A quadruped robot adaptive control method for uncertain load

Abstract

The application discloses a kind of four-legged robot adaptive control methods for uncertainty load, based on the motion state of four-legged robot, establish the dynamics system model, and further analyze the influence of load on system, establish the dynamics system model considering load disturbance;Based on the actual measurable motion state information design extended state observer, realize real-time estimation to load disturbance, and prove the fast stable tracking ability of the designed extended state observer in theory;Based on the dynamics system model and the obtained load disturbance estimation value design model predictive controller, and propose the corresponding optimization goal and constraint condition.The application can solve the stable motion control problem of large heavy hydraulic four-legged robot based on force servo under facing uncertainty load working condition.

Description

Technical Field

[0001] This invention belongs to the field of robot control technology, and in particular, it is an adaptive control method for quadruped robots facing uncertain loads. Background Technology

[0002] Heavy-duty hydraulic quadruped robots, a hot topic in biomimetic robotics research, are experiencing a surge in demand across numerous fields and scenarios due to their heavy-load capabilities, particularly in areas such as transporting supplies and injured personnel in the field. To enable robots to adapt to various complex and unstructured environments, legged robots are increasingly adopting force servo methods to design their underlying servo systems. Consequently, motion control systems have evolved to calculate the output forces of each joint based on dynamic models. Because of the introduction of dynamic models to design control laws, the motion control system requires accurate information on the robot's mass and inertia. However, in scenarios involving transporting supplies or injured personnel, the mass of the robot and its load is unknown and may even change in real-time for the motion control system. This can lead to calculated joint forces that are insufficient for stable robot movement under unknown load conditions. Therefore, the motion control system needs to possess the ability to adapt to uncertain loads.

[0003] Patent CN111070201A proposes a reactive robust control method for quadruped robots under sudden load changes based on ZMP theory. This patent can overcome the interference caused by sudden load changes to a quadruped robot's normal walking. However, it uses sliding mode control in the controller section, which is prone to introducing chattering during actual movement, and the control input obtained is not the energy-optimal solution compared to optimized methods. Patent CN116991083A proposes a control method and system for an intelligent agent to address the problem of robots easily misstepping or falling when facing risks such as sudden terrain changes or external impacts in complex environments. This patent uses an online training control strategy network, with local deployment and offline use of the strategy network. It has strict requirements for the online training environment; if the load in actual working conditions significantly exceeds the load variation range during online training, the control strategy cannot effectively cope. Therefore, in the prior art, the above methods still cannot effectively and robustly cope with load change disturbances. Summary of the Invention

[0004] To address the problem of stable motion control for heavy-duty hydraulic quadruped robots based on force servo under uncertain load conditions, this invention proposes an adaptive control method for quadruped robots under uncertain load conditions.

[0005] The technical solution to achieve the purpose of this invention is as follows:

[0006] An adaptive control method for a quadruped robot under uncertain loads is proposed. This method optimizes the objective and constraints, obtains the optimal plantar force control input using a solver, and then converts it into the desired torque for each joint using a force-Jacobi matrix, thereby achieving robot pose control.

[0007] The optimization objective is:

[0008] in, For the first System state reference value at time. and These are the weight matrices for the system state changes and the control inputs, respectively; f i Let be the three-dimensional force vector exerted by the ground on the i-th leg's foot, and let be the system state. ,in, , , , , , , These are the quadruped robot's attitude angles, center of mass displacement, angular velocity, center of mass linear velocity, and the additional force caused by the gravity of the load. The estimated additional torque caused by load gravity and offset distance. The estimation of gravitational acceleration.

[0009] The significant advantages of this invention compared to existing technologies are:

[0010] This invention proposes a control method combining high-frequency estimation and low-frequency optimization: a high-frequency response extended state observer enables rapid and stable tracking of the effects of uncertain load disturbances, while a low-frequency response model predictive control solves for the optimal foot force based on accurate load disturbance estimation. This method does not require prior knowledge of the load's privileged information or extensive trial-and-error training in simulations, yet still demonstrates good adaptability to unknown loads, exhibiting significant technical advantages and practical application value. Attached Figure Description

[0011] Figure 1 This is a schematic diagram illustrating the principle of an adaptive control method for a quadruped robot under uncertain loads.

