Self-adaptive limping control method for quadruped robot under leg failure condition

Abstract

The invention discloses a self-adaptive limping control method for a quadruped robot under a leg failure condition. The method comprises the following steps that S1, a system monitors the leg motion state of the robot in real time through sensor data; s2, judging whether the leg motion state is abnormal or not based on a preset key detection index, and if yes, executing the step S3; if not, directly executing the step S4 and the like. According to the method, self-adaptive limp control is achieved through algorithm level optimization in the whole process, a monitoring system can be constructed based on an existing sensor, model parameter adjustment and double-layer optimization control are achieved through software updating, and a hardware structure of the quadruped robot does not need to be transformed. The hardware cost of technology landing is reduced, the method can be adapted to different types of quadruped robots, the compatibility is high, and the method is convenient to popularize in existing quadruped robot application scenes such as emergency rescue and industrial inspection.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of robots, and particularly relates to a self-adaptive limping control method of a quadruped robot under a leg failure condition. BACKGROUND

[0002] As an important branch of robot technology, quadruped robots have been widely applied in key fields such as emergency rescue, terrain exploration and industrial inspection, due to their excellent adaptability and stability in complex terrains. In the emergency rescue scene, it can enter areas such as earthquake ruins and fire sites that are difficult for humans to reach, and perform survivor search and rescue and environmental monitoring tasks; in terrain exploration, it can shuttle in complex topography such as mountains and deserts, and collect geological and meteorological data; in industrial inspection, it can conduct all-weather and high-precision equipment state inspection and safety hazard investigation on large factories and power transmission lines.

[0003] However, leg failures (such as motor damage, joint jamming or mechanical structure fracture) of quadruped robots occur from time to time in actual operation. Such failures can significantly change the dynamic characteristics of the robot, making it difficult for traditional control methods to adapt, and thus causing problems such as robot motion instability and task interruption, and even causing damage to the robot itself, resulting in huge losses to related operations.

[0004] At present, domestic and foreign research institutions have carried out a large amount of research on the motion control of quadruped robots and have proposed various control schemes. Among them, model predictive control (MPC) as an advanced control method can explicitly handle system constraints and accurately predict the future state changes of the robot by solving optimization problems in a finite time domain online. When the structure of the quadruped robot is complete and the running state is normal, the standard MPC controller exhibits good trajectory tracking performance and can effectively ensure that the robot moves stably according to the preset path.

[0005] However, this method is highly dependent on the accuracy of the robot's dynamic model, and when the leg fails and the model parameters change significantly, its prediction accuracy will decrease significantly, and the control performance will also decrease. At the same time, the process of solving optimization problems online by MPC has a heavy computational burden, and in the special scenario of leg failure, which requires rapid re-planning of the motion strategy, the standard MPC cannot meet the real-time requirements, and it cannot dynamically adjust the constraint conditions according to the changes in the feasible motion range and working space of the robot after the leg failure, making it difficult for the robot to adapt to the running state after the failure.

[0006] On the other hand, reinforcement learning (RL) methods have demonstrated strong robustness and environmental adaptability in the field of robot control. Some studies have transferred learned control policies to physical robots through extensive training in simulation environments, enabling the robots to achieve stable movement in complex terrains. However, pure reinforcement learning methods typically require massive amounts of training data to support policy learning, and the lack of prior knowledge guidance during the learning process leads to long training cycles, high computational costs, difficulty in quickly responding to sudden leg failures, and inability to efficiently generate control policies adapted to fault states.

[0007] In summary, existing quadruped robot control methods all have significant shortcomings in dealing with leg failure issues and cannot meet the requirements for continuous and stable operation of robots under fault conditions in practical applications. There is an urgent need for a control method that can quickly and adaptively adjust under leg failure conditions to improve the reliability and task completion capability of quadruped robots. Summary of the Invention

[0008] The purpose of this invention is to overcome the shortcomings of existing quadruped robot control methods in dealing with leg failure, which cannot meet the requirements of continuous and stable operation of robots under fault conditions in practical applications. This invention provides an adaptive limp control method for quadruped robots under leg failure conditions that can quickly and adaptively adjust.

