Quadruped robot model prediction control method fused with extended state observer
By introducing an extended state observer (ESO) into the model predictive control of a quadruped robot, unknown disturbances are estimated and compensated, solving the problem of performance degradation of existing methods in complex environments and achieving higher robustness and trajectory tracking accuracy.
Patent Information
- Application Number
- CN202511130485.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-13
- Publication Date
- 2025-11-18
- Estimated Expiration
- 2045-08-13
AI Technical Summary
Existing model predictive control methods for quadruped robots exhibit performance degradation when faced with external unknown disturbances and model uncertainties. In particular, the disturbance observer relies on the bounded assumption of the disturbance derivative, making it difficult to accurately estimate higher-order non-stationary disturbances, resulting in incomplete compensation and insufficient stability.
Extended State Observer (ESO) is used to estimate unknown disturbances in translation and rotation directions respectively, and the estimated values are fed forward into the discrete MPC framework. A state-space model is constructed through a dual ESO structure to enhance the system's disturbance rejection capability and trajectory tracking accuracy. Robust control is achieved by combining the MPC optimization problem.
It achieves precise compensation for unknown disturbances, improves the motion robustness and control performance of quadruped robots in uncertain dynamic environments, enhances trajectory tracking accuracy and stability, and demonstrates stronger anti-disturbance capabilities.
Smart Images

Figure CN120972550A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The application belongs to the technical field of intelligent robot control, and particularly relates to a four-legged robot model predictive control method fusing an extended state observer. BACKGROUND
[0002] Four-legged robots play an important role in tasks such as inspection, transportation and rescue in complex environments due to their superior terrain adaptability and dynamic stability. In recent years, many advanced four-legged robot systems such as ANYmal, MiniCheetah, Boston Dynamics Spot and Unitree Go1 have emerged, driving the development of legged robot technology. However, in practical applications, uncertain factors such as complex terrain, load changes and external disturbances pose higher requirements on the robot control system. How to achieve robust, accurate and real-time motion control has become a key problem in the research of four-legged robots.
[0003] To address the above challenges, researchers have developed a variety of control methods. Early methods such as control strategies based on zero moment point (ZMP) and spring-loaded inverted pendulum (SLIP) model achieve basic gait stability by simplifying the dynamics model, but are difficult to cope with rapidly changing dynamic environments. Virtual model control (VMC) uses virtual springs and damping elements to construct control forces, enabling flexible adjustment of robot posture and trajectory, and has certain terrain adaptability. Whole-body control (WBC) optimizes the overall joint torque distribution in task space, enabling coordinated control in high-redundancy systems, and is suitable for multi-task parallel execution. Reinforcement learning (RL) methods have certain environmental perception and action migration capabilities due to their policy self-learning ability, but have practical limitations such as large training data volume, low sample efficiency and weak interpretability, making them difficult to be widely applied to high-safety task scenarios.
[0004] Among the above methods, model predictive control (MPC) has become the mainstream framework for four-legged robot dynamic control in recent years due to its ability to solve optimal control sequences online, compatibility with system constraints and state prediction capabilities. MPC outputs desired ground reaction forces or joint torques by constructing a state space model and a constraint optimization problem, and has been successfully applied to robot platforms such as Mini-Cheetah and ANYmal, enabling robots to achieve high-performance motions such as dynamic jumping and somersaults. However, MPC is extremely sensitive to model accuracy, and system parameter errors or external disturbances can significantly affect its control accuracy and stability.
[0005] To enhance the robustness of MPC, some researches have tried to combine it with robust control, adaptive control and other strategies. For example, robust MPC enhances the stability of control within the known disturbance boundary, but often sacrifices system performance due to conservative design; adaptive MPC can improve the adaptability of the system through online parameter estimation, but it is not enough to adapt to the rapid change of disturbance. Disturbance observer (DOB) as a typical disturbance compensation mechanism is also introduced into MPC to improve the anti-disturbance ability of the system, but its estimation accuracy usually depends on the assumption that the derivative of disturbance is bounded, which is difficult to cope with rapid changes or high-order nonlinear disturbances.
