A method and system for online compensation of joint clearance of a robot dog
By employing real-time monitoring and dynamic compensation mechanisms, the issues of motion accuracy and stability caused by joint gaps in the robot dog have been resolved, resulting in higher motion accuracy and stability.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-03-17
- Publication Date
- 2026-06-05
Smart Images
Figure CN122143008A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of robot dog joint control technology, and in particular to a method and system for online compensation of joint gaps in robot dogs. Background Technology
[0002] In the control system of modern robot dogs, precise joint control and motion compensation are key to improving the robot dog's motion performance and stability. The joints of a robot dog are usually composed of multiple drive units and load ends. These joints need to perform high-precision motion control in complex environments. However, due to the backlash and nonlinear characteristics of mechanical joints, phenomena such as backlash stagnation and rigid meshing may occur during the movement of the joints, which will affect the motion accuracy and stability of the robot dog.
[0003] Therefore, there is a need for an online joint space compensation method and system for robot dogs that can adjust joint motion compensation strategies in real time for different motion states of the robot dog. Summary of the Invention
[0004] This invention aims to provide a method and system for online compensation of joint clearance in robot dogs, which solves the problems of insufficient motion accuracy and poor stability in robot dogs under joint clearance and rigid meshing conditions. Through dynamic compensation and real-time load estimation, the joint control and motion compensation strategies are optimized.
[0005] A method for online compensation of joint gaps in a robot dog includes the following steps: The robot dog includes a joint drive end and a joint load end. While the robot dog maintains its motion at a preset speed, the system acquires the robot dog's joint dynamics and foot dynamics information. Based on the joint dynamics and foot dynamics information, the system performs dynamic topology decoupling on the robot dog to obtain several joint control units. The system monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation periods and joint rigid engagement periods. Based on the joint control units, the system estimates the load on the robot dog to obtain the equivalent load inertia of each joint control unit. When the real-time characteristic of joint motion is the joint space stagnation period, a virtual elastic force model of the joint is constructed based on the equivalent load inertia of the joint control unit; a joint dynamic compensation command is generated based on the joint virtual elastic force model; the robot dog motion compensation constraint is performed based on all joint dynamic compensation commands to obtain the robot dog joint online control command; the robot dog joint online control command is input into the corresponding joint control unit for control; when the real-time characteristic of joint motion is the joint rigid engagement period, the original motion state is maintained unchanged at the preset motion speed.
[0006] As a preferred embodiment of the present invention, the specific steps for load estimation of the robot dog based on the joint control unit include: The robot dog's contact Jacobian matrix is constructed based on joint dynamics information and foot dynamics information; for any joint control unit: the joint reflection inertia is obtained by dynamic projection along the motion transmission path of the current joint control unit according to the robot dog's contact Jacobian matrix; Obtain the inherent inertia of the joint motor of the current joint control unit; calculate the high-frequency dynamic change characteristics of the joint drive end and the joint load end based on the inherent inertia of the joint motor; generate a transmission state weighting factor with a value within a preset continuous range based on the high-frequency dynamic change characteristics and the real-time characteristics of joint motion; dynamically weight and fuse the joint reflection inertia and the inherent inertia of the joint motor through the transmission state weighting factor to obtain the equivalent load inertia of the joint control unit.
[0007] As a preferred embodiment of the present invention, the specific steps for constructing a virtual elastic force model of a joint based on the equivalent load inertia of each joint control unit include: Obtain the real-time phase displacement deviation and real-time phase velocity deviation of the joint drive end and the joint load end; calculate the relative penetration depth and remaining clearance margin of the current joint control unit during the joint clearance stagnation period based on the pre-calibrated clearance dead zone width; match the joint dynamic stiffness coefficient according to the remaining clearance margin. Extract the preset critical joint damping ratio; multiply the square root of the equivalent load inertia with the preset critical joint damping ratio to calculate the joint reference dissipation rate; construct a damping expansion operator with the remaining clearance margin as the denominator; multiply the joint reference dissipation rate with the damping expansion operator to obtain the dynamic damping coefficient of the current joint control unit. Multiply the real-time phase displacement deviation by the joint dynamic stiffness coefficient to obtain the joint displacement virtual elastic force term; multiply the real-time phase velocity deviation by the dynamic damping coefficient to obtain the joint velocity virtual damping force term; and superimpose the joint displacement virtual elastic force term and the joint velocity virtual damping force term to obtain the joint virtual elastic force model.
[0008] As a preferred embodiment of the present invention, the specific steps for generating joint dynamic compensation commands based on a joint virtual elasticity model include: The basic compensation torque of the joint is output based on the virtual elastic force model of the joint; the joint motor torque constant and joint electrical time constant corresponding to the current joint control unit are obtained based on the preset motion speed, and the basic compensation torque of the joint is linearly mapped to obtain the basic compensation current; The joint electrical time constant is used to perform first-order derivative feedforward compensation on the basic compensation current to obtain the predicted feedforward current; based on the remaining gap margin, a variable cutoff frequency smoothing filter is constructed with the remaining gap margin as the independent variable; wherein, the cutoff frequency of the variable cutoff frequency smoothing filter is dynamically reduced as the remaining gap margin decreases. The predicted feedforward current is input to a variable cutoff frequency smoothing filter for filtering, and the output current control signal serves as the final joint dynamic compensation command.
[0009] As a preferred embodiment of the present invention, the specific steps for performing motion compensation constraints on the robot dog based on all joint dynamic compensation commands include: Extract the joint motor torque constant corresponding to each joint control unit, and perform reverse conversion based on all joint dynamic compensation commands to obtain the equivalent joint compensation torque; Based on the transpose of the robot dog's contact Jacobian matrix, all equivalent joint compensation torques are dynamically mapped to the joint load end to obtain the expected additional contact force at the foot end; the robot dog's foot end support force at the current preset movement speed is obtained; the expected additional contact force at the foot end and the robot dog's foot end support force are superimposed in three dimensions to obtain the synthetic foot end contact force. A three-dimensional friction cone stable boundary is constructed at the joint load end; the combined foot contact force is judged to determine whether it escapes the robot dog cone stable boundary, and the comprehensive force judgment result is obtained. If the overall force judgment result does not escape, the joint dynamic compensation command will be output as the robot dog's online joint control command; if the overall force judgment result has a risk of escaping, the joint dynamic compensation command will be locally optimized until the overall force judgment result is satisfied.
[0010] As a preferred embodiment of the present invention, when the real-time joint motion characteristic of the joint control unit is the joint gap stagnation period, the transmission state weighting factor is used to smoothly decay the equivalent load inertia to the inherent inertia of the joint motor; when the real-time joint motion characteristic of the joint control unit is the joint rigid engagement period, the transmission state weighting factor is used to approximate the equivalent load inertia to the sum of the joint reflection inertia and the inherent inertia of the joint motor.
