described in detail, they will not be repeated here.

[0104] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A shock absorption and balance control method for a quadruped robot, characterized in that, The method includes the following steps: The linear velocity and angular velocity of the quadruped robot during the monitoring period are obtained. Based on the linear velocity and angular velocity, the difference between the actual vertical velocity and the theoretical vertical velocity of the quadruped robot are analyzed to determine the vertical velocity difference value during the monitoring period. The parameter variables in the vertical velocity constraint of model predictive control based on quadruped robots are used to determine the vertical constraint force index during the monitoring period. Based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index during the monitoring period, and combined with the balance of the vertical constraint force index, the deviation suppression weight during the monitoring period is obtained. By using deviation suppression weights to weight the solver cost function of the model predictive control algorithm, the optimal control torque is output for controlling the quadruped robot.

2. The vibration damping and balance control method for a quadruped robot according to claim 1, characterized in that, The deviation suppression weight for the monitoring period is obtained based on the deviation correlation between the vertical velocity difference value and the vertical constraint force index within the monitoring time period, combined with the balance degree of the vertical constraint force index. Specifically, this includes: Based on the consistency between the vertical velocity difference value at each moment within the monitoring period and the fluctuation trend of the vertical constraint force index at each moment, the positive feedback correlation coefficient of the monitoring period is obtained. Based on the balanced distribution of vertical velocity differences within the monitoring period, and combined with the positive feedback correlation coefficient, a linear mapping is performed to obtain the deviation suppression weight within the monitoring period.

3. The vibration damping and balance control method for a quadruped robot according to claim 2, characterized in that, The positive feedback correlation coefficient for the monitoring period is obtained by considering the consistency between the vertical velocity difference at each moment and the fluctuation trend of the vertical constraint force index at each moment within the monitoring period. Specifically, this includes: Based on the temporal characteristics and corresponding vertical velocity difference values ​​at each moment within the monitoring period, the weighted average velocity difference within the monitoring period is determined; based on the temporal characteristics and corresponding vertical constraint force indices at each moment within the monitoring period, the weighted average constraint force within the monitoring period is determined. Based on the vertical velocity difference and weighted average velocity difference at each moment within the monitoring period, as well as the vertical constraint force index and weighted average constraint force at each moment, the covariance is calculated to determine the positive feedback correlation coefficient for the monitoring period.

4. The shock absorption and balance control method for a quadruped robot according to claim 3, characterized in that, The step of obtaining the deviation suppression weight within the monitoring period by performing a linear mapping based on the balanced distribution of vertical velocity differences during the monitoring period and the positive feedback correlation coefficient specifically includes: The product of the normalized coefficient of the weighted mean velocity difference and the normalized coefficient of the positive feedback correlation coefficient during the monitoring period is negatively correlated to determine the bias suppression weight during the monitoring period.

5. The shock absorption and balance control method for a quadruped robot according to claim 3, characterized in that, The method for obtaining the weighted average velocity difference is as follows: using the time sequence corresponding to each moment within the monitoring period as the weight, the vertical velocity difference value at each moment is weighted and the average is calculated to obtain the weighted average velocity difference within the monitoring period. The method for obtaining the weighted average constraint force is as follows: using the time sequence corresponding to each moment within the monitoring period as the weight, the vertical constraint force index at each moment is weighted and the average is calculated to obtain the weighted average constraint force for the monitoring period.

6. The shock absorption and balance control method for a quadruped robot according to claim 1, characterized in that, The method involves analyzing the difference between the actual and theoretical vertical velocity of the quadruped robot based on linear velocity and angular velocity, respectively, to determine the vertical velocity difference value within the monitoring period. Specifically, this includes: Extract the vertical velocity component of the linear velocity of the quadruped robot at each moment during the monitoring period to obtain the actual vertical velocity during the monitoring period. Obtain the angle data of each leg of the quadruped robot at each time step and the rotation matrix of the quadruped robot relative to the world coordinate system; Based on the angle, the Jacobian matrix is ​​solved, and the angular velocity of the quadruped robot during the monitoring period is mapped using the Jacobian matrix and the rotation matrix to obtain the theoretical vertical velocity during the monitoring period. The absolute value of the difference between the actual vertical velocity and the theoretical vertical velocity at the same moment within the monitoring period is taken as the vertical velocity difference value at each moment within the monitoring period.

7. The shock absorption and balance control method for a quadruped robot according to claim 6, characterized in that, The process of solving the Jacobian matrix based on the angle, and using the Jacobian matrix and the rotation matrix to map the angular velocity of the quadruped robot during the monitoring period to obtain the theoretical vertical velocity during the monitoring period, specifically includes: For any given moment within the monitoring period; Obtain the angle of each leg of the quadruped robot at each joint point to construct the angle vector of each leg; obtain the angular velocity of each leg of the quadruped robot at each joint point to construct the angular velocity vector of each leg. The Jacobian matrix corresponding to each leg is solved based on the angle vector of each leg. The theoretical velocity vector of each leg is determined based on the dot product between the rotation matrix, the Jacobian matrix of each leg, and the angular velocity vector. The vertical component of the theoretical velocity vector of each leg is extracted to obtain the theoretical vertical velocity of each leg of the quadruped robot at any given time.

8. The vibration damping and balance control method for a quadruped robot according to claim 1, characterized in that, The parameter variables in the vertical velocity constraint of the model predictive control based on a quadruped robot, which determine the vertical constraint force index during the monitoring period, specifically include: During the monitoring period, the absolute values ​​of the Lagrange multipliers in the model predictive control solution process at the previous time point are obtained as the vertical constraint force index at the current time point.

9. A vibration damping and balance control method for a quadruped robot according to claim 3, characterized in that, The method further includes: When the weighted average velocity difference is less than the preset silent velocity threshold, the positive feedback correlation coefficient is set to 0.

10. A shock absorption and balance control system for a quadruped robot, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the computer program is executed by the processor, it implements the steps of the shock absorption and balance control method for a quadruped robot as described in any one of claims 1-9.

Images