[0006] Therefore, the application provides a quadruped robot model predictive control method fusing an extended state observer
[0007] In the application, the extended state observer (ESO) is a non-model-dependent observer that can estimate the system state and total disturbance in real time, and has the advantages of simple structure, high estimation accuracy and low computational cost. Integrating ESO into the MPC control framework is expected to achieve accurate compensation for unknown disturbances without increasing the burden of sensors, and to improve the motion robustness and control performance of quadruped robots in uncertain dynamic environments. Therefore, the ESO-based model predictive control method provides a new effective way to solve the bottleneck of existing MPC control in disturbance estimation and compensation. SUMMARY
[0008] The application aims to provide a quadruped robot model predictive control method fusing an extended state observer, which can solve the problems of motion control precision decline and insufficient disturbance resistance of quadruped robots in complex dynamic environments, especially the problem of control performance degradation of existing model predictive control (MPC) methods when facing external unknown disturbances and model uncertainties. The existing methods based on disturbance observer (DOB) often rely on the assumption that the derivative of disturbance is bounded, which makes it difficult to accurately estimate high-order non-stationary disturbances, resulting in incomplete compensation and insufficient stability.
[0009] The technical solutions adopted by the application are as follows:
[0010] According to the characteristics of the extended state observer (ESO) that can estimate the total disturbance without knowing the disturbance model, the application provides a quadruped robot model predictive control method in uncertain dynamic environments based on an extended state observer. The method uses a double-ESO structure to estimate unknown disturbances in the translation and rotation directions respectively, and feeds the estimated values into the discrete MPC framework to improve the anti-disturbance ability and trajectory tracking accuracy of the system. The method comprises the following steps:
[0011] The method establishes a single rigid body dynamics model of the quadruped robot with the center of mass as a reference, and models unknown disturbance forces and disturbance moments on the body as system disturbance terms to form a state space model with disturbances, thereby providing a basis for subsequent observer and controller design;
[0012] The extended state observer of the method designs a multi-order position ESO and an attitude ESO, respectively;
[0013] The position ESO is used to estimate external disturbance forces in the x, y and z directions of the robot;
[0014] The attitude ESO is used to estimate disturbance moments around the roll, pitch and yaw axes;
[0015] In the observer design, appropriate bandwidth parameters are selected to ensure that the disturbance estimation error is bounded, and the stability and convergence of the linear system method are analyzed;
[0016] The MPC state space of the method introduces the disturbance terms estimated by the ESO as compensation into the discrete state space model, thereby achieving robustness enhancement of the controller without increasing the burden of additional sensors.
[0017] The MPC optimization problem of the method is specifically a cost function defined to minimize the desired trajectory tracking error, combined with system operation safety and actuator physical constraint conditions, to construct the following optimization constraints: the ground reaction force satisfies the friction cone constraint; the swing leg contact force is zero; the foot force and joint torque are limited by the upper and lower limits of the system; and the state variables need to satisfy the system physical boundary conditions.
[0018] The MPC rolling optimization solution of the method is specifically a rolling horizon control strategy, which solves the finite time optimal control problem with disturbance compensation in each control period to obtain the desired foot force or joint torque, which is used as the output of the controller to act on the quadruped robot.
[0019] The simulation verification of the method is specifically a trajectory tracking experiment on the Unitree Go1 quadruped robot on the gazebo simulation platform of the Ubuntu 20.04 system, which verifies that, when carrying unmodeled disturbance loads, the method has a significant improvement in body height error, desired speed error and attitude fluctuation compared with the traditional DOB-MPC controller, and has stronger anti-disturbance ability and control stability.
[0020] The technical effects achieved by the method are:
[0021] The method of the application realizes real-time estimation of unknown disturbance force and torque by constructing a double ESO structure, and introduces it into an MPC controller, realizes high-precision trajectory tracking and system robustness enhancement. It can effectively overcome the problems of poor estimation accuracy and strong dependence on model of traditional disturbance observation methods, and has the advantages of simple structure, good real-time performance, strong anti-interference performance and the like.
[0022] The application can be widely applied to high-precision motion control tasks of various quadruped robot systems in uncertain environments such as complex terrain, load change, unknown disturbance, and is suitable for application scenes such as search and rescue, field inspection, intelligent transportation, and has good engineering practicability and popularization value. BRIEF DESCRIPTION OF DRAWINGS
[0023] Figure 1 is a model predictive control framework based on an extended state observer of the application;
[0024] Figure 2 is a walking parameter diagram of Go1 in a simulation environment carrying a 6kg payload;
[0025] Figure 3 is a turning parameter diagram of Go1 in a simulation environment carrying a 6kg payload. DETAILED DESCRIPTION
[0026] In order to make the purpose and advantages of the application clearer and more apparent, the application will be specifically described below in combination with embodiments. It should be understood that the following text is only used to describe one or several specific embodiments of the application, and does not strictly limit the specific protection scope requested by the application.