[0011] A robot dog joint gap online compensation system, comprising: The robot dog motion monitoring module includes a mechanical analysis unit; the mechanical analysis unit is used to monitor the robot dog, which has joint drive ends and joint load ends; when the robot dog maintains a motion state at a preset motion speed, it acquires the joint dynamics information and foot dynamics information of the robot dog; based on the joint dynamics information and foot dynamics information, it performs dynamic topology decoupling on the robot dog to obtain several joint control units; The robot dog motion compensation module includes a joint compensation unit and a comprehensive compensation unit. The joint compensation unit monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation and joint rigid engagement. Based on the joint control units, the robot dog's load is estimated to obtain the equivalent load inertia of each joint control unit. The comprehensive compensation unit constructs a virtual joint elasticity model based on the equivalent load inertia of the joint control units when the real-time joint motion characteristic is joint gap stagnation. It generates joint dynamic compensation commands based on the joint virtual elasticity model. Based on all joint dynamic compensation commands, the robot dog's motion compensation constraints are applied to obtain online joint control commands. The online joint control commands are input to the corresponding joint control units for control. When the real-time joint motion characteristic is joint rigid engagement, the original motion state is maintained at a preset motion speed.
[0012] The present invention has the following advantages: 1. This invention dynamically acquires joint dynamics information and foot dynamics information, enabling real-time monitoring of the robot dog's motion state in complex environments, thus providing a precise basis for subsequent compensation; by decoupling the robot dog's dynamic topology, the complex multi-joint system is decomposed into multiple independent joint control units, improving control accuracy and response speed, and making the system more stable under different load conditions; by distinguishing in real time whether the joint is in the gap stagnation period or the rigid engagement period, the compensation strategy can be dynamically adjusted, and different compensation measures can be taken for different motion states, thereby significantly improving motion accuracy and stability.
[0013] 2. This invention uses a load estimation method combined with a virtual joint elasticity model and dynamic damping coefficient to accurately calculate and compensate for the influence of joint clearance, avoiding the shortcomings of traditional methods in complex dynamic environments. Based on joint dynamic compensation commands, the invention performs motion compensation constraints on the robot dog, ensuring that the robot dog can maintain a stable motion trajectory under complex loads and external disturbances, reducing motion errors and instabilities caused by joint clearance. Attached Figure Description
[0014] Figure 1 This is a schematic diagram of the structure of an online joint clearance compensation system for a robot dog used in an embodiment of the present invention. Detailed Implementation
[0015] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of this invention.
[0016] Example 1: A method for online compensation of joint gaps in a robot dog, comprising the following steps: The robot dog includes a joint drive end and a joint load end. While the robot dog maintains its motion at a preset speed, the system acquires the robot dog's joint dynamics and foot dynamics information. Based on the joint dynamics and foot dynamics information, the system performs dynamic topology decoupling on the robot dog to obtain several joint control units. The system monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation periods and joint rigid engagement periods. Based on the joint control units, the system estimates the load on the robot dog to obtain the equivalent load inertia of each joint control unit. Specifically, regarding the definitions of the joint drive end and the joint load end, the joint drive end refers to the rotor output shaft of the motor inside the robot dog's joint, that is, the power source side before the reducer gear mechanism. It represents the active force-generating part of the joint and the object on which the control algorithm directly applies current. The joint load end refers to the mechanical linkage connected after the reducer gear mechanism, such as the mechanical structure of the robot dog's thigh or lower leg. It directly bears the weight of the robot dog and the reaction force of the terrain in the external environment. Since there are inevitably machining tolerances and tooth surface clearances inside the joint reducer, the actual physical positions of the joint drive end and the joint load end are not always strictly synchronized at the microscopic level. The relative displacement difference between the two when the force changes is the core physical phenomenon that the subsequent clearance compensation algorithm needs to deal with.
[0017] The specific steps for decoupling the robot dog's dynamic topology based on joint and foot dynamic information to obtain several joint control units are a low-level mathematical processing procedure that breaks down the robot dog's complex multibody system into independent single-joint control channels. Specifically, firstly, the system collects real-time state data such as the angles and angular velocities of each joint of the robot dog, as well as three-dimensional contact force data transmitted back by the foot sensors, and establishes a whole-machine dynamic mathematical model including a global mass matrix, centrifugal force, and Coriolis force matrices. Subsequently, using the contact Jacobian matrix describing the mapping relationship between the foot and joint positions, the ground reaction force felt by the foot is inversely and equivalently transformed into... The load torques acting on each joint axis; matrix block and diagonal projection calculations are performed on the global mass matrix of the robot dog, and the motion coupling force of adjacent links and the cross inertia of other joints are regarded as external independent disturbance terms of the current joint. Thus, the originally intertwined dynamic equations of the whole machine are forcibly separated into independent dynamic equations for a single specific joint (such as the left front knee joint); each separated independent dynamic equation and its corresponding drive control loop constitute an independent joint control unit, enabling the system to eliminate the interference of the overall attitude change and perform highly targeted micro clearance compensation calculations for a single joint that is prone to gear collision.
[0018] To clearly illustrate the specific calculation process of the above-mentioned dynamic topology decoupling, this embodiment adopts the disturbance observation compensation method based on inverse dynamics feedforward. Specifically, the system first constructs a complete dynamic model of the whole machine according to the Lagrange equation. In the decoupling step, the system extracts the independent inertia components corresponding to the current joint on the diagonal of the whole machine mass matrix as the nominal physical inertia of the joint control unit. At the same time, the off-diagonal elements (representing inertial coupling between joints), Coriolis force and centrifugal force vector, gravity vector, and Jacobian mapping components from foot contact force in the mass matrix are all uniformly summarized and merged into a total disturbance observation term. In each control cycle, the value of the total disturbance observation term is estimated using the sensor feedback data of the previous moment, and it is added as a feedforward compensation quantity to the torque balance equation of a single joint. In this way, the originally complex multivariable coupling equation is reconstructed into a decoupled single-input single-output system in the form of nominal inertia multiplied by angular acceleration equal to driving torque minus total disturbance torque, thus making it possible to calculate the clearance compensation for a single joint.
[0019] The joint gap stagnation period indicates that the transmission gears inside the reducer are in a microscopic idling state where they are out of contact. In this state, when the motor at the joint drive end reverses or accelerates, the motor rotor moves within the dead zone of the physical tooth gap. The motor's power cannot be effectively transmitted to the joint load end, causing the robot dog's leg link to experience a brief positional pause or slippage due to external force. This period is also the dangerous zone where the gears are most likely to violently impact at the end. The joint rigid engagement period indicates that the mechanical tooth gap has been completely crossed, and the transmission tooth surfaces at the joint drive end and the joint load end are in a tightly pressed rigid contact state. In this state, the motor's output torque can be directly transmitted to the leg link to support the ground without delay. The system as a whole returns to an ideal rigid body dynamics transmission structure. At this time, the controller only needs to maintain normal motion commands without injecting additional anti-collision compensation torque.
[0020] Real-time joint motion characteristics are used to characterize whether the robot dog's joint control unit is currently in a deep joint gap stagnation period, has transitioned to the joint rigid engagement period, or is in a micro-elastic critical contact state between the two. Mathematically, this characteristic is not a black-and-white discrete logic switch state, but exists as a kinematic state vector composed of multi-dimensional continuous physical quantities. This kinematic state vector mainly contains two core data components: the real-time relative phase displacement deviation and the real-time relative angular velocity deviation between the joint drive end (motor rotor) and the joint load end (mechanical linkage). These two continuous variables can accurately quantify the specific penetration depth of the motor rotor inside the dead zone of the reducer's mechanical cavity and the relative motion trend of the impending collision.