[0009] The objective of this invention is achieved through the following technical solution: an adaptive limp control method for a quadruped robot under leg failure conditions, comprising the following steps: S1. The system monitors the robot's leg movement status in real time through sensor data; S2. Determine whether the leg movement status is abnormal based on preset key detection indicators. If yes, proceed to step S3; otherwise, proceed directly to step S4. S3. The system uses a state classifier based on a neural network to determine the fault type and triggers the model update mechanism. The system then initiates the online estimation process of model parameters and completes the constraint update. S4, Policy Networks for High-Level Reinforcement Learning Receive input information and process the data, then output the MPC parameter adjustment amount; S5. The controller converts the optimal input sequence calculated by MPC into joint torque commands and executes them.

[0010] Furthermore, in step S2, "judging whether the leg movement state is abnormal based on preset key detection indicators", the key detection indicators include at least joint torque abnormality, foot contact failure, and body posture deviation; wherein, joint torque abnormality refers to the torque output of a certain leg not matching the expectation, foot contact failure refers to the foot sensor detecting an unexpected lifting or dragging phenomenon, and body posture deviation refers to the IMU detecting body tilt or continuous shift of the center of gravity.

[0011] The "system determines the fault type by a state classifier based on a neural network" mentioned in step S2 includes single-leg complete failure, single-leg partial failure, multi-leg complete failure, and multi-leg partial failure.

[0012] Step S3, the "system startup model parameter online estimation process", specifically includes the following steps: S31. Recalculate the robot's total mass, center of mass position, and inertia tensor according to the fault type; S32. Update the Jacobian matrix and dynamic equations; S33. Redefine motion constraints based on the workspace of the robot's remaining effective legs.

[0013] Furthermore, the "policy network" described in step S4 The expression for " is: ,in, It is the set of parameters for the policy network. Indicates based on input information and network parameters The calculated prediction time domain, Indicates based on input information and network parameters The calculated state weight matrix, Indicates based on input information and network parameters The calculated input weight matrix, Indicates based on input information and network parameters The calculated function of the constraint slack variables, the In robot state, For reference commands, This is privileged information.

[0014] Furthermore, the "MPC parameter adjustment amount" mentioned in step S4 includes the prediction time domain N, the state weight matrix Q, the input weight matrix R, and the constraint slack variables.

[0015] Furthermore, the calculation formula for the "joint torque command" mentioned in step S5 is as follows: Where J is the Jacobian matrix, F is the optimal ground reaction force calculated by MPC, Kp is the proportionality coefficient, Kd is the differential coefficient, and θ_d es For the desired joint angle, θ_c ur dθ_d represents the current joint angle. es / dt represents the desired joint angular velocity, dθ_c ur / dt represents the current joint angular velocity.

[0016] As a preferred embodiment, the "sensor" mentioned in step S1 includes at least: a joint encoder, an IMU unit, and a foot contact sensor.

[0017] The "motion constraint" mentioned in step S33 includes at least the joint angle constraint, joint angular velocity constraint, and foot end workspace constraint of the remaining effective leg.

[0018] Compared with the prior art, the present invention has the following advantages and beneficial effects: (1) This invention combines real-time monitoring by multiple sensors with preset key detection indicators to quickly identify abnormal leg movements; then, by using a state classifier based on neural networks to subdivide the fault types and trigger a targeted model update mechanism, it can avoid robot shutdown due to fault judgment delay or misjudgment, and effectively ensure that the robot can continue to perform tasks after leg failure.

[0019] (2) The model update mechanism of the present invention will recalculate the total mass, centroid position and other mass attributes based on the fault type, update the Jacobian matrix and dynamic equation, and adjust the motion constraints (joint angle, angular velocity, foot end workspace constraints) according to the remaining effective leg workspace, so as to ensure that the model parameters are highly matched with the actual dynamic characteristics of the robot after the leg failure, solve the control deviation problem caused by the fixed model in the traditional control method, and significantly improve the accuracy of the control strategy.

[0020] (3) The high-level reinforcement learning policy network of the present invention By combining robot status, reference commands, and privileged information from the training phase, the MPC parameter adjustment amount is dynamically output. The underlying MPC solves the optimization problem based on the adjusted parameters, forming a collaborative mode of "high-level decision-making and low-level execution". This enables the robot to flexibly adjust its motion strategy and become more adaptable even in complex scenarios such as leg failure and changes in motion constraints.

[0021] (4) This invention directly converts the optimal input sequence calculated by MPC into joint executable torque commands through a clear joint torque command calculation formula. No additional complex conversion process is required, reducing command transmission errors and ensuring the stability of robot movement after leg failure, avoiding problems such as body tilt and gait disorder.