[0027] As shown in Figure 1 , a quadruped robot model predictive control method fusing an extended state observer, characterized in that it comprises the following steps:
[0028] Step 1: Construct a quadruped robot dynamics model, establish a single rigid body kinematics equation with the center of mass as the reference point, introduce a system disturbance term, and form a state space model containing disturbance;
[0029] The quadruped robot dynamics model essentially describes the dynamic coupling relationship between internal forces and motion state changes. In order to realize the expected motion trajectory of the robot, the dynamics balance between the driving joint torque and the foot contact force needs to be calculated. The application establishes a quantitative mapping between the foot contact force and the body motion parameters by constructing mathematical and physical equations, which provides core algorithm support for subsequent control algorithms.
[0030] In the application, the robot kinematics model modeling method of step 1 specifically comprises:
[0031] To describe the dynamic behavior of the whole quadruped robot under disturbance, the single rigid body kinematics modeling method with the center of mass as the reference point is adopted. This model ignores the flexibility of the leg joints and mainly depicts the six-degree-of-freedom rigid body dynamics behavior of the robot body, which has the advantages of simple modeling and easy controller design. Its expression form is as follows:
[0032]
[0033] wherein, is the mass of the robot; is the moment of inertia of the robot; is the gravity vector; is the base angular velocity of the robot; is the center of mass position; i = 1, 2, 3, 4 is the foot position; is the contact force of each leg;
[0034] Considering the influence of unknown disturbance on the robot during movement, the single rigid body dynamics model incorporating disturbance terms is established as follows
[0035]
[0036] wherein, represents the uncertainty of the robot in the translational direction, represents the uncertainty of the robot in the rotational direction; at this time, if the influence of ω b ×I G ω b is ignored, the above formula can be expanded as follows
[0037]
[0038] It is worth noting that this approximation method is widely used in the motion control of quadruped robots, and this approximation is considered reasonable, especially when the angular velocity of the object is small, ω b ×I G ω b is ignored;
[0039] To complete the design of the observer, the following assumptions are proposed:
[0040] Assumption 1. F b (t), d(t) and their time derivatives have upper bounds,
[0041] i.e. ‖F b (t)‖2≤γ1, ‖d(t)‖2≤γ2, and wherein, γ1, γ2, δ1 and δ2 are random normal numbers.
[0042] Step 2: Design the dual extended state observer (ESO), including:
[0043] Position ESO for real-time estimation of unknown disturbance forces in the translational directions of the robot;
[0044] Attitude ESO for real-time estimation of disturbance torques around the roll, pitch, and yaw axes;
[0045] Based on the ESO knowledge as described in Step S1, the steps of the quadruped robot extended state observer design can be divided into two steps: extended state observer design based on disturbance forces and extended state observer design based on disturbance torques. The specific design process is shown below:
[0046] A nonlinear control method using ESO is used to estimate the unknown disturbance forces F b and uncertain torques d acting on the robot. After estimating the two disturbance terms, they are integrated into the MPC framework to ensure the friction constraints of the ground reaction forces. Finally, the desired foot forces and the correction forces generated by the swing leg dynamics are optimized together to convert into joint output torques τ d .
[0047] Define an input output system of the form u (n) = Δ + x, where, is the unknown term of the system. The output u can be measured by sensors. Let u1 = u, u n = u (n) . Define u n+1 as the extended state variable, and u n+1 = Δ. The original controlled object is now described as
[0048] The purpose of the ESO is to estimate Δ using the measurable output u. Let be the estimated value of u1,..., u n+1 , then the ESO is designed as follows:
[0049]
[0050] where a1 = w0α1, α i = (n+1)! / [i! (n+1-i)!], i = 1, 2,..., n+1. Set ω0 as the observer bandwidth, and ω0 > 0. Let (i = 1,..., n+1) represent the estimation error. Combining it with equation (1.33), the error dynamics expression is
[0051]
[0052] From the stability analysis, if is bounded, i.e. γ o is a random constant, then there exists a constant r i > 0 and a finite time T1> 0 such that for all t≥ T1≥ 0 and ω0> 0.
[0053] Let the constant r i be set as follows:
[0054]
[0055] where k is a positive integer, h i is a constant. Therefore, from equation (1.35), the estimation error bound of ESO is determined by the initial estimation error and the observer bandwidth ω0; since the unknown term Δ cannot be measured, the initial estimation error and γ o cannot be used to reduce r i , while the gain ω0is inversely proportional to ; therefore, when the value of ω0in ESO is larger, a smaller r i value can be obtained.
