Motion control method for dual-drive self-switching high-adaptability quadruped robot

Abstract

The invention discloses a dual-drive self-switching strong-adaptability quadruped robot motion control method, which comprises the following steps of: estimating current topographic features by using sensing signals such as body state feedback, IMU (Inertial Measurement Unit) and sole force, and predicting a falling probability according to the topographic features, so that a quadruped robot is intelligently switched between two driving modes of MPC (Moving Picture Control) independent control and MPC-DRL (Moving Picture Control-Digital Regulation Language) cooperative control; dynamic behavior parameters are introduced to influence the gait and step height of robot movement, the optimization efficiency and smoothness of MPC and the disturbance adaptive capacity of DRL are efficiently fused, and the high-robustness, high-efficiency and self-adaptive movement capacity of the quadruped robot in various terrains under the condition that external sensing information is not relied on is enhanced.

Description

Technical Field

[0001] This invention belongs to the field of quadruped robot motion control, and particularly relates to a terrain-adaptive, dual-drive self-switching quadruped robot motion control method. Background Technology

[0002] Quadruped robots, with their multi-legged support structure and omnidirectional mobility, can traverse uneven terrain (steep slopes, steps, gravel, etc.), making them promising candidates for applications in complex scenarios such as special operations, disaster relief, and field exploration. However, the unstructured and diverse terrain in complex scenarios poses significant challenges to the adaptability and robustness of quadruped robot motion control methods. Enhancing the terrain adaptability of quadruped robots in complex scenarios has become a research hotspot and a difficult issue in this field.

[0003] Faced with diverse terrains in complex scenarios, breakthroughs in the motion control capabilities of quadruped robots without external sensors primarily rely on methods based on Model Predictive Control (MPC) and Deep Reinforcement Learning (DRL). MPC-based methods perform exceptionally well in structured or predictable terrains (flat ground, regular stairs), offering high computational efficiency and smooth trajectories. However, they are extremely sensitive to model errors and unstructured disturbances (gravel, soft soil), easily leading to instability. DRL-based methods have demonstrated strong adaptability to diverse terrains in simulation training, capable of autonomously learning to cope with unknown disturbances. However, DRL-based methods suffer from drawbacks such as high real-time computational load, poor interpretability, and difficulty in Sim-to-Real transfer, often limiting their application in scenarios requiring high efficiency and safety. Some methods attempt to combine the advantages of MPC and DRL, but the mechanisms are mostly static parallel or simple rule switching, lacking a comprehensive assessment of the robot's real-time state (such as tilt angle and foot gliding) and environmental characteristics. The switching timing is rigid and lagging, making it difficult to ensure efficient, continuous and robust maneuverability in highly dynamic unstructured terrain.

[0004] In summary, without relying on external sensors, these methods suffer from problems such as the difficulty in overcoming the limitations of a single control paradigm and the lack of contextualized intelligent switching mechanisms, which prevent quadruped robots from maintaining efficient movement and body stability when the terrain changes. Summary of the Invention

[0005] The purpose of this invention is to provide a terrain-adaptive dual-drive self-switching quadruped robot motion control method, so as to achieve high robustness, high efficiency and adaptive motion capability of quadruped robots in diverse terrains.

[0006] The technical solution to achieve the purpose of this invention is as follows:

[0007] A motion control method for a quadruped robot with dual-drive self-switching and strong adaptation includes:

[0008] Step 1: Acquire joint motor, IMU, and foot force data of the quadruped robot;

[0009] Step 2: Use IMU data for pre-integration to obtain the robot's center of mass posture at the current moment; input the joint data and IMU data into the state estimator to obtain the robot's center of mass position and linear velocity in the world coordinate system, and the robot's three-axis rotational angular velocities in the body coordinate system.

[0010] Step 3: The robot's center of mass posture, center of mass position, center of mass linear velocity, rotational angular velocity, and foot force are combined into a one-dimensional vector, which is defined as the state in the Markov decision process;

[0011] Step 4: Input the current state into the terrain feature extraction module to obtain the feature vector of the quadruped robot's stability margin;

[0012] Step 5: Input the current feature vector into the behavior parameter adjustment strategy module to obtain the behavior parameter adjustment amount, including the increment of the quadruped robot's expected foot contact and step height.