[0021] The specific steps for load estimation of the robot dog based on the joint control unit include: The robot dog's contact Jacobian matrix is constructed based on joint dynamics information and foot dynamics information; for any joint control unit: the joint reflection inertia is obtained by dynamic projection along the motion transmission path of the current joint control unit according to the robot dog's contact Jacobian matrix; Obtain the inherent inertia of the joint motor of the current joint control unit; calculate the high-frequency dynamic change characteristics of the joint drive end and the joint load end based on the inherent inertia of the joint motor; generate a transmission state weighting factor with a value within a preset continuous range based on the high-frequency dynamic change characteristics and the real-time characteristics of joint motion; dynamically weight and fuse the joint reflection inertia and the inherent inertia of the joint motor through the transmission state weighting factor to obtain the equivalent load inertia of the joint control unit. When the real-time joint motion characteristic of the joint control unit is the joint gap stagnation period, the transmission state weighting factor is used to smoothly decay the equivalent load inertia to the inherent inertia of the joint motor; when the real-time joint motion characteristic of the joint control unit is the joint rigid engagement period, the transmission state weighting factor is used to approximate the equivalent load inertia to the sum of the joint reflection inertia and the inherent inertia of the joint motor.
[0022] The rotation angles of each joint and the grounding status of the foot sensors are read in real time. Using the forward kinematics formula of spatial mechanics, the robot dog's contact Jacobian matrix, representing the linear mapping relationship between joint rotation speed and foot three-dimensional spatial linear velocity, is derived. Essentially, the robot dog's contact Jacobian matrix is a mathematical bridge connecting the robot dog's internal joint movements with its interaction with the external environment. It accurately describes how the ground reaction force is transmitted in reverse along the mechanical structure from the lower leg, thigh, to the torso when the leg steps on the ground to form a closed-loop physical constraint. Subsequently, for any independently defined joint control unit... The algorithm extracts the global rigid body mass matrix of the entire robot dog and combines it with the contact Jacobian matrix and its transpose constructed above. It then performs mathematical projection calculations on the link mass matrix of the entire leg and the equivalent mass of adhesion caused by the foot contacting the ground, along the mechanical motion transmission path of the currently controlled joint. The core of the projection calculation is to solve for the total inertial drag imposed by the entire leg and even the environment that needs to be overcome if the joint rotates slightly under the current system posture. This allows the extraction of the diagonal element components of the matrix in the specific motion direction, which are the joint reflected inertia describing the magnitude of the external load.
[0023] The inherent inertia of a joint motor refers to the pure physical rotational inertia of the rotor of the drive motor inside the joint. It is a static constant parameter that is determined at the factory calibration and represents the magnitude of the inherent physical inertia of the motor when it rotates without any external mechanical load connection. In actual operation, due to the frequent disengagement and re-impact of the reducer gears within the gap, this abrupt change in mechanical state causes the force between the drive end (motor rotor) and the load end (leg linkage) to change drastically in a very short time. To capture this phenomenon, the real-time angular acceleration of the motor rotor is read using high-frequency sampling and multiplied by the inherent inertia of the joint motor to calculate the reference torque theoretically required for the rotor to maintain its current motion. Then, the actual electromagnetic control torque output by the motor driver in real time is differentially compared with this reference torque. By setting up high-frequency feature extraction algorithms such as bandpass filters, the low-frequency smooth force part representing the normal walking gait is specifically filtered out. Thus, the sharp peak fluctuation data caused by the instantaneous collision or disengagement of the gears crossing the gap is accurately isolated from the difference signal. This peak fluctuation data constitutes the high-frequency abrupt change feature of the power characteristic of transmission discontinuity.
[0024] The high-frequency abrupt change characteristics of the power and the real-time characteristics of joint motion, such as the current relative speed and position of the joint, are input into a preset smooth nonlinear mapping function. The above discrete and violent mechanical state signals are transformed into a transmission state weight factor whose value is strictly limited to continuously changing between zero and one. The transmission state weight factor no longer adopts the traditional black-and-white logic judgment, but is used to continuously characterize the physical transition probability of the current reducer gear from a completely idling dead zone disengaged state, to a small elastic deformation of the tooth surface, and finally to a rigid transmission state of complete contact and compression.
[0025] Using the transmission state weighting factor as a harmonic ratio, the previously calculated joint reflected inertia representing the force on the entire leg and the joint motor's inherent inertia representing only the rotor itself are weighted and summed. Specifically, when the transmission state weighting factor approaches zero, indicating a state of complete backlash idling, the transmission of external load inertia is cut off in the mathematical model, so that the final output equivalent load inertia is smoothly reduced to only the joint motor's inherent inertia. When the transmission state weighting factor approaches one as the gears gradually mesh, indicating a state of complete rigid transmission, the algorithm drives the equivalent load inertia to gradually approach and equal to the sum of the joint motor's inherent inertia and the joint reflected inertia. Through the above dynamic weighted fusion method, the system can provide the controller with an equivalent load inertia value that transitions smoothly inside and outside the gear backlash, fundamentally preventing the control algorithm from sending overshoot compensation current to the motor due to instantaneous changes in inertia parameters, which would cause severe oscillations.
[0026] In the specific algorithmic steps of incorporating real-time joint motion characteristics into the function to solve for the transmission state weight factor, the system constructs a nonlinear boundary mapping function with smooth saturation characteristics at the control layer, such as a continuous activation model based on a hyperbolic tangent function or an S-curve. First, the real-time relative phase displacement deviation in the state vector is extracted and normalized by difference with the pre-calibrated maximum mechanical clearance dead zone width value in the system memory. This is used to calculate the remaining clearance margin between the current motor rotor and the actual rigid gear collision. This remaining clearance margin is used as the basic independent variable of the mathematical model, while the high-frequency abrupt change characteristics of the power are used as the dynamic adjustment gain term or feedforward bias term of the function, and substituted into the nonlinear boundary mapping function. The mapping function is used for joint solution. When the calculation shows that the remaining clearance margin is sufficient and the high-frequency abrupt change in dynamic characteristics is extremely small, the input term of the mapping function falls in the central flat region, so that the algebraic solution is strictly suppressed and smoothly converges to zero. Thus, the system is completely in the joint clearance stagnation period through mathematical expression. However, when the rotor continues to move, causing the relative phase displacement deviation to continuously approach the physical dead zone width boundary, and accompanied by the high-frequency abrupt change in dynamic characteristics showing the trend of micro-extrusion stress on the tooth surface, the input term of the mapping function quickly cuts into the steep stretching region of the curve, driving the mathematical solution to show a nonlinear exponential climb and finally smoothly saturate to a value of one. In this way, the full weight factor characterizing that the system has entered the joint rigid meshing period is continuously and without signal jumps is derived.