[0022] (5) The present invention achieves adaptive limp control through algorithm-level optimization throughout the entire process. It can build a monitoring system based on existing sensors and achieve model parameter adjustment and two-layer optimization control through software updates, without the need to modify the hardware structure of the quadruped robot. This not only reduces the hardware cost of technology implementation, but also adapts to different types of quadruped robots, has strong compatibility, and is easy to promote in existing quadruped robot application scenarios such as emergency rescue and industrial inspection. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall process structure of the present invention.

[0024] Figure 2 This is a flowchart illustrating the online estimation process of system startup model parameters according to the present invention. Detailed Implementation

[0025] The present invention will be further described in detail below with reference to embodiments, but the implementation of the present invention is not limited thereto.

[0026] Example

[0027] like Figure 1 As shown in the figure, the adaptive limp control method for a quadruped robot under leg failure conditions described in this embodiment includes steps S1 to S5.

[0028] Step S1 involves the system monitoring the robot's leg movement status in real time using sensor data. The "sensors" include at least a joint encoder, an IMU unit, and a foot contact sensor. Other types of sensors can be added to this embodiment as needed.

[0029] The joint encoder is a sensor used to directly monitor the motion state of the robot's leg joints. Its core function is to collect joint-related data, providing basic parameters for fault detection and control calculations. It can collect the torque output data of the leg joints in real time and transmit it to the system. The system compares this data with a preset normal torque range to determine whether there are abnormal joint torque conditions such as "a certain leg's torque output is consistently zero" or "the torque is abnormally high," providing direct data for judging key detection indicators in step S2 and helping to initially identify signs of leg faults. Simultaneously, in addition to real-time torque data collection, the joint encoder also collects motion parameters such as joint angles and angular velocities. These parameters serve as the robot's state (…). As an important component, it will be fed into the policy network of high-level reinforcement learning. This provides crucial data support for the strategy network to analyze the robot's real-time motion state and accurately output MPC parameter adjustment amounts, ensuring the rationality of subsequent control strategies.

[0030] The IMU (Inertial Measurement Unit) is mainly used to monitor the robot's overall posture and center of gravity. It is the core basis for state judgment in the detection of body posture deviations and control decisions. Specifically, its functions are as follows: First, the IMU unit can collect data such as the robot's tilt angle (e.g., roll angle, pitch angle), tilt rate, center of gravity offset, and offset trend in real time. By analyzing this data, the system determines whether there are body posture deviations such as "body tilt exceeding the normal motion threshold" or "center of gravity continuously shifting in a certain direction." This data is an important component of the key detection indicators in step S2 and can help identify the robot's overall motion imbalance caused by leg failure. Second, the body posture and center of gravity data collected by the IMU unit, together with the joint parameters collected by the joint encoder, constitute a complete robot state ( Input to the policy network The policy network, by combining this data, can gain a more comprehensive understanding of the robot's posture changes caused by leg failure, thereby optimizing the adjustment of MPC parameters and ensuring that the control strategy generated by the underlying MPC can adapt to the robot's posture adjustment needs and avoid motion instability.

[0031] The foot contact sensor is used to directly monitor the contact state between the foot and the ground, and is crucial for foot contact failure detection and motion constraint adjustment. Specifically, it has two main functions: First, the foot contact sensor can collect data such as contact pressure, contact duration, and ground-lift frequency in real time. The system analyzes this data to determine if foot contact failures have occurred, such as "unexpectedly long ground-lift time" or "abnormal contact pressure during mopping." This data is the core content of the key detection indicators in step S2, directly reflecting whether the leg is unable to interact normally with the ground due to failure, thus aiding in accurate identification of leg malfunctions. Second, in the model update stage (step S3), the foot contact state data collected by the foot contact sensor can help the system determine the actual working range of the remaining effective legs.

[0032] S2. Determine whether the leg movement status is abnormal based on preset key detection indicators. If yes, proceed to step S3; otherwise, proceed directly to step S4.

[0033] The "judging whether the leg movement state is abnormal based on preset key detection indicators" mentioned above includes at least joint torque abnormality, foot contact failure and body posture deviation.

[0034] Specifically, the abnormal joint torque refers to the torque output of a certain leg not matching the expectation; the foot contact failure refers to the foot sensor detecting an unexpected lift-off or dragging phenomenon; and the body posture deviation refers to the IMU detecting body tilt or continuous shift of the center of gravity.

[0035] S3. The system uses a state classifier based on a neural network to determine the fault type and triggers a model update mechanism. The system then initiates the online estimation process of model parameters and completes the constraint update.

