Wheel-legged hybrid robot motion self-adaptive control method and system

Abstract

The application provides a wheel-leg composite robot motion adaptive control method and system, relates to the technical field of robot control, and comprises the following steps: constructing a multi-mode dynamics model and performing sensor parameter calibration; performing accurate sensing of a terrain, an obstacle and ground adhesion characteristics, dynamic obstacle trajectory prediction and terrain mutation sensing; presetting a multi-parameter fusion switching threshold value containing single-wheel, single-leg and wheel-leg combined three motion modes; formulating a dynamic switching decision; constructing an adaptive control model, performing detailed control of the three motion modes and mode switching transition control; and performing multi-source error collaborative compensation on errors in the three motion modes. Through the combination of multi-mode dynamics modeling and cross-mode fusion sensing, the application accurately captures robot dynamics characteristics and complex environment parameters, greatly enhances the adaptability of the robot to diversified terrains such as flat roads, slopes and steps and dynamic obstacles, and realizes rapid identification of terrain mutations and dynamic obstacle trajectory prediction.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, and in particular to a method and system for adaptive motion control of a wheel-legged composite robot. Background Technology

[0002] As human exploration and operations continue to expand into complex terrain environments, scenarios such as field exploration, disaster relief, and military reconnaissance place higher demands on the environmental adaptability of robots, driving the evolution of mobile robot configurations from single-mode to composite modes. Wheeled and legged mobility technologies, as the two core technological foundations of ground mobile robots, have formed mature technological systems. On this basis, to overcome the limitations of single-mode configurations in terrain adaptability, the wheel-legged composite robot configuration has emerged. Its technological development relies on the cross-integration of multiple disciplines such as mechanical structure design, multi-degree-of-freedom drive, and sensor fusion, representing an integration and innovation of traditional mobile robot technologies.

[0003] Wheel-legged hybrid robots, possessing multiple movement modes such as wheeled rolling and legged walking, and needing to adapt to switching between movement states in different terrains, face challenges in motion control, including multi-mode coordination, complex system dynamics, and uncertainties in environmental disturbances. Against this backdrop, motion adaptive control technology, based on the development of intelligent control theory and supported by cutting-edge technologies such as GPU / TPU computing power enhancement, 5G / TSN network collaboration, and digital twin simulation, has gradually become a core research direction in the field of wheel-legged hybrid robot control. It aims to achieve real-time perception of environmental changes and dynamic adjustment of control strategies to address system uncertainties. Its technological formation and development have been intertwined with the collaborative evolution of technologies in multiple fields, including motion control, artificial intelligence, and sensing detection.

[0004] Existing motion control for wheeled-legged hybrid robots mostly employs preset trajectory optimization or MPC to achieve mode switching. The switching time is usually long, and the switching modes lack wheel-legged collaborative modes. Motion mode switching decisions are mostly based on fixed threshold control, lacking a dynamic optimization mechanism based on reinforcement learning. The posture fluctuations are large and the impact is obvious during mode switching. It is impossible to balance the efficiency of wheeled motion with the terrain adaptability of legged motion, and there is even a risk of switching failure. Summary of the Invention

[0005] This invention provides a motion adaptive control method and system for a wheel-legged hybrid robot, which solves the shortcomings of existing technologies that cannot simultaneously achieve the high efficiency of wheeled motion and the terrain adaptability of legged motion.

[0006] On one hand, the present invention provides a motion adaptive control method for a wheel-legged hybrid robot, comprising:

[0007] S1: Based on the structural parameters of the wheel-leg hybrid robot, a multi-mode dynamic model is constructed and sensor parameters are calibrated to obtain a unified dynamic equation and initial control parameters;

[0008] S2: Based on a cross-modal fusion perception framework, it collects and processes synchronous data from multiple sensors to accurately perceive terrain, obstacles, and ground adhesion characteristics, predict dynamic obstacle trajectories, and perceive terrain abrupt changes, outputting perception results containing multi-dimensional ground characteristic parameters.

[0009] S3: Based on structural parameters and perception results, preset multi-parameter fusion switching thresholds including three motion modes: single wheel, single leg, and wheel-leg combination;

[0010] S4: Based on the perception results and the multi-parameter fusion switching threshold, formulate dynamic switching decisions based on reinforcement learning, and output the mode switching timing and dynamic weight coefficients;

[0011] S5: A three-stage coupled control strategy is adopted, and an adaptive control model is constructed by combining mode switching timing, dynamic weight coefficients and unified dynamic equations to perform detailed control of three motion modes and transition control between modes.