[0027] Regarding the method for obtaining the pre-calibrated dead zone width, this embodiment does not rely solely on theoretical gear drawing data, but rather uses a specific offline experimental calibration process to obtain the actual physical value. The specific calibration method is as follows: During the debugging phase before the robot dog leaves the factory or during each power-on self-test phase, the system controls the joint drive motor to output a small sinusoidal scanning torque, while using a high-precision encoder to record the difference between the angular displacement of the motor rotor and the angular displacement of the joint load end (mechanical linkage). When the motor rotor rotates but the load end has not yet produced a corresponding movement, this rotation range is determined to be within the dead zone. The system records the total angular difference between the motor's forward drive to the contact boundary and its reverse drive to the contact boundary, takes the average value of multiple measurements as the specific dead zone width parameter of the joint, and stores it in the controller's non-volatile memory for subsequent real-time retrieval.
[0028] To specifically implement the aforementioned nonlinear boundary mapping function, this embodiment employs a continuous saturation curve model based on the characteristics of the hyperbolic tangent function (tanh) or the sigmoid function. The specific calculation logic is as follows: the transmission state weight factor is calculated based on the functional relationship between the preset gap sensitivity coefficient, the normalized remaining gap margin, and the high-frequency abrupt change characteristics of the power. In this functional relationship, the remaining gap margin serves as the main independent variable, used to determine the basic value of the weight factor; the high-frequency abrupt change characteristics of the power serve as a suppression or penalty term, used to adjust the steepness of the curve or the saturation rate. During calculation, the system first obtains the ratio of the remaining gap margin to the gap dead zone width, maps it to a symmetrical interval centered at zero, and then performs a nonlinear transformation in conjunction with the high-frequency abrupt change characteristics, so that the calculated transmission state weight factor has the following characteristics: when the remaining gap margin is large and the abrupt change characteristics are small, the factor approaches the saturation value (e.g., 1); while as the remaining gap margin decreases or the abrupt change characteristics increase, the factor rapidly decays according to a smooth nonlinear gradient and approaches zero, thereby achieving a seamless soft switch from rigid transmission logic to gap idling logic. It is important to note that although the system distinguishes between the gap stagnation period and the rigid engagement period in macroscopic logic, the continuous and smooth transition between the two stages is achieved through the aforementioned transmission state weighting factor in the specific execution of the underlying algorithm. When the weighting factor is 1, the system is fully in the compensation logic of the gap stagnation period, and outputs the compensation torque calculated by the virtual elastic force model in full. When the weighting factor is 0, the system is fully in the rigid engagement period, and the compensation torque is zero. When the weighting factor is between 0 and 1 (i.e., in the microscopic transition zone before contact), the system actually executes a weighted sum of the gap compensation strategy and the rigidity maintenance strategy. This design corrects the control discontinuity problem caused by state abrupt changes in traditional logic, enabling the robot dog to maintain the mathematical continuity of torque output at the moment of virtual and real contact.
[0029] When the real-time characteristic of joint motion is the joint space stagnation period, a virtual elastic force model of the joint is constructed based on the equivalent load inertia of the joint control unit. The specific steps for constructing the virtual elastic force model of the joint based on the equivalent load inertia of each joint control unit include: Obtain the real-time phase displacement deviation and real-time phase velocity deviation of the joint drive end and the joint load end; calculate the relative penetration depth and remaining clearance margin of the current joint control unit during the joint clearance stagnation period based on the pre-calibrated clearance dead zone width; match the joint dynamic stiffness coefficient according to the remaining clearance margin. Regarding the specific process of matching the joint dynamic stiffness coefficient based on the remaining clearance margin, this embodiment adopts a linear or piecewise linear function calculation logic based on the relative penetration depth, rather than a simple table lookup method. The specific matching process follows the following logic: the joint dynamic stiffness coefficient is obtained by multiplying the currently calculated relative penetration depth by a preset basic stiffness gain. The basic stiffness gain is a preset proportional coefficient (its unit is Newton-meter / radian) used to define the magnitude of the reaction torque generated per unit penetration depth. When the remaining clearance margin is positive (i.e., not in contact within the gap), the stiffness coefficient is matched to zero or a very small weak guiding stiffness. When the remaining clearance margin is exhausted and the calculation result shows that virtual penetration occurs, the dynamic stiffness coefficient increases linearly proportionally with the increase of penetration depth, thereby simulating the physical contact elastic characteristics that conform to Hooke's Law, ensuring that the virtual torque can accurately resist physical collisions.
[0030] The instantaneous rotation angle and angular velocity of the joint drive end (i.e., the motor rotor shaft) and the joint load end (i.e., the leg connecting rod) are synchronously acquired by a high-frequency encoder, and the real-time position misalignment and velocity difference are obtained by subtracting the two; the maximum range of physical backlash, i.e., the pre-calibrated dead zone width, is retrieved from the data obtained in advance at the factory or during the calibration stage.
[0031] Based on this, by dividing the current real-time phase displacement deviation by or comparing it with the dead zone width of the gap, the specific travel distance that the motor rotor has now penetrated into the dead zone of the mechanical cavity is calculated, i.e., the relative penetration depth. At the same time, by subtracting the penetration depth from the dead zone width, the remaining physical space available for the motor rotor to move before the next rigid gear impact is accurately calculated, i.e., the remaining gap margin. In order to prevent the motor from breaking the reducer gear due to excessive force when crossing the gap, the system constructs a mapping function that decays nonlinearly with spatial distance in the underlying control algorithm based on the calculated remaining gap margin. Based on this, a joint dynamic stiffness coefficient is dynamically matched and output. The joint dynamic stiffness coefficient is used to provide strong traction force when the gap margin is sufficient to enable the rotor to quickly cross the cavity, and to rapidly decay the force when the gap margin is about to be exhausted, thereby removing the driving source that leads to rigid impact from the mechanical source.
[0032] Extract the preset critical joint damping ratio; multiply the square root of the equivalent load inertia with the preset critical joint damping ratio to calculate the joint reference dissipation rate; construct a damping expansion operator with the remaining clearance margin as the denominator; multiply the joint reference dissipation rate with the damping expansion operator to obtain the dynamic damping coefficient of the current joint control unit. The preset critical joint damping ratio is a reference physical constant used to measure the system's absorption of oscillation energy and prevent back-and-forth overshooting and spasms. The equivalent load inertia, which includes changes in terrain and body state, is mathematically squared, and the square root result is multiplied by the critical joint damping ratio to calculate a stable braking reference capability that can cope with the current leg weight, namely the joint reference dissipation rate. Regarding the aforementioned preset critical damping ratio of the joint, this parameter is a dimensionless scalar used to define the desired dynamic response characteristics of the system. In this embodiment, in order to ensure that the robot dog's joints do not oscillate or produce excessive stickiness during the gap transition, the damping ratio is usually set to a fixed value range between 0.7 and 1.1 (for example, preferably set to 1.0, i.e., the critical damping state). Although the robot dog's thigh joint and lower leg joint have different physical inertia due to their different mechanical structures, at the control algorithm level, this preset critical damping ratio serves as a unified performance indicator and is preset based on the control system's consistent stability requirements for all joints. The system will then combine this fixed damping ratio with the real-time estimated equivalent load inertia and reference stiffness to calculate the specific physical damping coefficient, thereby automatically adapting to the physical differences of different joints.