[0012] Figure 2 This is a schematic diagram of the forces acting on a quadruped robot. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0014] The principle of the adaptive control method for quadruped robots under uncertain loads designed in this invention is as follows: Figure 1 As shown, it is divided into three parts: model predictive control, extended state observer, and robot dynamics system model. Therefore, the technical solution of this invention will be described in the order of dynamics modeling, extended state observer design, and model predictive controller design.

[0015] 1. Determine the motion state parameters of the quadruped robot and establish a complete dynamic system model of the quadruped robot. The specific modeling steps are as follows:

[0016] Firstly, for robot motion control systems, the primary focus is on the overall motion state of the robot, such as its pose and speed, rather than the motion state of individual components. Therefore, reference... Figure 2 The force diagram of the quadruped robot shown first establishes a center-of-mass coordinate system. The origin of the center-of-mass coordinate system coincides with the center of mass of the robot body. The X-axis coincides with the forward direction of the robot, the Y-axis points to the left side of the robot, and the Z-axis points vertically upward. The quadruped robot's attitude angles are then used as a guide. Displacement of the center of mass angular velocity and the linear velocity of the center of mass The dynamic model established for the system state is as follows

[0017]

[0018] in, To find the derivative with respect to the state variables, For time, vector ,in , and Let X be the roll angle of the fuselage about the X-axis, the pitch angle about the Y-axis, and the yaw angle about the Z-axis; vector. ,in , and Let X be the displacement of the robot's center of mass along the X, Y, and Z axes, respectively; vector. ,in , and The angular velocities of the fuselage around the X, Y, and Z axes are respectively; the velocity vector of the center of mass is... ,in , and Let X be the linear velocities of the robot's center of mass along the X, Y, and Z axes, respectively. It is a rotation matrix; To find the inverse of the robot's rotational inertia matrix, where, , and These are the moments of inertia of the robot about the X, Y, and Z axes, respectively. , , and Let be the three-dimensional displacement vectors pointing from the ends of the right foreleg, left foreleg, right hindleg, and left hindleg towards the robot's center of mass. Taking the three-dimensional vector of the right foreleg as an example, the antisymmetric form of the vector is: ,in, , and Three-dimensional vectors The magnitude of the projection values ​​on the X, Y, and Z axes; , , and These are the three-dimensional force vectors exerted by the ground on the right foreleg foot, left foreleg foot, right hind leg foot, and left hind leg foot, respectively. The vector of gravitational acceleration. The mass of the robot does not include the load; and These are a three-dimensional identity matrix and a three-dimensional all-zero matrix, respectively.

[0019] The above dynamic system model only considers the robot's own gravity. When the robot needs to carry a certain load during task execution, the load's center of mass cannot be exactly on the same axis as the robot's center of mass. Therefore, the load's center of mass is offset relative to the robot's center of mass. This offset causes the robot to be affected in two ways: ① the additional force caused by the gravity of the load. Impact, among which ① Uncertainty about load mass; ② Additional torque caused by load gravity and offset distance. Impact, among which Let be the distance vector from the center of mass of the uncertain load to the center of mass of the robot. When using model predictive control to obtain the optimal plantar force, the disturbance caused by the load must be considered; otherwise, the actual robot motion may be biased towards the load side or even become unstable, directly preventing the robot from completing its intended task. Since the load disturbance is not known to the motion control system, the dynamic model considering the uncertain load is as follows:

[0020]

[0021] 2. Based on directly measurable state information, an extended state observer is designed to observe load disturbances, obtain accurate observation values ​​of load disturbances, and theoretically prove the stability of the designed extended state observer closed-loop system. The specific design steps are as follows:

[0022] Firstly, for the extended state observer, the dynamic model system considering the uncertain load can be divided into a second-order system, and the corresponding system state can also be divided into fuselage pose vectors. and fuselage velocity vector ,in, , , and These are the estimated values ​​for the fuselage attitude angles, the center of mass displacement, the angular velocity, and the linear velocity of the center of mass, respectively.