[0013] Step 6: Input the current feature vector into the overturning risk assessment module to obtain the current robot fall probability;

[0014] Step 7: Input the fall probability into a switching function, and use the switching function to control the switching between the two driving modes of "MPC individual control" and "MPC-DRL cooperative control";

[0015] Step 8: Input the user speed command and behavior parameter adjustment into the model predictive controller. Based on the user speed command, generate the reference center of mass position, reference center of mass velocity, reference center of mass attitude angle and reference center of mass angular velocity based on the spring-loaded inverted pendulum model, which together constitute the center of mass reference trajectory.

[0016] Step 9, based on the reference centroid velocity The stride frequency is calculated with the preset stride length. A gait cycle table is established based on the gait cycle. The reference leg state at the current moment can be obtained by looking up the table. The final foot contact state is obtained by performing an "OR" operation between the corresponding elements of the behavior parameter adjustment.

[0017] Step 10: Based on the centroid reference trajectory and the final foot contact state, use the MPC solver to calculate the expected plantar force of the supporting leg.

[0018] Step 11: Based on the expected plantar force of the supporting leg calculated by the MPC solver, the plantar force of each supporting leg is converted into the joint torque of each joint of the supporting leg through the Jacobian matrix;

[0019] Step 12, based on the final foot contact state Increment of step height Construct the foot trajectory of the swing leg to obtain the foot trajectory of the swing term in the world coordinate system. shaft and Position component along the axis;

[0020] Step 13, based on the foot of the oscillating term in the world coordinate system shaft and The position component along the axis is obtained by inverse kinematics calculation to obtain the joint position and joint angular velocity of the thigh and lower leg joints in the swing term;

[0021] Step 14: Driven by the torque of the supporting neck joint and the position of the swinging neck joint, the quadruped robot moves in accordance with the user's speed command.

[0022] The significant advantages of this invention compared to existing technologies are:

[0023] This invention addresses the inherent limitations of single control paradigms based on MPC or DRL. Without relying on external sensors, it extracts terrain features from the robot's sensor data and utilizes the "memory" property of LSTM networks to assess the risk of falls caused by changes in terrain in real time. Based on the magnitude of the fall risk, it intelligently switches between two driving modes: "MPC independent control" and "MPC-DRL cooperative control." By introducing dynamic behavioral parameters to influence the robot's gait and step height, it efficiently integrates the optimization efficiency and smoothness of MPC with the disturbance adaptation capability of DRL, thereby enhancing the quadruped robot's robust, efficient, and adaptive movement capabilities in diverse terrains without relying on external sensor information. Attached Figure Description

[0024] Figure 1 This is a flowchart of a motion control method.

[0025] Figure 2 This is a schematic diagram of the coordinate orientation of a quadruped robot.

[0026] Figure 3 Extracting network structure diagrams for terrain features.

[0027] Figure 4 This is a flowchart of motion control based on MPC. Detailed Implementation

[0028] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0029] like Figure 1 As shown, the present invention provides a terrain-adaptive, dual-drive self-switching quadruped robot motion control method, which specifically includes the following steps:

[0030] Step 1: Acquire sensor data from the quadruped robot's joint motors, IMU, and foot force using the robot's body sensors. In this embodiment, the positive direction of robot movement is defined as... Positive axis direction shaft and The positive axis follows the right-hand rule. Initially, the world coordinate system coincides with the body coordinate system, which is located at the quadruped robot's center of mass. The foot force of the quadruped robot is defined as... Among them, the force of the right foreleg foot sole is (Number 1), Left foreleg foot plantar force is (Number 2), right hind leg plantar force is (Number 3), Left hind leg plantar force is (Number 4), such as Figure 2 As shown.

[0031] Step 2: Use IMU data for pre-integration to obtain the robot's center of mass pose at the current moment. , The pitch angle, For roll angle, As the yaw angle, the joint data and IMU data are input into the state estimator to obtain the position of the robot's center of mass in the world coordinate system. and linear velocity Angular velocity of the robot's three axes in the body coordinate system .

[0032] Step 3, adjust the robot's center of mass orientation. Location of the center of mass linear velocity of the center of mass angular velocity of rotation With foot strength We can construct a one-dimensional vector and define it as the state in the Markov decision process. .

[0033] Step 4, record the current state. The input terrain feature extraction module obtains feature vectors containing implicit terrain features such as slope, roughness, and continuity, along with the stability margin of the quadruped robot. In this example, the terrain feature extraction module is implemented using a multilayer perceptron. The multilayer perceptron has a network architecture with hidden layers of [128, 64] and the activation function is Tanh, as shown below. Figure 3 As shown.