[0056] Based on the above knowledge of ESO, the steps of designing the extended state observer of a quadruped robot can be divided into two steps: the design of an extended state observer based on disturbance force and the design of an extended state observer based on disturbance torque, and the specific design process is as follows:
[0057] Step 201: The design of the extended state observer based on disturbance force in step 2 specifically includes:
[0058] According to equation (1.2), we have:
[0059]
[0060] where is the control input term; is the disturbance force acting on the center of mass of the robot;
[0061] Design an n-order position ESO to estimate the unknown term Δ Let represent the estimates of p, v, f Δ,1 ,..., f Δ,n , respectively; then the n-order ESO is designed as:
[0062]
[0063] where They represent The i-th element; x v,i For x v The i-th element; These are adjustable parameters for the ESO; according to the ESO design and formula (1.3), the estimated error is bounded, i.e. From formula (1.3), we can see that r p,1 ,r p,2 ,r p,3 With ω p,1 ,ω p,2 ,ω p,3 It is inversely proportional to the k-th power;
[0064] At the same time, use a large gain vector ω p To obtain a smaller estimation error bound, we can use the position of ESO; at this point, let r p Indicates r p,1 ,r p,2 ,r p,3 Three times the maximum value, i.e., r p =3max{r p,1 ,r p,2 ,r p,3 If}, then there exists a limit.
[0065] The closed-loop dynamics of the position loop are divided into two sub-loops: position and velocity. This dual-sub-loop structure is matched to the cascaded structure of a quadruped robot. Let the desired position and position error of the base be represented respectively. Then the expected dynamics of the two sub-loops are as follows:
[0066]
[0067] in, It is a diagonal matrix; the position error is defined as s. v =v c -v r v c v is the actual speed. r The desired velocity reference trajectory; to achieve the desired position sub-loop, v r Designed as follows:
[0068]
[0069] v r and The relationship is shown in formula (1.24); at the same time, the velocity error s is obtained. v The relationship with position error is as follows:
[0070]
[0071] From equation (1.8), the desired position sub-loop can be implemented by r v → 0; To illustrate the relationship between s v and the stability of position dynamics, the following technical lemma 1, assumption 2 and lemma 2 are introduced:
[0072] Lemma 1: Let r1 be a constant, if ||s v ||≤ r1, t→∞, then t→∞, where, denotes the trace of ;
[0073] Assumption 2: If f Δ (n) is bounded, then is bounded, i.e.
[0074] Lemma 2: According to assumption 2, if the position control law is given by equation (1.9) and the position ESO is given by equation (1.5), then the position tracking error is bounded, i.e. t→∞, where, r p is the upper bound of the position ESO estimation error;
[0075]
[0076] Step 202: The design of the disturbance-based extended state observer of step 2 specifically includes:
[0077] According to equations (1.2) and (1.3), the dynamics model can be rewritten as:
[0078]
[0079] where, is the control input; is the disturbance torque;
[0080] An n-order attitude ESO is designed to estimate the unknown disturbance term Let , τ b , …, τ Δ,1 , respectively, represent the estimates of ω Δ,n ; then the n-order attitude ESO is designed as:
[0081]
[0082] where, x ω,i is the i-th element of x ω , and is the adjustable bandwidth vector of the attitude ESO, respectively and the i-th element of ; it is known from the above that the estimation error is bounded, i.e. Since r ω,1 , r ω,2 , r ω,3 are inversely proportional to the k-th power of ω ω,1 , ω ω,2 , ω ω,3 , therefore, the position ESO with larger gain vector obtains smaller estimation error bound; let r ω denote three times the maximum of r ω,1 , r ω,2 , r ω,3 , i.e., r ω = 3max{r ω,1 , r ω,2 , r ω,3}, then there exists a limit Let ω d denote the desired angular velocity of the quadruped robot, which can be obtained by the following formula:
[0083]
[0084] Define the error rotation matrix as Define the error quaternion where and
[0085]
[0086] Let be the derivative of the error rotation matrix, which can be obtained by the following formula:
[0087]
[0088] where [·] × denotes the cross product operation;
[0089] In order to realize the motion control of the quadruped robot, the closed-loop dynamics of the attitude loop is divided into two sub-loops of attitude and angular velocity; the design dynamics of the two sub-loops are given as follows:
[0090]
[0091] where and are positive diagonal matrices, k β > 0 is a constant gain; in addition, s ω = ω - ω r , ω r is the desired angular velocity reference trajectory;
[0092] To achieve the desired attitude sub-loop, ω r is designed as:
[0093]
[0094] Pushes out
[0095]
[0096] Then the desired attitude sub-loop is achieved as s ω → 0.