[0033] However, in actual robot dog movement, constant damping cannot cope with the impending instantaneous impact; therefore, a damping expansion operator is introduced. The damping expansion operator uses the remaining clearance margin as the denominator in the division operation. When the remaining clearance margin approaches zero, the value of the damping expansion operator will increase exponentially. The joint reference dissipation rate, which represents the basic braking capability, is multiplied with the damping expansion operator in real time. The above calculation process makes the dynamic damping coefficient of the current joint control unit generated by the system exhibit asymmetrical spatial characteristics: when the motor rotor is wandering in the center of the clearance, the damping is minimal, but at the moment when the rotor is extremely close to the target meshing tooth surface and the remaining clearance margin is almost exhausted, the dynamic damping coefficient will explode, forming an extremely viscous virtual air cushion in the control space, controlling the relative movement of the motor rotor.
[0034] Multiply the real-time phase displacement deviation by the joint dynamic stiffness coefficient to obtain the joint displacement virtual elastic force term; multiply the real-time phase velocity deviation by the dynamic damping coefficient to obtain the joint velocity virtual damping force term; and superimpose the joint displacement virtual elastic force term and the joint velocity virtual damping force term to obtain the joint virtual elastic force model.
[0035] The real-time phase displacement deviation, representing the positional misalignment, is algebraically multiplied with the previously known joint dynamic stiffness coefficient, which exhibits distance-dependent decay characteristics. This product generates a virtual traction torque within the control algorithm, known as the joint displacement virtual elastic force term. This virtual elastic force term acts like a spring with automatically weakening tension, suspended between the motor rotor and the leg linkage, providing gentle yet rapid bridging power during the early stages of the gap stagnation. Simultaneously, the real-time phase velocity deviation, representing the speed at which the two approach each other, is extracted and multiplied with the dynamic damping coefficient, which produces an exponential braking effect at the end, to calculate the joint velocity virtual damping force term specifically used to dissipate excess kinetic energy.
[0036] The joint virtual elasticity model is a nonlinear dynamic control mechanism constructed in the underlying control algorithm through pure mathematical and physical formulas. It is not a real mechanical spring or shock absorber in the physical world, but a spatial force potential field simulated inside the computer microcontroller.
[0037] Generating joint dynamic compensation commands based on a joint virtual elasticity model; the specific steps for generating joint dynamic compensation commands based on a joint virtual elasticity model include: The basic compensation torque of the joint is output based on the virtual elastic force model of the joint; the joint motor torque constant and joint electrical time constant corresponding to the current joint control unit are obtained based on the preset motion speed, and the basic compensation torque of the joint is linearly mapped to obtain the basic compensation current; The algorithm extracts the required joint basic compensation torque for the current control cycle from the constructed virtual joint elasticity model in real time. The joint basic compensation torque represents the theoretical mechanical rotational force that must be applied to enable the motor rotor to achieve a soft landing without impact. Based on the preset movement speed of the robot dog, the algorithm addresses and matches the internal hardware parameter matrix to obtain two core physical attributes of the motor corresponding to the current joint control unit: the joint motor torque constant and the joint electrical time constant. The joint motor torque constant is an inherent proportional conversion physical constant that characterizes how many Newton-meters of mechanical torque the motor can generate for every ampere of current. The joint electrical time constant represents the inherent physical response delay time caused by the inductive impedance effect of the stator winding inside the motor, which prevents the current from being established instantaneously. After obtaining these physical parameters, the algorithm performs a linear mapping operation, that is, directly dividing the joint basic compensation torque representing the mechanical requirement by the joint motor torque constant, thereby accurately converting the abstract mechanical torque requirement into the ideal current value that the motor theoretically needs to pass through. This current value constitutes the basic compensation current, which serves as the reference input signal for further electrical high-frequency compensation and optimization.
[0038] Regarding the specific method for determining the dynamic damping coefficient in the aforementioned virtual elasticity model of the joint, the dynamic damping coefficient is calculated based on the joint critical damping ratio, the product of the system's equivalent stiffness and load inertia, and a damping expansion operator based on the reciprocal relationship. The damping expansion operator is calculated based on the ratio of the preset dead zone width to the current remaining clearance margin. Specifically, this calculation logic follows the characteristics of an inverse proportional function: the dead zone width is used as the numerator, and the current remaining clearance margin (usually superimposed with a very small zero-prevention constant) is used as the denominator. When the remaining clearance margin is sufficient, the operator value is small, and the calculated dynamic damping coefficient remains at a low level, allowing the motor to move rapidly. When the remaining clearance margin approaches zero, due to the decrease in the denominator, the operator value increases exponentially or geometrically, causing the calculated dynamic damping coefficient to rise rapidly. This creates a highly damped virtual buffer force at the boundary where the physical clearance is about to be exhausted, i.e., the virtual air cushion effect mentioned in the manual.
[0039] The joint electrical time constant is used to perform first-order derivative feedforward compensation on the basic compensation current to obtain the predicted feedforward current; based on the remaining gap margin, a variable cutoff frequency smoothing filter is constructed with the remaining gap margin as the independent variable; wherein, the cutoff frequency of the variable cutoff frequency smoothing filter is dynamically reduced as the remaining gap margin decreases. The specific steps for using the joint electrical time constant to perform first-order derivative feedforward compensation of the basic compensation current and construct a variable cutoff frequency smoothing filter are as follows: In the real physical world, due to the inductance of the motor stator coil, when the control system issues a drastically changing current command such as emergency braking, the actual current rise will lag significantly behind the command, causing the motor to brake too late and triggering gear collision. In order to completely offset this electrical physical delay, the first-order derivative mathematical operation of the basic compensation current obtained in the previous step is performed, that is, the difference between the basic compensation current of the current control cycle and the previous cycle is calculated and divided by the sampling time to obtain the instantaneous rate of change of the current.
[0040] The rate of change of the current is multiplied by the joint electrical time constant, which reflects the inherent characteristics of physical delay, to obtain a lead drive compensation component specifically used to overcome inductive resistance. This component is then added to the original basic compensation current to synthesize a predictive feedforward current with lead active response capability. However, in actual calculations, the lead compensation is prone to amplifying tiny step jitters in the signal, causing harsh electromagnetic excitation noise. To address this, a spatial distance dimension is further introduced into the robot dog's control loop. The remaining clearance margin between the motor rotor and the collision boundary is extracted as an independent variable, and a variable cutoff frequency smoothing filter is dynamically constructed in the underlying code. The filter is specially configured in a spatial linkage mode: when the remaining gap margin is large, the filter maintains a high cutoff frequency to allow the predicted feedforward current to pass through quickly and without loss; as the rotor approaches the tooth surface and the remaining gap margin continues to decrease, the algorithm linkage forces a reduction in the cutoff frequency of the filter, using the microscopic physical spatial distance to dynamically compress and isolate the high-frequency band of the electrical signal.