[0034] Step 5, convert the current feature vector Input behavior parameter adjustment strategy module to obtain behavior parameter adjustment amount The desired foot contact of the quadruped robot, output by the behavior parameter adjustment strategy module, is defined as follows: The desired contact state of the right foreleg foot is The desired contact state of the left foreleg foot is The desired contact state of the right hind leg is as follows: The desired contact state of the left hind leg is as follows: .when At that time, it indicates the expected number of the behavioral strategy. The legs are for the aerial event, when At that time, it indicates the expected number of the behavioral strategy. The legs serve as the supporting element. This represents the increment of the step height. In this example, the behavior parameter adjustment strategy is implemented by a multilayer perceptron, with a network architecture of [64, 128, 128, 64] hidden layers and ReLU activation function.

[0035] Step 6, convert the current feature vector Input the tipping risk assessment module to obtain the current probability of the robot falling over. In this example, the overturning risk assessment module consists of two LSTM layers. A one-dimensional convolutional layer is added before the first LSTM layer. Residual connections are introduced between the layers. A fully connected layer with a sigmoid activation function is used as the output layer, directly outputting a scalar value between 0 and 1, which is the probability of falling.

[0036] Step 7: Input the fall probability into a switching function. This switching function controls the switching between two drive modes: "MPC standalone control" and "MPC-DRL collaborative control." In this example, the switching function... for:

[0037]

[0038] in, To switch the threshold, in this example .when This indicates that the disturbance caused by the current terrain change poses a significant risk of the robot falling over. This indicates that the DRL-based strategy intervenes in the control loop, adapting to disturbances caused by terrain changes by influencing the gait and step height of the MPC-based controller, thus enabling the quadruped robot's motion to enter an "MPC-DRL joint control" mode. When the terrain is relatively flat and the robot's movement is stable, this indicates that the terrain is relatively flat and the robot's movement is stable. This indicates that the DRL-based strategy does not affect the robot's motion, and the quadruped robot's motion enters the "MPC individual control" mode.

[0039] Step 8: Input the user speed command and behavior parameter adjustment into the model predictive controller, wherein the user speed command is defined as... , In the fuselage coordinate system Desired linear velocity in the axial direction, In the fuselage coordinate system Desired linear velocity in the axial direction, For the fuselage The rotational angular velocity of the shaft. The centroid reference trajectory generation module generates the trajectory based on the user's velocity command. The reference center of mass position is generated based on the spring-loaded inverted pendulum model. Reference centroid velocity Reference centroid attitude angle angular velocity relative to reference center of mass Together, they form the reference trajectory of the center of mass.

[0040] Step 9, the gait planning module calculates the reference centroid velocity. With preset stride ,pass Calculate the step frequency. For step frequency, For reference center of mass velocity of Axial direction component. In this example, m, the basic gait adopted is a diagonal gait with a duty cycle of 0.5, based on the gait cycle. A gait cycle table can be created, and the reference leg state at the current moment can be obtained by looking up the table. The reference foot contact state of the right foreleg is as follows: The reference foot contact state of the left foreleg is as follows: The reference foot contact state of the right hind leg is as follows: The reference foot contact state of the left hind leg is as follows: .when When, it indicates the number. The legs are for the aerial event, when When, it indicates the number. The legs serve as the supporting element. This is achieved through adjustments with behavioral parameters. To perform an OR operation between elements in the same position, this example uses " "This indicates that the final foot contact state has been obtained." :

[0041]

[0042] in, The final contact state of the right foreleg foot is The final contact state of the left foreleg's foot is The final contact state of the right hind leg's foot is The final contact state of the left hind leg's foot is .when When, it indicates the number. The legs are for the aerial event, when When, it indicates the number. The legs serve as support.

[0043] Step 10: Based on the centroid reference trajectory and the final foot contact state The desired foot force of the supporting leg is calculated using the MPC solver. In this example, the controlled state of the quadruped robot is set. The variation of this variation with control input can be expressed as a control-oriented model, namely: , among which is It's the kinematics part. This is the state matrix, which describes the transformation relationships between differentials; It is the dynamics part. The control matrix describes the influence of forces on the rigid body's motion state. To control the input, the dynamic model is expanded as follows:

[0044]

[0045] In the formula Indicates yaw angle The rotation matrix, This indicates that all elements are 0 and the size is 0. The matrix, This indicates that all elements are 0 and the size is 0. The matrix, This indicates that all elements are 1 and the size is 1. The matrix, It is a 3D identity matrix. Indicates the distance from the centroid to the number. The vector at the foot of the leg, Represented by vectors Antisymmetric matrix constructed from elements To provide sufficient support for the expected foundation of the supporting items, This is the inverse of the moment of inertia of the fuselage in the world coordinate system. For robot quality, This is the vector of gravitational acceleration.