[0097] To achieve the desired angular velocity sub-loop, the quadruped robot base attitude controller is designed as:
[0098]
[0099] where,
[0100]
[0101] Substitute the control law (1.19) into equation (1.10) to get:
[0102]
[0103] According to is bounded; under the control law (1.19), the desired angular velocity sub-loop is achieved bounded, then the control torque vector T B is:
[0104]
[0105] Here it is worth mentioning that in order to illustrate the stability of the attitude loop; introduce the following assumptions 3, lemma 3;
[0106] Assumption 3: Assume is bounded, then we can get is bounded, that is
[0107] Lemma 3: According to assumption 3, if the attitude controller is designed as equation (1.21) and the attitude ESO is equation (1.11), then the attitude tracking error is bounded, that is t→∞, where, r q is the upper bound of the attitude tracking error.
[0108] Step 3: Embed the disturbance estimation output by the ESO as a feedforward compensation item into the constructed discrete-time state-space model to form a model predictive control (MPC) system containing disturbance compensation;
[0109] The step considers the single rigid body dynamics model and the design of the extended state observer described above, and for the convenience of the controller design, the application further constructs a state space expression form of the quadruped robot system, which is used for the subsequent optimization modeling of the model predictive controller.
[0110] The disturbance compensation containing model predictive control system of step 3 specifically comprises:
[0111] According to formula (1.4) and formula (1.16), the continuous time state space model of the quadruped robot system is expressed as:
[0112]
[0113] The discrete time state space model thereof can be shown as:
[0114]
[0115] Wherein, the variables of formula (1.23) are defined as follows:
[0116] State variables are:
[0117]
[0118] Wherein, is the position of the robot centroid; is the attitude angle; such as Euler angle; is the linear velocity of the centroid; is the angular velocity of the body; is the disturbance force and disturbance torque estimated by the position ESO and the attitude ESO;
[0119] Input control represents the expected ground reaction force of the quadruped robot, and the expression is:
[0120]
[0121] Wherein, represents the expected ground reaction force of the nth leg;
[0122] Gravity term vector represents the gravity acceleration term, and the expression is:
[0123]
[0124] ESO estimation term vector The ESO estimation value containing the disturbance term and its derivative are used for feedforward compensation, and the expression is:
[0125]
[0126] Transition matrix The basic dynamics update of the system, including position-velocity integration, attitude update, etc., is expressed as:
[0127]
[0128] Input matrix The expression is:
[0129]
[0130] Where m is the mass of the robot, I G is the body inertia matrix, [p i -p c ] × Delta t, i = 1, 2, 3, 4 is the position vector of the foot relative to the center of mass;
[0131] Disturbance compensation matrix The disturbance term estimated by ESO is introduced into the discrete model to realize feedforward disturbance compensation, and the expression is:
[0132]
[0133] The above is the detailed implementation method of the discrete state space model constructed in the present application. By modeling the system state, control input and disturbance compensation term in matrix form, a rigorous and programmable numerical basis is provided for the solution of the subsequent MPC optimizer. This form has the characteristics of clear structure, clear physical meaning of parameters, strong scalability, etc., and is suitable for unified implementation in simulation and actual controller deployment.
[0134] After the discrete state space model containing the disturbance compensation term is constructed, the present application formulates a linear model predictive control (MPC) problem in a finite time domain based on the model, with a prediction range length of n, i.e. the system behavior in the future n steps is predicted in each control period of the optimizer, and the optimal control input sequence is solved online. The linear MPC problem can be expressed as follows:
[0135]
[0136] Where, represents the system state at time i, represents the desired state of the system at the same time. F i is the ground reaction force calculated at time step i, R i and S i are diagonal positive semi-definite matrices. The components of the contact force vector of each foot are represented as F ix , F iy and F izFurthermore, μ denotes the friction coefficient between the contact and the ground, K i is an inequality constraint matrix, D i is a force constraint matrix of the swing leg.
[0137] The linear MPC problem constructed by the application combines a disturbance real-time estimation mechanism based on an ESO, effectively overcoming the problems of strong dependence on model accuracy and insufficient robustness to disturbances of traditional MPC. The optimization model has the following advantages: accurate state prediction, adaptation to complex trajectory changes, disturbance feedforward compensation, improved disturbance rejection capability, support for multiple constraint conditions, adaptation to actual robot physical limitations, clear algorithm structure, and ease of engineering implementation and real-time deployment. At the same time, the controller structure has been verified in a simulation platform and an actual quadruped robot system, showing excellent trajectory tracking performance and robustness. The specific verification method is described in S5.