[0023]

[0024] The above dynamic model can then be rewritten as follows:

[0025]

[0026] Among them, matrix ,matrix , To control the input vector, This is the load disturbance vector.

[0027] By combining the IMU sensors and joint sensors on the fuselage, the fuselage pose can be measured quickly and accurately. Therefore, using the measured fuselage pose information and the estimated fuselage pose value from the extended state observer as the error value input for the extended state observer, the extended state observer system can be designed as follows:

[0028]

[0029] in, To extend the state observer to the fuselage pose vector The estimated value, For the extended state observer to the fuselage velocity vector The estimated value, For the extended state observer to the load disturbance vector The estimated value, The measured fuselage position and attitude information.

[0030] Secondly, it is proven that the estimated value of the aforementioned extended state observer can converge quickly to the actual state value. For the aforementioned extended state observer, let the state of the dynamic system be... Then the above dynamic system model can be converted into a first-order state-space form.

[0031]

[0032] Among them, matrix ,matrix ,matrix ,matrix ,matrix , For a 6-dimensional matrix with all elements equal to zero, It is a 6-dimensional identity matrix. Given a 6×12 dimensional matrix with all elements equal to zero, the extended state observer can be similarly converted to...

[0033]

[0034] in, System status The estimated value of the derivative, System status The estimated value, For the tracking gain of the extended state observer, where, , and The gain value to be set for the state observer. Input the plantar force. and the measured actual fuselage position As a combined input to the extended state observer system Then the above formula can be transformed into

[0035]

[0036] To enable the designed extended state observer to quickly approximate the actual system state, its tracking gain needs to be designed so that the closed-loop system poles of the extended state observer are in the negative half-plane, and the farther the poles are from the imaginary axis, the faster the observer tracks. Assuming its poles are in... place, If the value is arbitrarily set to be positive, then the characteristic polynomial can be designed as follows:

[0037]

[0038] in, To represent the eigenvalues ​​of the matrix, The above equation, where is a three-dimensional identity matrix, is used to solve for the eigenvalues ​​of the state equation of the closed-loop system with the extended state observer. That is, when the tracking gain of the extended state observer... This ensures that the system state estimated by the designed extended state observer can quickly and stably approach the actual system state.

[0039] 3. An adaptive motion control method based on model predictive control is designed to solve for the optimal plantar force input. The specific design steps are as follows:

[0040] Model predictive control requires solving for the optimal foot force input using a dynamic system model, and it also requires prior knowledge of the load disturbance. In the previous step, an extended state observer was used to estimate the load disturbance. Therefore, the dynamic system model for model predictive control is...

[0041]

[0042] The aforementioned quadruped robot dynamics system model is complex and nonlinear. Model predictive control (MDC), based on this complex nonlinear dynamics system model and using a QP solver to solve for the optimal foot force control input, is prone to poor dynamic performance due to high computational complexity, ultimately affecting the overall motion performance of the robot. Therefore, there are two optimizations when using MDC: ① designing a solution step size independent of the motion control system to reduce the solution frequency; ② transforming the system model into a linear form and then using a simplified linear MDC method. Combining these two methods can achieve the fastest possible acquisition of the optimal foot force control input with limited computational resources. Therefore, the aforementioned nonlinear dynamics model is transformed into a linear form, i.e., the system state is... ,in, and The extended state observer is given the load perturbation vector. and If the observed values ​​are obtained, the above dynamic model can be transformed into

[0043]

[0044] Among them, matrix System status The first derivative with respect to time, matrix ,matrix .

[0045] Then, the linearized dynamic model is discretized.