[0036] The "failure type" includes four situations: complete failure of a single leg, partial failure of a single leg, complete failure of multiple legs, and partial failure of multiple legs.

[0037] The term "single-leg complete failure" refers to the presence of a combination of features in the input feature vector: "a certain leg has a continuously zero torque, no contact signal at the foot, and the body is significantly tilted towards that leg." When multiple legs exhibit this combination of features, it is considered a multi-leg complete failure. For example, this quadruped robot has a left foreleg, a left hindleg, a right foreleg, and a right hindleg. If only the left foreleg exhibits the combination of continuously zero torque, no contact signal at the foot, and a significant tilt of the body towards that leg, while the left hindleg, right foreleg, and right hindleg do not, then it is a single-leg complete failure. If only the left foreleg and left hindleg simultaneously exhibit the combination of continuously zero torque, no contact signal at the foot, and a significant tilt of the body towards that leg, then it is a multi-leg complete failure.

[0038] The aforementioned single-leg partial failure refers to the input feature vector displaying a combination of features such as "abnormal torque fluctuation in a certain leg, intermittent failure of foot contact signal, and small body posture deviation that can be briefly recovered"; when multiple legs simultaneously exhibit the above combination of features, it is considered a multi-leg partial failure.

[0039] The process of the "system startup model parameter online estimation process" is as follows: Figure 2 As shown, the specific steps include: S31. Recalculate the robot's total mass, center of mass position, and inertia tensor based on the fault type. This step is essentially a re-estimation of mass attributes, which corrects the robot's core mass attributes based on the fault type (such as complete failure of a single leg, partial failure of a single leg, complete failure of multiple legs, and partial failure of multiple legs) to eliminate the interference of the failed leg on the model calculation.

[0040] When recalculating the total mass, if the fault type is "complete failure," the system will remove the preset mass of the failed leg from the total mass of the robot, obtaining the total mass of the remaining effective structure. For example, if the total mass of the robot is M and the mass of a single leg is m, then the total mass after a single leg complete failure will be updated to Mm.

[0041] If the fault type is "partial failure", since the failed leg is still part of the robot structure (not completely detached or losing all mass contribution), the mass of the leg is not removed when calculating the total mass. Only its complete mass is included in the total mass of the whole machine to avoid over-correction that would cause the mass attribute to be inconsistent with the actual mass.

[0042] The recalculation of the center of mass position is based on the robot body coordinate system (usually with the geometric center of the body as the origin). The system first obtains the mass and initial coordinates (the preset coordinates of each component under normal conditions) of each effective leg (and the main body), and then recalculates the coordinates of the center of mass of the whole machine through the center of mass calculation formula.

[0043] For example, when a single leg fails completely, after removing the mass and coordinates of the failed leg, the centroid position will shift towards the side where the effective legs are concentrated: assuming the centroid coordinates in the normal state are (x0, y0, z0), and the coordinates of the failed leg in the coordinate system are (x1, y1, z1), then the updated x-axis centroid coordinates are... Similarly, the y-axis and z-axis coordinates are determined to ensure that the position of the center of mass can reflect the displacement of the robot's center of gravity after failure, thus providing an accurate reference for subsequent dynamic calculations.

[0044] The system recalculates the inertia tensor, which reflects the robot's rotational inertia about each axis of the coordinate system. Its calculation depends on the total mass, the mass distribution of each component, and the position of the center of mass. Based on the updated total mass and center of mass position, and combined with the mass distribution parameters of the effective legs, the system recalculates the nine components of the inertia tensor using the inertia tensor matrix formula.

[0045] For example, if a single leg fails completely, the rotational inertia around the axis perpendicular to the body plane will decrease due to the reduced mass on one side. The system corrects the inertia tensor components to ensure that the subsequent dynamic equations can accurately reflect the changes in the robot's rotational characteristics, thus avoiding errors in the calculation of control torque due to deviations in inertial parameters.

[0046] The above formulas for calculating the centroid and the inertia tensor matrix are existing formulas and contents, and will not be repeated in this embodiment.

[0047] S32. Update the Jacobian matrix and the dynamic equations.