[0138] Step S4: Implementation of experimental verification and analysis
[0139] To verify the effectiveness and robustness of the quadruped robot model predictive control method (ESO-MPC) proposed by the application, a simulation experiment is designed to test it on a virtual ring, evaluate its motion control performance under complex disturbance conditions, and compare and analyze it with the traditional disturbance observer combined MPC method (DOB-MPC).
[0140] The application builds an experimental environment in the Gazebo simulation platform, and the operating system is Ubuntu 20.04. In the experiment, the URDF model description file of the Unitree Go1 quadruped robot is used to ensure the consistency of the simulation environment and the real platform in structure and parameters. The controller is implemented in C++ language, and the real-time driving and interaction of the simulation robot are realized through the control interface under the ROS framework.
[0141] This experiment aims to evaluate the control ability of the Go1 robot in maintaining the desired center of mass height, linear velocity and angular velocity under disturbance conditions when carrying a payload with unknown mass. The ESO-MPC control method proposed by the application and the traditional DOB-MPC method are used to perform linear velocity tracking tasks on the robot, and the actual motion trajectory of the robot under different control strategies is compared with the set reference trajectory, thereby verifying the effectiveness and robustness of the proposed control framework.
[0142] In the simulation experiment of the application, the following two groups of typical simulation experiment scenes are designed:
[0143] Scene one: straight walking experiment with 6kg payload
[0144] In this experiment, the robot's center of mass height (CoM) is set to 0.32 m, and the desired forward speed is set to 1.5 m / s. The robot's high-speed straight-line motion under constant load disturbance is simulated in the simulation platform. The ESO-MPC control method proposed in the present application and the traditional DOB-MPC control method are used for control respectively, and the actual tracking curves of CoM height and CoM speed are recorded, and the simulation comparison results are shown in Figure 2 .
[0145] From the experimental data, it can be observed that both control methods can achieve accurate center of mass speed tracking tasks, with good disturbance rejection ability and stability. Among them, the ESO-MPC method is slightly better than the DOB-MPC method in tracking error control, showing higher speed tracking accuracy and control smoothness, verifying the control advantage of the present application method in dealing with payload disturbance.
[0146] Scenario two: in-place rotation experiment with 6 kg payload
[0147] In this experiment, the robot's center of mass height is still set to 0.32 m, and the desired rotational angular velocity is set to 2 rad / s. The robot performs an in-place rotation task without moving the position, and records the CoM height and base angular velocity to test the attitude tracking ability and disturbance torque compensation ability, and the simulation comparison results are shown in Figure 3 .
[0148] The experimental results show that both methods can achieve relatively stable angular velocity output in the in-place rotation control task, and the ESO-MPC control method has faster response and smaller oscillation in disturbance compensation, showing higher attitude control stability and robustness.
[0149] Error analysis and quantitative result comparison
[0150] According to the recorded body height tracking errors e1 and e2 data in the simulation process, their L2 norms are calculated respectively, and the results are shown in Table 1. The experimental data show that under the same disturbance conditions and desired input, the ESO-MPC method is better than the DOB-MPC method in the tracking error of the body height.
[0151]
[0152] Table 1 L2 norms of body height tracking errors e1 and e2 in the simulation experiment of translation and rotation.
[0153] The above merely describes the preferred embodiments of the present application, and it should be pointed out that those skilled in the art can make several improvements and refinements without departing from the principles of the present application, and these improvements and refinements should also be considered as the protection scope of the present application. The structures, devices and operation methods not specifically described and explained in the present application are implemented according to the conventional means in the art, unless specifically described and limited.
Claims
1. A predictive control method for a quadruped robot model incorporating an extended state observer, characterized in that: It includes the following steps: Step 1: Construct a dynamic model of the quadruped robot, establish the kinematic equations of a single rigid body with the center of mass as the reference point, introduce system perturbation terms, and form a state-space model with perturbation. Step 2: Design the dual extended state observer (ESO), including: Position ESO is used to estimate unknown disturbance forces in the robot's translational direction in real time; Attitude ESO is used to estimate disturbance moments about the roll, pitch, and yaw axes in real time. Step 3: The disturbance estimate output by ESO is used as a feedforward compensation term and embedded into the constructed discrete-time state-space model to form a model predictive control (MPC) system with disturbance compensation. Step 4: Construct a finite-time linear MPC optimization problem, using the minimum trajectory tracking error as the cost function, while considering the following constraints: the ground reaction force satisfies the friction cone constraint; the contact force of the swing leg is zero; the foot force and joint torque are within the upper and lower limits; and the state variables satisfy the physical boundary conditions. Step 5: Solve the MPC problem online using the rolling optimization method to obtain the optimal foot reaction force and joint torque, and apply them to the quadruped robot system to complete robust trajectory tracking control under disturbance environment.