[0046]

[0047] in, For positive integers, For the first The system state at any given moment. For the first The system state at any given moment. For the first Time-based control input, matrix ,matrix ,in It is an 18-dimensional identity matrix. Let be the single-step solution time of the model predictive controller, and let the prediction step size of the model predictive controller be . If we take one step, then we obtain the prediction model for all step lengths as follows:

[0048]

[0049] in, This is the initial state of the system. For the system in the first The state transition matrix of the step, the above prediction model can be simplified to

[0050]

[0051] Among them, matrix ,matrix ,matrix ,matrix .

[0052] After obtaining the discretized prediction model, before introducing the QP solver to solve for the optimal foot force control input, it is necessary to add optimization objectives and constraints. The optimization objectives are generally simply set as a smooth transition of the system state and minimum energy consumption; therefore, the optimization objective can be written as:

[0053]

[0054] in, For the first System state reference value at time. and These are the system state weight matrix and the control weight matrix, respectively. and The weights in the matrix are respectively the first and second weights. The weight matrix contains diagonal element values, each pre-defined by the user. The L2 norm is calculated using the following formula: There are two types of constraints: equality constraints and inequality constraints. Equality constraints are based on the dynamic model described above, while inequality constraints arise because the optimal foot force is not infinitely large and must satisfy certain practical physical constraints.

[0055]

[0056] in, Let be the coefficient of friction between the ground and the foot. Then it is the first leg strength The three components on the X, Y, and Z axes, and These represent the minimum and maximum values ​​of the foot force in the vertical direction, respectively.

[0057] After obtaining the optimization objective, equality constraints (prediction equations), and inequality constraints, the optimal plantar force control input can be obtained using a QP solver. Finally, the plantar force needs to be converted into the desired torque of each joint using a force-Jacobi matrix. This force-Jacobi matrix is ​​closely related to the robot's size and configuration, exhibiting unique characteristics. Furthermore, this invention focuses on a motion control method that can be combined with force-Jacobi matrices of different configurations; therefore, the calculation process of the force-Jacobi matrix will not be described in detail here.

Claims

1. An adaptive control method for a quadruped robot under uncertain loads, characterized in that, The optimal plantar force control input is obtained by optimizing the objective and constraints and using a solver. Then, it is converted into the desired torque of each joint through the force Jacobian matrix to achieve robot pose control. in The optimization objective is: in, For the first System state reference value at time. and These are the weight matrices for the system state changes and the control inputs, respectively; f i Let be the three-dimensional force vector exerted by the ground on the i-th leg's foot, and let be the system state. ,in, , , , , , , These are the quadruped robot's attitude angles, center of mass displacement, angular velocity, center of mass linear velocity, and the additional force caused by the gravity of the load. The estimated additional torque caused by load gravity and offset distance. The estimation of gravitational acceleration.

2. The adaptive control method for a quadruped robot according to claim 1, characterized in that, The constraints are equality constraints: matrix ,matrix ,matrix ,matrix ; in For the first The system state at any given moment. For the first The state transition matrix at time t, For the control input at time n-1, the matrix , To determine the single-step solution time of the prediction equation, the matrix... in, The inverse of the robot's rotational inertia matrix. For the first The antisymmetric form of the three-dimensional displacement vector of each leg pointing towards the robot's center of mass. These correspond to the tips of the right foreleg, left foreleg, right hindleg, and left hindleg, respectively.

3. The adaptive control method for a quadruped robot according to claim 2, characterized in that, The inverse of the robot's rotational inertia matrix is: in, , and These are the moments of inertia of the robot about the X, Y, and Z axes, respectively.

4. The adaptive control method for a quadruped robot according to claim 1, characterized in that, in: in For uncertain load quality, The vector of gravitational acceleration. The distance vector from the centroid of the uncertain load to the centroid of the robot. The antisymmetric form.

5. The adaptive control method for a quadruped robot according to claim 1, characterized in that, The constraints are inequality constraints: in Let be the coefficient of friction between the ground and the foot. Then it is the first leg strength The three components on the X, Y, and Z axes, and These represent the minimum and maximum values ​​of the foot force in the vertical direction, respectively.

Images