[0041] The predicted feedforward current is input to a variable cutoff frequency smoothing filter for filtering, and the output current control signal is used as the final joint dynamic compensation command. Specifically: Within each microsecond-level real-time control cycle, the predicted feedforward current superimposed with the lead-drive component is used as the original excitation waveform signal and input to a variable cutoff frequency smoothing filter whose cutoff frequency parameters have just been updated in the previous step. At the execution level of the underlying algorithm, this filtering process manifests as an iterative operation of first-order or multi-order discrete difference equations. The filter automatically calculates and increases the trust weight of historical smoothed current data based on the current cutoff frequency, which is dynamically lowered due to the narrowing gap, while drastically reducing the acceptance weight of the severe predicted feedforward current glitches input at the current instant. Through this digital filtering mechanism that dynamically tightens the control bandwidth with spatial distance, glitches are accurately identified and forcibly eliminated. Steep current spikes and high-frequency oscillations caused by damped nonlinear expansion or lead derivative amplification; the current signal processed by a variable cutoff frequency smoothing filter effectively retains the low-frequency drive components that enable the motor rotor to achieve smooth deceleration and soft landing, while accurately eliminating high-frequency oscillation waveforms that easily cause electromagnetic noise and abnormal winding heating in the motor; this signal processing generates a smooth and safe current control curve based on successfully offsetting the physical inductance delay of the motor; this smooth current control signal is directly output to the underlying driver of the joint motor as the final joint dynamic compensation command, precisely controlling the robot dog's joints to smoothly cross gear gaps without mechanical impact or electromagnetic noise.
[0042] Based on all joint dynamic compensation commands, the robot dog's motion compensation constraints are performed to obtain online joint control commands; these commands are then input to the corresponding joint control units for control; when the real-time joint motion characteristic is during the rigid engagement phase, the original motion state is maintained at a preset speed; the specific steps for performing motion compensation constraints based on all joint dynamic compensation commands include: Extract the joint motor torque constant corresponding to each joint control unit, and perform reverse conversion based on all joint dynamic compensation commands to obtain the equivalent joint compensation torque; The specific steps of extracting the joint motor torque constant corresponding to each joint control unit and performing reverse conversion are a verification prerequisite process for restoring the underlying electrical signals to physical and mechanical parameters. The system main control unit extracts all the previously filtered joint dynamic compensation commands from the underlying drivers of each joint. These commands are physically represented as the control current values to be injected into the motor. In parallel, the pre-calibrated joint motor torque constants of all actuators of the robot dog are extracted. These constants characterize the mechanical rotational torque that can be generated by each unit current in a specific motor winding. Basic algebraic multiplication is performed, that is, multiplying the current command value of each joint by the torque constant of its corresponding motor to complete the reverse conversion from the electrical domain to the force domain. Through this reverse conversion algorithm, the system successfully quantizes and restores the local current compensation signals, which were originally generated independently only to compensate for the microscopic mechanical gaps of each joint, into the real physical torque acting on the rotation axis of each joint, that is, the equivalent joint compensation torque. This provides an accurate mechanical data basis for subsequent evaluation of whether these local compensation actions will affect the overall motion posture of the machine.
[0043] Based on the transpose of the robot dog's contact Jacobian matrix, all equivalent joint compensation torques are dynamically mapped to the joint load end to obtain the expected additional contact force at the foot end; the robot dog's foot end support force at the current preset movement speed is obtained; the expected additional contact force at the foot end and the robot dog's foot end support force are superimposed in three dimensions to obtain the synthetic foot end contact force. The contact Jacobian matrix, previously constructed during the dynamic topology decoupling phase, is invoked and its mathematical transpose is performed. In robot statics and virtual work principles, the transpose of the Jacobian matrix is a core mathematical tool that accurately projects the torque vector within the multi-degree-of-freedom joint space into a linear force acting on the end effector in a three-dimensional Cartesian coordinate system. The equivalent joint compensation torques of all joints are combined into a multidimensional torque vector, which is then multiplied on the left by the transpose of the Jacobian matrix. Through matrix-vector multiplication, the calculation is performed to determine how these compensation torques, originally acting only within the joints, would be transmitted along the robotic leg skeleton to the foot. The additional spatial force generated at the contact point between the foot and the ground is the expected additional contact force of the foot. At the same time, the main control unit of the system obtains the nominal foot support force that the robot dog must originally exert to maintain standing or walking, based on the preset movement speed and overall gait planning that the robot dog is currently executing. In the three-dimensional spatial coordinate system, the original foot support force vector and the newly generated expected additional contact force vector of the foot are combined by real-time linear vector addition to calculate the final three-dimensional force vector that the foot will actually exert on the ground after considering the interference of the underlying gap compensation, that is, the synthesized foot contact force.
[0044] A three-dimensional friction cone stability boundary is constructed at the joint load end. The combined foot contact force is judged to determine whether it escapes the robot dog's cone stability boundary, and the comprehensive force judgment result is obtained. The joint load end here specifically refers to the foot contact end where the robot dog directly interacts with the complex terrain. First, the normal support pressure and the estimated static friction coefficient of the ground are obtained in real time through the foot sensor. According to the classical Coulomb's law of friction, the algorithm takes the foot contact point as the vertex, the direction of the normal support pressure as the central axis, and uses the static friction coefficient as the tangent of the cone angle to draw a cone-shaped physical force limit envelope in the computer virtual three-dimensional space. This envelope is the three-dimensional friction cone stability boundary. As long as any total force vector applied by the foot falls inside this cone, the ground can provide enough friction to prevent the foot from slipping; otherwise, relative sliding will occur.
[0045] Based on the combined foot contact force, its three-dimensional vector coordinates are compared with the geometric boundary equations of the cone using algebraic inequalities. By calculating the angle between the combined force vector and the normal axis, and determining whether this angle is greater than the critical half-cone angle of the friction cone, the system can determine in real time whether the additional force generated by the bottom joint to eliminate gaps will break the existing ground grip limit, thus obtaining a comprehensive force judgment result on whether the force state is safe.
[0046] If the overall force judgment result does not escape, the joint dynamic compensation command will be output as the robot dog's joint online control command. When the algorithm determines that the overall force judgment result does not escape, it means that even if each joint is superimposed with anti-collision clearance compensation torque, the overall force on the foot is still steadily within the friction cone safety zone. At this time, the robot dog can not only eliminate gear noise, but also prevent the foot from slipping. Therefore, the system directly allows the output and outputs the original joint dynamic compensation command as the final robot dog joint online control command to the underlying actuator.
[0047] If the comprehensive force judgment result has a risk of escape, the joint dynamic compensation command is locally optimized until the comprehensive force judgment result is not escaped. When the comprehensive force judgment result shows a risk of escape, it means that blindly issuing the current compensation current will directly cause the robot dog's feet to lose grip and fall. At this time, a local reduction mechanism based on optimization constraints is immediately activated, with the highest mathematical optimization goal being to re-converge the synthetic foot contact force to the inside of the three-dimensional friction cone stable boundary. Numerical optimization algorithms such as quadratic programming are used to correct the specific joint current command that causes the surge in tangential force. The system iterates the above Jacobian mapping and force verification process in a microsecond-level control cycle, forcibly suppressing out-of-bounds commands, until the recalculated synthetic foot force is completely retracted within the safe friction cone. This ensures that when the robot dog walks in complex terrain, it always sacrifices some local quiet performance to ensure the safety of the control command that prevents the whole machine from falling.