2. The quadruped robot model predictive control method according to claim 1, characterized in that, The robot kinematics modeling method in step 1 specifically includes: The kinematic modeling method using a single rigid body with the center of mass as the reference point is adopted; its representation is as follows: in, For the quality of the robot; Let be the rotational inertia of the robot's center of mass; It is the gravity vector; The base angular velocity of the robot; The location of the center of mass; i = 1, 2, 3, 4 represents the position of the foot tip; For the contact force of each leg; The single rigid body dynamic model incorporating disturbance terms is established as follows: in, This indicates the uncertainty of the robot in the translational direction. This represents the uncertainty in the robot's rotational direction; in this case, if ω is ignored... b ×I G ω b Due to the influence of the above formula, it can be expanded into the following form. When the angular velocity of the object is small, ω b ×I G ω b Negligible; To complete the observer design, the following assumptions are proposed: Assumption 1.F b d(t), d(t), and their time derivatives have upper bounds. That is, ||F b (t)‖2≤γ1,‖d(t)‖2≤γ2, and Where γ1, γ2, δ1 and δ2 are random positive constants.
3. The predictive control method for a quadruped robot model with an integrated extended state observer according to claim 2, characterized in that, The design of the extended state observer based on disturbance force in step 2 specifically includes: According to formula (1.2): in, To control input items; This refers to the positional disturbance force acting on the robot's center of mass; Design an n-order position ESO to estimate the unknown term. make Let p, v, and f represent respectively. Δ,1 ,...,f Δ,n The estimate; then the nth-order ESO design is: in, They represent The i-th element; x v,i For x v The i-th element; These are adjustable parameters for the ESO; according to the ESO design and formula (1.3), the estimated error is bounded, i.e. From formula (1.3), we can see that r p,1 ,r p,2 ,r p,3 With ω p,1 ,ω p,2 ,ω p,3 It is inversely proportional to the k-th power; At the same time, use a large gain vector ω p To obtain a smaller estimation error bound, we can use the position of ESO; at this point, let r p Indicates r p,1 ,r p,2 ,r p,3 Three times the maximum value, i.e., r p =3max{r p,1 ,r p,2 ,r p,3 If}, then there exists a limit. The closed-loop dynamics of the position loop are divided into two sub-loops: position and velocity. This dual-sub-loop structure is matched to the cascaded structure of a quadruped robot. Let the desired position and position error of the base be represented respectively. Then the expected dynamics of the two sub-loops are as follows: in, It is a diagonal matrix; the position error is defined as s. v =v c -v r v c v is the actual speed. r The desired velocity reference trajectory; to achieve the desired position sub-loop, v r Designed as follows: v r and The relationship is shown in formula (1.24); at the same time, the velocity error s is obtained. v The relationship with position error is as follows: As shown in formula (1.8), the desired positional sub-loop can be obtained through r. v →0 is used to achieve this; in order to illustrate s v and To understand the relationship between position dynamics and the stability of position dynamics, we introduce the following technical lemma 1, assumption 2, and lemma 2: Lemma 1: Let r1 denote a constant, if ||s v If ||≤r1, t→∞, then t→∞, where, express traces; Assumption 2: Assumption If it is bounded, then we can obtain Bounded, that is Lemma 2: According to Assumption 2, if the position control law is as shown in Formula (1.9) and the position ESO is as shown in Formula (1.5), then the position tracking error is... Bounded, that is t→∞, where r p This is the upper bound of the position ESO estimation error; 4. The predictive control method for a quadruped robot model with an integrated extended state observer according to claim 3, characterized in that, The design of the disturbance-based extended state observer in step 2 specifically includes: According to formulas (1.2) and (1.3), the dynamic model can be rewritten as: in, For control input; For disturbance torque; Design an n-order attitude ESO to estimate unknown disturbance terms. make They represent ω respectively b ,τ Δ,1 ,...,τ Δ,n The estimation of the nth-order attitude ESO is then performed as follows: Where, x ω,i It is x ω The i-th element, For the attitude ESO adjustable bandwidth vector, They are respectively and The i-th element; as mentioned above, the estimation error is bounded, i.e. And because r ω,1 ,r ω,2 ,r ω,3 With ω ω,1 ,ω ω,2 ,ω ω,3 The value is inversely proportional to the k-th power of the gain vector; therefore, the position of the ESO with a larger gain vector obtains a smaller estimation error bound; let r ω Indicates r ω,1 ,r ω,2 ,r ω,3 Three times the maximum