[0048] The core function of the Jacobian matrix J is to establish the mapping relationship between joint angular velocity and foot linear velocity or angular velocity. When a leg fails, the joints of the failed leg cannot move normally (complete failure) or have limited movement (partial failure), requiring targeted adjustments to the Jacobian matrix. If it is a "complete failure," the columns corresponding to the joints of the failed leg in the Jacobian matrix are directly deleted, retaining only the columns corresponding to the joints of the effective leg, forming a new Jacobian matrix adapted to the movement of the effective leg. This ensures that the matrix accurately reflects the mapping relationship between effective joint motion and foot velocity. If it is "partially failed", the column elements in the Jacobian matrix corresponding to the failed joint are set to 0 (indicating that the angular velocity of that joint does not contribute to the foot velocity), while the columns corresponding to other effective joints of the leg are retained to avoid interference from the failed joint in the velocity mapping calculation.

[0049] The core form of the dynamic equation is: ,in, For joint torque, The inertia matrix, The Coriolis-centrifugal force matrix, For the gravity matrix, is the effective joint angular acceleration vector, J is the updated Jacobian matrix, and F is the foot ground reaction force vector.

[0050] S33. Redefine motion constraints based on the workspace of the robot's remaining effective legs.

[0051] This step involves resetting the constraint boundaries of the robot's motion based on the physical motion limits of the remaining effective legs. This prevents control commands from failing to execute or the robot from being damaged due to a mismatch between the constraints and the actual capabilities of the effective legs. The specific constraint content and definition logic are as follows: Regarding joint angle constraints, the system first obtains the physical activity limits of each joint in each effective leg (the maximum / minimum angles of each joint need to be stored in advance, such as the hip joint's rotation range around the x-axis being -30° to 60°), and then determines the actual usable range of the effective joints based on the failure type. If it is a "complete failure," all joint angle constraints of the failed leg are directly removed, and only the angle constraints of the effective leg joints are retained; if it is a "partial failure," the angle constraints of the effective joints of the failed leg are retained, and fixed angle constraints are set for the failed joints to ensure that the constraints can reflect the actual motion capability of the joints.

[0052] Constraints on joint angular velocity. Based on the effective leg joint's driving capability, the system resets the upper and lower limits of the effective joint's angular velocity. For example, if a single leg fails completely, to prevent the effective leg from exceeding its angular velocity limit due to increased load, the maximum angular velocity constraint of the effective joint can be appropriately reduced (e.g., from 10 rad / s in the normal state to 8 rad / s) to ensure that the joint movement is within a safe driving range.

[0053] Constraints on the foot-end workspace. The foot-end workspace refers to the area reachable by the effective leg's foot in space. The system redefines the constraints as follows: First, based on the joint angle constraints and link lengths of the effective leg, the reachable range of the foot in the fuselage coordinate system (e.g., the range of movement in the x-axis direction and the height range in the z-axis direction) is calculated using forward kinematics. Then, combined with the ground environment, a secondary constraint is applied to the foot-end workspace. This ultimately forms the foot-end position constraint (x... min ≤x foot ≤ x max y min ≤y foot ≤y max , z min ≤z foot ≤ z maxThis ensures that MPC plans movement only within the effective reach when generating foot trajectories, avoiding invalid or dangerous foot position commands.

[0054] S4, Policy Networks for High-Level Reinforcement Learning Receive input information and process the data, then output the MPC parameter adjustment amount.

[0055] The policy network The input information needs to comprehensively reflect the robot's real-time state, motion targets, and environmental priors during the training phase to ensure the targeted and accurate nature of parameter adjustment decisions. This "policy network" The expression for " is: ,in, It is the set of parameters for the policy network. Indicates based on input information and network parameters The calculated prediction time domain, Indicates based on input information and network parameters The calculated state weight matrix, Indicates based on input information and network parameters The calculated input weight matrix, Indicates based on input information and network parameters The calculated function of the constraint slack variables, the In robot state, For reference commands, This is privileged information.

[0056] The "MPC parameter adjustment amount" includes the prediction time domain N, the state weight matrix Q, the input weight matrix R, and the constraint slack variables.

[0057] S5. The controller converts the optimal input sequence calculated by MPC into joint torque commands and executes them.

[0058] This step connects control decisions with robot motion execution. Its core function is to transform the abstract optimization results output by the underlying MPC into physical drive signals that can be directly executed by the robot's leg joints, ensuring that the robot can still complete the movement stably and accurately even if the legs fail.