[0048] It should be noted that, for the local optimization and numerical optimization algorithms involved, this embodiment specifically adopts a constraint-based quadratic programming solution logic. The optimization objective function of this algorithm is calculated based on the Euclidean distance or weighted sum of squared deviations between the originally calculated joint dynamic compensation command and the optimized command to be solved, aiming to find a correction torque that minimizes the deviation from the original motion intention. At the same time, the constraints of this optimization process are jointly constructed based on the foot force vector obtained by mapping the robot dog's contact Jacobian matrix and the three-dimensional friction cone inequality boundary determined by the surface friction coefficient. In each control cycle, the solver minimizes the above objective function under the premise that the synthesized foot force vector is strictly within the friction cone linear inequality constraint range, thereby calculating the optimal joint torque correction value. If the original command does not cause the foot force to escape from the friction cone, the optimization result is the original command; if there is a risk of escape, the optimization result is the optimal projection value of the original command on the friction cone stability boundary.
[0049] Meanwhile, regarding the aforementioned local reduction mechanism based on optimization constraints, this embodiment specifically employs a minimum norm projection algorithm based on quadratic programming, rather than simple linear scaling. The specific execution logic is as follows: When the comprehensive force judgment result indicates that the original joint torque command will cause the foot contact force to escape from the friction cone boundary, the system constructs a quadratic programming problem. The optimization objective of this problem is to minimize the Euclidean distance between the corrected torque vector and the original torque vector (i.e., to find the solution closest to the original intention), while strictly constraining that the foot contact force must be located inside the three-dimensional friction cone. The solver iterates and calculates within microseconds, outputting an optimal corrected torque vector. Physically, this corrected vector automatically reduces the normal or tangential torque components that cause the friction force to exceed the boundary, while maintaining the original motion trend direction as much as possible. This achieves a mathematical projection truncation from the boundary-crossing state to the safe boundary, ensuring that the robot dog does not slip while eliminating gaps.
[0050] When the real-time characteristics of joint movement are determined to be in the rigid engagement period of the joint, at the micro-physical level, it means that the transmission gear inside the robot dog's reducer has completely crossed the physical cavity dead zone. The drive tooth surface of the motor rotor and the load tooth surface of the leg mechanical linkage are in close contact and mutually pressed together, forming an ideal rigid body structure that can transmit torque without delay. At the underlying control logic level, since the mechanical system has completely left the dangerous zone where gear collision occurs, the virtual elasticity model and various micro-compensation commands originally used to guide the collision avoidance soft landing will automatically stop intervening and smoothly return to zero. The control system then completely returns the absolute dominance of the motor torque to the upper-level main gait controller of the robot dog, so that the robot dog is no longer affected by the additional torque interference of the underlying gap compensation algorithm, but strictly follows the preset movement speed issued by the macro system, and smoothly and steadily maintains its original normal walking or standing gait cycle with an ideal rigid transmission posture.
[0051] In practical applications, due to the involvement of multiple physical fields such as mechanical kinematics, electromagnetics, and dynamics, the system inevitably faces the problem of inconsistent data units, orders of magnitude, and even dimensions collected by different sensors. To ensure the good state of the computation matrix and prevent floating-point overflow, the system adopts an international standard unit base normalization and dimensional homogeneous scaling mechanism to solve this problem. The specific implementation method is as follows: During the initialization phase, the system mandates that all core algorithm variables involved in multibody dynamics and impedance control calculations must be converted and locked to international standard base units. For example, regardless of whether the encoder's underlying feedback is pulse count or degree, it is uniformly scaled and converted to radians through a preset gear ratio and resolution coefficient; angular velocity is uniformly converted to radians per second; the ADC voltage or digital quantity of the foot force sensor is converted to Newtons; and torque is uniformly converted to Newton-meters. This step completely eliminates the dimensional chaos caused by heterogeneous data sources before entering the algorithm.
[0052] For proportional parameters used to continuously adjust the algorithm's strength, such as the transmission state weighting factor and damping expansion operator, since they should be purely mathematical weights without any physical dimensions, the algorithm employs extreme value normalization processing of physically sourced data when calculating these parameters. For example, when calculating the mapping function involving the remaining clearance margin, the system does not directly use the remaining millimeter or radian value as the independent variable. Instead, it divides the current absolute value of the remaining clearance by the absolute value of the factory-calibrated maximum clearance dead zone width. After dividing these two physical quantities with the same unit, their dimensions are automatically canceled out, thus generating a pure proportional value that is absolutely dimensionless and takes values in the interval [0, 1]. This allows it to be safely used as a mathematical exponent or multiplier in subsequent virtual damping and stiffness calculations.
[0053] When the algorithm reaches the critical point of transitioning from the torque domain to the current domain (i.e., generating the base compensation current from the joint base compensation torque), the system introduces a joint motor torque constant containing composite physical dimensions, with units of Newton-meters per ampere. The algorithm, by dividing the dividend (desired torque) by this constant, not only performs the division numerically but also achieves precise dimensional cancellation, naturally and correctly deriving the ampere-level current command. Furthermore, since the torque calculation result may contain huge peak values, and the effective numerical ranges of different physical quantities in the microcontroller's fixed-point or floating-point registers may differ by several orders of magnitude, this is addressed to prevent… To prevent matrix singularity or truncation errors during Jacobian matrix construction or quadratic programming solutions, the system incorporates diagonal preprocessing scaling at the algorithm's underlying layer. This method assigns specific scaling gains to matrix rows of different physical types. While ensuring that the mathematical relationships on both sides of the physical constraint equations remain unchanged, it flattens all data values involved in the calculation to a similar order of magnitude for high-speed solution. After obtaining the final current command result, it multiplies back by the inverse scaling factor to restore the actual physical command for execution. The above data processing process can be adapted and adjusted by those skilled in the art; the above only describes one feasible data processing method.
[0054] Example 2: An online joint gap compensation system for a robot dog, see [link / reference] Figure 1 As shown, it includes: The robot dog motion monitoring module includes a mechanical analysis unit; the mechanical analysis unit is used to monitor the robot dog, which has joint drive ends and joint load ends; when the robot dog maintains a motion state at a preset motion speed, it acquires the joint dynamics information and foot dynamics information of the robot dog; based on the joint dynamics information and foot dynamics information, it performs dynamic topology decoupling on the robot dog to obtain several joint control units; The robot dog motion compensation module includes a joint compensation unit and a comprehensive compensation unit. The joint compensation unit monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation and joint rigid engagement. Based on the joint control units, the robot dog's load is estimated to obtain the equivalent load inertia of each joint control unit. The comprehensive compensation unit constructs a virtual joint elasticity model based on the equivalent load inertia of the joint control units when the real-time joint motion characteristic is joint gap stagnation. It generates joint dynamic compensation commands based on the joint virtual elasticity model. Based on all joint dynamic compensation commands, the robot dog's motion compensation constraints are applied to obtain online joint control commands. The online joint control commands are input to the corresponding joint control units for control. When the real-time joint motion characteristic is joint rigid engagement, the original motion state is maintained at a preset motion speed.
[0055] It should be understood that those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims. Parts not described in detail in this specification are prior art known to those skilled in the art.