[0012] S6: Based on the attention mechanism, combined with real-time mapping data from the digital twin and initialization control parameters, multi-source error collaborative compensation is performed on the errors in the three motion modes, and error compensation commands are output and fed back to the adaptive control model.

[0013] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, in step S1, the unified dynamic equation is expressed as follows:

[0014]

[0015] Where M(q) is the inertia matrix, Let G(q) be the Coriolis force and centrifugal force matrix, G(q) be the gravity term vector, and q be the joint angle vector. The joint angular velocity vector. Let τ be the joint angular acceleration vector, τ be the joint torque vector output by the controller, J(q) be the Jacobian matrix, and F be the joint torque vector. e Let T be the contact force vector between the wheel leg and the ground, and T is the transpose.

[0016] According to the motion adaptive control method for a wheel-legged hybrid robot provided by the present invention, step S2, the step of outputting the perception result containing multi-dimensional ground characteristic parameters, includes:

[0017] S21: Based on the cross-modal fusion sensing framework, data from lidar, vision sensor, IMU and contact force sensor are synchronously acquired through time synchronization to obtain raw sensing data;

[0018] S22: Preprocess the raw sensing data to obtain processed sensing data;

[0019] S23: A multimodal fusion perception algorithm is adopted to fuse visual image features, lidar 3D terrain features and IMU attitude features in the processed perception data, and to classify and identify terrain types through semantic segmentation and obtain terrain geometric parameters through 3D reconstruction.

[0020] S24: Calculate the normal and tangential forces on the ground based on contact force sensor data, and obtain ground hardness parameters through pressure distribution detection;

[0021] S25: LSTM neural network is used to learn from the historical trajectory data of dynamic obstacles to predict their trajectory; based on terrain geometry parameters and ground hardness parameters, a sliding window detection algorithm is used to monitor changes in terrain parameters in real time and identify areas and degrees of terrain change.

[0022] S26: Based on the obstacle movement trajectory and the terrain change area and degree of change, establish a state-environment coupled perception model, correlate and analyze robot posture fluctuations, wheel and leg contact force data and terrain texture features, motion response parameters, correct robot state error and terrain assessment error, and output perception results containing multi-dimensional ground characteristic parameters.

[0023] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S3, the step of presetting the multi-parameter fusion switching threshold, includes:

[0024] S31: Physical constraint parameters for determining motion modes based on robot structural parameters;

[0025] S32: Combine multi-dimensional ground characteristic parameters to extract key perception parameters that affect motion mode switching;

[0026] S33: The weights of physical constraint parameters and key sensing parameters are determined by the analytic hierarchy process (AHP), and a multi-parameter fusion evaluation function is constructed.

[0027] S34: Through a robot simulation platform, conduct multiple sets of motion tests in typical terrain scenarios to obtain the optimal motion performance data of three motion modes under different scenarios;

[0028] S35: Based on the multi-parameter fusion evaluation function and optimal motion performance data, determine the multi-parameter fusion switching threshold under different mode switching scenarios and form a threshold matrix.

[0029] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S4, the step of formulating dynamic switching decisions based on reinforcement learning, includes:

[0030] S41: Construct a reinforcement learning agent, define the action space, and set the reward function;

[0031] S42: The reinforcement learning agent is trained using a deep deterministic policy gradient algorithm. Combined with a multi-parameter fusion switching threshold, the action decision-making strategy is optimized through joint training in a robot simulation environment and a physical testing environment to obtain a mode-switching agent.

[0032] S43: Input the current state space data in real time, and output the mode switching timing judgment result through the mode switching agent. When the state variable meets the switching threshold, output the motion mode switching command and dynamic weight coefficient.

[0033] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S5, the step of constructing the adaptive control model includes:

[0034] S51: Determine the core stages of the three-stage coupling control strategy, including the pre-switching preparation stage, the mode switching execution stage, and the post-switching stabilization stage;

[0035] S52: In the pre-switching preparation stage, based on the mode switching timing judgment result and dynamic weight coefficient, the control parameters of the current motion mode are pre-adjusted, and the initial control quantity of the target motion mode is calculated according to the unified dynamic equation.