[0059] The optimal input sequence refers to the underlying MPC (Model Predictive Control) based on the high-level policy network. The output parameter adjustment amount, combined with the robot dynamics model updated in step S3, is used to obtain abstract control target data after solving the optimization problem online. The calculation formula for the "joint torque command" is: Where J is the Jacobian matrix, F is the optimal ground reaction force calculated by MPC, Kp is the proportionality coefficient, Kd is the differential coefficient, and θ_d es For the desired joint angle, θ_c ur dθ_d represents the current joint angle. es / dt represents the desired joint angular velocity, dθ_c ur / dt represents the current joint angular velocity.

[0060] When a leg fails, the robot needs to switch to a limp gait. This gait requires extremely high real-time and continuous motion commands. Delays or interruptions in command transmission can cause gait transitions to become choppy, affecting motion stability. This embodiment employs a "real-time calculation-continuous output" working mode. On one hand, this ensures high-frequency data interaction between the controller and the MPC, allowing the controller to obtain the optimal input sequence updated by the MPC in real time, avoiding torque command lag due to data latency. On the other hand, the controller uses a smoothing algorithm to optimize the calculated joint torques. The process ensures continuous torque output variation, smoothly rotating the effective leg joints and making the support and swing phases of the limping gait naturally connected. This avoids the robot body from shaking or the feet from impacting the ground due to sudden changes in commands, thus adapting to the limping movement needs after leg failure.

[0061] As described above, the present invention can be well implemented.

Claims

1. An adaptive limp control method for a quadruped robot under leg failure conditions, characterized in that, Includes the following steps: S1. The system monitors the robot's leg movement status in real time through sensor data; S2. Determine whether the leg movement status is abnormal based on preset key detection indicators. If yes, proceed to step S3; otherwise, proceed directly to step S4. S3. The system uses a state classifier based on a neural network to determine the fault type and triggers a model update mechanism. The system then initiates the online estimation process of model parameters and completes the constraint update. S4, Policy Networks for High-Level Reinforcement Learning Receive input information and process the data, then output the MPC parameter adjustment amount; S5. The controller converts the optimal input sequence calculated by MPC into joint torque commands and executes them.

2. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 1, characterized in that, In step S2, "judging whether the leg movement state is abnormal based on preset key detection indicators", the key detection indicators include at least joint torque abnormality, foot contact failure, and body posture deviation; wherein, joint torque abnormality refers to the torque output of a certain leg not matching the expectation, foot contact failure refers to the foot sensor detecting an unexpected lifting or dragging phenomenon, and body posture deviation refers to the IMU detecting body tilt or continuous shift of the center of gravity.

3. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 2, characterized in that, The "system determines the fault type based on the state classifier of the neural network" mentioned in step S3 includes single-leg complete failure, single-leg partial failure, multi-leg complete failure, and multi-leg partial failure.

4. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 3, characterized in that, The "system startup model parameter online estimation process" described in step S3 specifically includes the following steps: S31. Recalculate the robot's total mass, center of mass position, and inertia tensor according to the fault type; S32. Update the Jacobian matrix and dynamic equations; S33. Redefine motion constraints based on the workspace of the robot's remaining effective legs.

5. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 4, characterized in that, The "policy network" mentioned in step S4 The expression for " is: ,in, It is the set of parameters for the policy network. Indicates based on input information and network parameters The calculated prediction time domain, Indicates based on input information and network parameters The calculated state weight matrix, Indicates based on input information and network parameters The calculated input weight matrix, Indicates based on input information and network parameters The calculated function of the constraint slack variables, the In robot state, For reference commands, This is privileged information.

6. An adaptive limp control method for a quadruped robot under leg failure conditions according to any one of claims 1 to 5, characterized in that, The "MPC parameter adjustment amount" mentioned in step S4 includes the prediction time domain N, the state weight matrix Q, the input weight matrix R, and the constraint slack variables.

7. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 6, characterized in that, The calculation formula for the "joint torque command" mentioned in step S5 is as follows: Where J is the Jacobian matrix, F is the optimal ground reaction force calculated by MPC, Kp is the proportionality coefficient, Kd is the differential coefficient, and θ_d es For the desired joint angle, θ_c ur dθ_d represents the current joint angle. es / dt represents the desired joint angular velocity, dθ_c ur / dt represents the current joint angular velocity.

8. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 6, characterized in that, The "sensor" mentioned in step S1 includes at least: a joint encoder, an IMU unit, and a foot contact sensor.

9. The adaptive limp control method for a quadruped robot under leg failure conditions according to claim 4, characterized in that, The "motion constraint" mentioned in step S33 includes at least the joint angle constraint, joint angular velocity constraint, and foot end workspace constraint of the remaining effective leg.

Images