[0046] Step 11: Calculate the expected plantar force of the supporting leg using the MPC solver. , for middle The leg numbers are used to convert the plantar force of each supporting leg into joint torque of each joint of the supporting leg through the Jacobian matrix.

[0047] Step 12, based on the final foot contact state Increment of step height Construct the foot trajectory of the swing leg when When, it indicates the number. The leg is the swinging term. In this example, the output is achieved through a behavior parameter strategy. To increase the oscillation trajectory in the world coordinate system The height in the axial direction is determined by the speed command. fuselage position feedback Airframe linear velocity feedback and gait cycle It can calculate the world coordinate system. Point of contact of the foot of the swinging leg in the axial direction :

[0048]

[0049] In the formula This represents the position of the hip joint of the swinging leg in the world coordinate system. For the fuselage in The actual velocity component in the axial direction, For the speed deviation adjustment amount, in this example, To reduce the impact between the foot and the ground upon landing, the foot swing trajectory is typically designed to ensure that the longitudinal velocity is zero at landing, takeoff, and the highest point. Simultaneously, it is desirable that the position and velocity curves of the foot motion do not exhibit abrupt changes. Based on these requirements, the reference foot trajectory for the swing phase is planned as follows:

[0050]

[0051]

[0052] In the formula For the foot of the oscillating term in the world coordinate system Position components along the axial direction For the foot of the oscillating term in the world coordinate system Position components on the axis, For the foot of the oscillating term in the world coordinate system Initial position along the axis, For the initial moment of the swing Velocity components in the axial direction In the world coordinate system Position components along the axial direction For time, As a preset step height, in this example, =0.1m.

[0053] Step 13, based on the position of the foot and The joint positions and angular velocities of the thigh and lower leg joints in the swing term are obtained by inverse kinematics calculation.

[0054] Step 14: Driven by the torque of the supporting neck joint and the position of the swinging neck joint, the quadruped robot moves in accordance with the user's speed command.

[0055] This invention discloses a dual-drive, self-switching, highly adaptive motion control method for quadruped robots. This method estimates current terrain features (slope, roughness, continuity, etc.) using body state feedback, IMU, and foot force sensors without relying on external sensors. Based on these terrain features, it predicts the probability of falling, enabling the quadruped robot to intelligently switch between two drive modes: "MPC solo control (no risk of falling)" and "MPC-DRL cooperative control (risk of falling)." This method efficiently integrates the optimization efficiency and smoothness of MPC with the disturbance adaptation capability of DRL, achieving highly robust, efficient, and adaptive motion capabilities for the quadruped robot in diverse terrains.

Claims

1. A motion control method for a quadruped robot with dual-drive self-switching and strong adaptability, characterized in that, include: Step 1: Acquire joint motor, IMU, and foot force data of the quadruped robot; Step 2: Use IMU data to perform pre-integration to obtain the robot's center of mass pose at the current moment; Input the joint data and IMU data into the state estimator to obtain the robot's center of mass position and linear velocity in the world coordinate system, and the robot's three-axis rotational angular velocities in the body coordinate system. Step 3: The robot's center of mass posture, center of mass position, center of mass linear velocity, rotational angular velocity, and foot force are combined into a one-dimensional vector, which is defined as the state in the Markov decision process; Step 4: Input the current state into the terrain feature extraction module to obtain the feature vector of the quadruped robot's stability margin; Step 5: Input the current feature vector into the behavior parameter adjustment strategy module to obtain the behavior parameter adjustment amount, including the increment of the quadruped robot's expected foot contact and step height. Step 6: Input the current feature vector into the overturning risk assessment module to obtain the current robot fall probability; Step 7: Input the fall probability into a switching function, and use the switching function to control the switching between the two driving modes of "MPC individual control" and "MPC-DRL cooperative control"; Step 8: Input the user speed command and behavior parameter adjustment into the model predictive controller. Based on the user speed command, generate the reference center of mass position, reference center of mass velocity, reference center of mass attitude angle and reference center of mass angular velocity based on the spring-loaded inverted pendulum model, which together constitute the center of mass reference trajectory. Step 9, based on the reference centroid velocity The stride frequency is calculated by comparing the stride with the preset stride length. A gait cycle table is established based on the gait cycle. The reference leg state at the current moment can be obtained by looking up the table. The final foot contact state is obtained by performing an "OR" operation between the corresponding elements of the behavior parameter adjustment. Step 10: Based on the centroid reference trajectory and the final foot contact state, use the MPC solver to calculate the expected plantar force of the supporting leg. Step 11: Based on the expected plantar force of the supporting leg calculated by the MPC solver, the plantar force of each supporting leg is converted into the joint torque of each joint of the supporting leg through the Jacobian matrix; Step 12, based on the final foot contact state Increment of step height Construct the foot trajectory of the swing leg to obtain the foot trajectory of the swing term in the world coordinate system. shaft and Position component along the axis; Step 13, based on the foot of the oscillating term in the world coordinate system shaft and The position component along the axis is obtained by inverse kinematics calculation to obtain the joint position and joint angular velocity of the thigh and lower leg joints in the swing term; Step 14: Driven by the torque of the supporting neck joint and the position of the swinging neck joint, the quadruped robot moves in accordance with the user's speed command.

2. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, The foot of the oscillating term in the world coordinate system shaft and The position components along the axial direction are as follows: ； ； In the formula For the foot of the oscillating term in the world coordinate system Position components along the axial direction For the foot of the oscillating term in the world coordinate system Position components on the axis, For the fuselage in The actual velocity component in the axial direction, For the foot of the oscillating term in the world coordinate system Initial position along the axis, For the initial moment of the swing Velocity components in the axial direction In the world coordinate system Position components along the axial direction The preset step height, Let T be the gait period, and T be the time period. In the world coordinate system The point of contact of the foot of the swinging leg in the axial direction. This is the increment of the step height.

3. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 2, characterized in that, World coordinate system Point of contact of the foot of the swinging leg in the axial direction for: ； in For the hip of the swing leg in the world coordinate system In the fuselage coordinate system Desired linear velocity joint position in the axial direction. For the fuselage in The actual velocity component in the axial direction, This is the speed deviation adjustment amount.

4. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, The final foot contact state is as follows: ； ； ； in The reference leg state at the current moment is obtained by looking up a table; the reference foot contact state of the right foreleg is as follows: The reference foot contact state of the left foreleg is as follows: The reference foot contact state of the right hind leg is as follows: The reference foot contact state of the left hind leg is as follows: ,when When, it indicates the number. The legs are for the airborne event, when When, it indicates the number. The legs are the supporting element; C π The desired foot contact of the quadruped robot output by the parameter adjustment strategy module is as follows: The desired foot contact state of the right foreleg is... The desired contact state of the left foreleg foot is The desired foot contact state of the right hind leg is The desired contact state of the left hind leg is as follows: ,when At that time, it indicates the expected number of the behavioral strategy. The legs are for the aerial event, when At that time, it indicates the expected number of the behavioral strategy. The legs serve as the support.

5. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, Gait cycle , ; For step frequency, For reference center of mass velocity of Axial component, This is the preset stride length.

6. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, The switching function is: ； in, To switch the threshold, when the robot falls... When this occurs, it indicates that the DRL-based strategy intervenes in the control loop, adapting to the disturbances caused by terrain changes by influencing the gait and step height of the MPC-based controller; the quadruped robot's movement enters the "MPC-DRL joint control" mode; when When this occurs, it indicates that the DRL-based strategy does not affect the robot's movement, and the quadruped robot's movement enters the "MPC individual control" mode.

7. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 6, characterized in that, The overturning risk assessment module consists of two LSTM layers. A one-dimensional convolutional layer is added before the first LSTM layer. Residual connections are introduced between the layers. A fully connected layer with a sigmoid activation function is used as the output layer, which directly outputs a scalar value between 0 and 1, i.e., the probability of falling.

8. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, The parameter adjustment strategy module is implemented by a multilayer perceptron. The hidden layer of the multilayer perceptron is a network architecture of [64, 128, 128, 64], and the activation function of the network is ReLU.

9. The dual-drive self-switching strongly adaptive quadruped robot motion control method according to claim 1, characterized in that, The terrain feature extraction module is implemented by a multilayer perceptron, with a hidden layer network architecture of [128, 64] and the activation function of the network being Tanh.

Images