value, i.e., r ω =3max{r ω,1 ,r ω,2 ,r ω,3 If}, then there exists a limit. Let ω d The desired angular velocity of the quadruped robot is represented by the following formula: Define the error rotation matrix as Define error quaternions in, and set up The derivative of the error rotation matrix can be obtained using the following formula: in,[·] × This represents the cross product operation; To achieve motion control of the quadruped robot, the closed-loop dynamics of the attitude loop are divided into two sub-loops: attitude and angular velocity. The design dynamics of the two sub-loops are given below: in, and Let k be a diagonal matrix. β >0 represents a constant gain; furthermore, s ω =ω-ω r ω r Provide a reference trajectory for the required angular velocity; To achieve the desired attitude sub-loop, ω r Designed as follows: roll out The desired attitude sub-loop is then s. ω →0; To achieve the desired angular velocity sub-loop, the quadruped robot base attitude controller is designed as follows: in, Substituting the control law (1.19) into formula (1.10) yields: according to Bounded; under control law (1.19), the desired angular velocity subloop is realized in a bounded manner, then the control torque vector T B for: To illustrate the stability of the attitude loop, we introduce the following assumption 3, Lemma 3; Assumption 3: Assumption If it is bounded, then it can be obtained. Bounded, that is Lemma 3: According to Assumption 3, if the attitude controller is designed as Equation (1.21) and the attitude ESO is Equation (1.11), then the attitude tracking error is bounded, i.e. t→∞, where r q This is the upper bound of the attitude tracking error.
5. The quadruped robot model predictive control method according to claim 4, characterized in that, The construction of the model predictive control system with disturbance compensation in step 3 specifically includes: According to formulas (1.4) and (1.16), the continuous-time state-space model of the quadruped robot system is expressed as: Its discrete-time state-space model can be represented as: The variables in formula (1.23) are defined as follows: State variables for: in, The location of the robot's center of mass; The attitude angle; The linear velocity of the center of mass; The angular velocity of the machine body; The disturbance force and disturbance moment are estimated from the position ESO and attitude ESO. Input control The expression for the expected ground reaction force of a quadruped robot is: in, Let it be expressed as the expected ground reaction force of the nth leg; Gravity term vector The expression for gravitational acceleration is: ESO estimator vector The ESO estimate containing the interference term and its derivative, used for feedforward compensation, are expressed as follows: Transition matrix The fundamental dynamic updates of the system, including position and velocity integrals and attitude updates, are expressed as follows: Input matrix The expression is: Where m is the robot mass, I G For the body's inertia matrix, [p i -p c ] × Δt, i = 1, 2, 3, 4 is the position vector of the foot relative to the center of mass; Disturbance compensation matrix The expression for incorporating the ESO-estimated disturbance term into the discrete model to achieve feedforward disturbance compensation is as follows:
6. The quadruped robot model predictive control method according to claim 5, characterized in that, The construction and solution of the MPC optimization problem specifically includes: After constructing a discrete state-space model including disturbance compensation terms, a finite-time linear model predictive control (MPC) problem is formulated based on this model. The prediction range length is n, meaning that the optimizer predicts the system behavior for the next n steps within each control cycle, and solves for the optimal control input sequence online. This linear MPC problem can be expressed as follows: in, This represents the system state at time i. F represents the desired state of the system at any given time; i The ground reaction force R is calculated at time step i. i and S i It is a diagonal positive semi-definite matrix; the components of the contact force vector for each foot are represented by F. ix F iy and F iz Furthermore, μ represents the coefficient of friction between the contact point and the ground, K i It is the inequality constraint matrix, D i It is the force constraint matrix of the swing leg.
Citation Information
Patent Citations
Pneumatic tail end self-adaptive force control method based on double extended state observers
CN117193021A
Adaptive robust path tracking control method based on ESO-DFLOS guidance law
CN117950321A
Cable inspection robot attitude control method based on fusion of MPC and ADRC
CN120469215A
Stratospheric airship trajectory tracking method based on reinforcement learning optimal control
WO2024216870A1
Cited By
Small robot dynamic balance control system based on DDPG
CN121179442A