Claims
1. A method for online compensation of joint gaps in a robot dog, characterized in that, Includes the following steps: The robot dog includes a joint drive end and a joint load end. While the robot dog maintains its motion at a preset speed, the system acquires the robot dog's joint dynamics and foot dynamics information. Based on the joint dynamics and foot dynamics information, the system performs dynamic topology decoupling on the robot dog to obtain several joint control units. The system monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation periods and joint rigid engagement periods. Based on the joint control units, the system estimates the load on the robot dog to obtain the equivalent load inertia of each joint control unit. When the real-time characteristic of joint motion is the joint space stagnation period, a virtual elastic force model of the joint is constructed based on the equivalent load inertia of the joint control unit. Dynamic joint compensation commands are generated based on the virtual elasticity model of the joints; motion compensation constraints for the robot dog are performed based on all joint dynamic compensation commands to obtain online control commands for the robot dog joints. Input the online control commands for the robot dog's joints into the corresponding joint control unit for control; When the real-time characteristics of joint movement are in the rigid engagement phase, the original movement state is maintained at a preset movement speed.
2. The method for online compensation of joint clearance in a robot dog according to claim 1, characterized in that, The specific steps for load estimation of the robot dog based on the joint control unit include: The robot dog's contact Jacobian matrix is constructed based on joint dynamics information and foot dynamics information; for any joint control unit: the joint reflection inertia is obtained by dynamic projection along the motion transmission path of the current joint control unit according to the robot dog's contact Jacobian matrix; Obtain the inherent inertia of the joint motor of the current joint control unit; calculate the high-frequency dynamic change characteristics of the joint drive end and the joint load end based on the inherent inertia of the joint motor; generate a transmission state weighting factor with a value within a preset continuous range based on the high-frequency dynamic change characteristics and the real-time characteristics of joint motion; dynamically weight and fuse the joint reflection inertia and the inherent inertia of the joint motor through the transmission state weighting factor to obtain the equivalent load inertia of the joint control unit.
3. The method for online compensation of joint gaps in a robot dog according to claim 2, characterized in that, The specific steps for constructing a virtual elastic force model of a joint based on the equivalent load inertia of each joint control unit include: Obtain the real-time phase displacement deviation and real-time phase velocity deviation of the joint drive end and the joint load end; calculate the relative penetration depth and remaining clearance margin of the current joint control unit during the joint clearance stagnation period based on the pre-calibrated clearance dead zone width; match the joint dynamic stiffness coefficient according to the remaining clearance margin. Extract the preset critical joint damping ratio; multiply the square root of the equivalent load inertia with the preset critical joint damping ratio to calculate the joint reference dissipation rate; construct a damping expansion operator with the remaining clearance margin as the denominator; multiply the joint reference dissipation rate with the damping expansion operator to obtain the dynamic damping coefficient of the current joint control unit. Multiply the real-time phase displacement deviation by the joint dynamic stiffness coefficient to obtain the joint displacement virtual elastic force term; multiply the real-time phase velocity deviation by the dynamic damping coefficient to obtain the joint velocity virtual damping force term; and superimpose the joint displacement virtual elastic force term and the joint velocity virtual damping force term to obtain the joint virtual elastic force model.
4. The method for online compensation of joint gaps in a robot dog according to claim 3, characterized in that, The specific steps for generating joint dynamic compensation commands based on a joint virtual elastic model include: The basic compensation torque of the joint is output based on the virtual elastic force model of the joint; the joint motor torque constant and joint electrical time constant corresponding to the current joint control unit are obtained based on the preset motion speed, and the basic compensation torque of the joint is linearly mapped to obtain the basic compensation current; The joint electrical time constant is used to perform first-order derivative feedforward compensation on the basic compensation current to obtain the predicted feedforward current; based on the remaining gap margin, a variable cutoff frequency smoothing filter is constructed with the remaining gap margin as the independent variable; wherein, the cutoff frequency of the variable cutoff frequency smoothing filter is dynamically reduced as the remaining gap margin decreases. The predicted feedforward current is input to a variable cutoff frequency smoothing filter for filtering, and the output current control signal serves as the final joint dynamic compensation command.
5. The method for online compensation of joint gaps in a robot dog according to claim 4, characterized in that, The specific steps for implementing motion compensation constraints for the robot dog based on all joint dynamic compensation commands include: Extract the joint motor torque constant corresponding to each joint control unit, and perform reverse conversion based on all joint dynamic compensation commands to obtain the equivalent joint compensation torque; Based on the transpose of the robot dog's contact Jacobian matrix, all equivalent joint compensation torques are dynamically mapped to the joint load end to obtain the expected additional contact force at the foot end; the robot dog's foot end support force at the current preset movement speed is obtained; the expected additional contact force at the foot end and the robot dog's foot end support force are superimposed in three dimensions to obtain the synthetic foot end contact force. A three-dimensional friction cone stable boundary is constructed at the joint load end; the combined foot contact force is judged to determine whether it escapes the robot dog cone stable boundary, and the comprehensive force judgment result is obtained. If the overall force judgment result does not escape, the joint dynamic compensation command will be output as the robot dog's online joint control command; if the overall force judgment result has a risk of escaping, the joint dynamic compensation command will be locally optimized until the overall force judgment result is satisfied.
6. The method for online compensation of joint gaps in a robot dog according to claim 5, characterized in that, When the real-time joint motion characteristic of the joint control unit is the joint gap stagnation period, the transmission state weighting factor is used to smoothly decay the equivalent load inertia to the inherent inertia of the joint motor; when the real-time joint motion characteristic of the joint control unit is the joint rigid engagement period, the transmission state weighting factor is used to approximate the equivalent load inertia to the sum of the joint reflection inertia and the inherent inertia of the joint motor.
7. An online joint clearance compensation system for a robot dog, characterized in that, The system employs an online joint clearance compensation method for a robot dog according to any one of claims 1-6, comprising: The robot dog motion monitoring module includes a mechanical analysis unit; the mechanical analysis unit is used to monitor the robot dog, which has joint drive ends and joint load ends; when the robot dog maintains a motion state at a preset motion speed, it acquires the joint dynamics information and foot dynamics information of the robot dog; based on the joint dynamics information and foot dynamics information, it performs dynamic topology decoupling on the robot dog to obtain several joint control units; The robot dog motion compensation module includes a joint compensation unit and a comprehensive compensation unit. The joint compensation unit monitors the robot dog's joint motion state in real time to obtain real-time joint motion characteristics. The real-time joint motion characteristics of the robot dog's joint control units include joint gap stagnation and joint rigid engagement. Based on the joint control units, the robot dog's load is estimated to obtain the equivalent load inertia of each joint control unit. The comprehensive compensation unit constructs a virtual joint elasticity model based on the equivalent load inertia of the joint control units when the real-time joint motion characteristic is joint gap stagnation. It generates joint dynamic compensation commands based on the joint virtual elasticity model. Based on all joint dynamic compensation commands, the robot dog's motion compensation constraints are applied to obtain online joint control commands. The online joint control commands are input to the corresponding joint control units for control. When the real-time joint motion characteristic is joint rigid engagement, the original motion state is maintained at a preset motion speed.