[0036] S53: Based on the initial control input, during the mode switching execution phase, a dynamic weighted coupling control law is constructed, and the dynamic weight coefficients are integrated into the control model.

[0037] S54: During the stabilization phase after switching, the robot's motion state data is collected in real time and compared with the predicted values ​​of the unified dynamic equation to adjust the control parameters.

[0038] S55: An adaptive control model is constructed based on a three-stage coupled control strategy and a unified dynamic equation.

[0039] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S52, the step of pre-adjusting the control parameters of the current motion mode, includes:

[0040] S521: Calculate the adjustment amount of the control parameters based on the dynamic weighting coefficient, and use the exponential smoothing method to gradually adjust the control parameters;

[0041] S523: Based on the unified dynamic equation, substitute the dynamic parameters of the target motion mode to calculate the initial joint angle, angular velocity and torque command of the target motion mode as the initial control quantity;

[0042] S524: Perform physical constraint verification on the initial control quantity to ensure that the initial control quantity does not exceed the torque, speed and angle limits of the robot joint. If it does, perform trimming and adjustment.

[0043] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S53, the step of constructing a dynamic weighted coupling control law includes:

[0044] S531: Set the general form of the dynamic weighted coupling control law to obtain the general control law;

[0045] S532: For single-wheel motion mode, a model predictive control algorithm is used for speed and steering control;

[0046] S533: For single-leg movement mode, a fuzzy PID algorithm is used for joint angle tracking control;

[0047] S534: For the wheel-leg combined motion mode, a coupled PID algorithm is used for wheel-leg coordinated motion control;

[0048] S535: Substitute the control torque of the three modes into the general control law, and combine it with the dynamic weighting coefficient to obtain the final control torque output.

[0049] According to the adaptive motion control method for a wheel-legged hybrid robot provided by the present invention, step S6, the step of performing multi-source error collaborative compensation includes:

[0050] S61: Analyze the sources of multi-source errors and output the error analysis results;

[0051] S62: Based on the error analysis results, compare the real-time mapping data of the digital twin with the predicted output data of the adaptive control model to extract the error characteristics of the error source;

[0052] S63: Weighting the error features of the error source through an attention mechanism;

[0053] S64: Combine the initial control parameters to establish an error prediction model, predict the error change trend based on historical error data and error characteristics of error sources, and generate multi-source error collaborative compensation quantity;

[0054] S65: Convert the multi-source error collaborative compensation quantity into error compensation command and feed it back to the adaptive control model.

[0055] This invention also provides a motion adaptive control system for a wheel-legged hybrid robot, comprising: a dynamic modeling module, a parameter initialization module, a cross-modal fusion sensing module, a mode switching threshold preset module, a dynamic decision-making module, a three-stage coupled control module, and a multi-source error compensation module; the dynamic modeling module is used to construct a multi-mode dynamic model based on the structural parameters of the wheel-legged hybrid robot; the parameter initialization module is used to calibrate sensor parameters based on the structural parameters of the wheel-legged hybrid robot; the cross-modal fusion sensing module is used to collect and process multi-sensor synchronous data based on the cross-modal fusion sensing framework, to accurately perceive terrain, obstacles, and ground adhesion characteristics, and to predict dynamic obstacle trajectories and perceive terrain abrupt changes, outputting perception results containing multi-dimensional ground characteristic parameters; the mode switching threshold preset module... The system is designed to perform multi-parameter fusion switching based on structural parameters and perception results, pre-setting thresholds for three motion modes: single-wheel, single-leg, and wheel-leg combination. A dynamic decision-making module, based on perception results and multi-parameter fusion switching thresholds, formulates dynamic switching decisions using reinforcement learning, outputting the mode switching timing and dynamic weight coefficients. A three-stage coupled control module employs a three-stage coupled control strategy, combining mode switching timing, dynamic weight coefficients, and a unified dynamic equation to construct an adaptive control model, performing detailed control of the three motion modes and transitional control between modes. A multi-source error compensation module, based on an attention mechanism and combining real-time digital twin mapping data and initial control parameters, performs multi-source error collaborative compensation for errors in the three motion modes, outputting error compensation commands and feeding them back to the adaptive control model.

[0056] This invention provides a motion adaptive control method and system for a wheel-legged hybrid robot. By combining multi-mode dynamic modeling with cross-modal fusion perception, it accurately captures the robot's dynamic characteristics and complex environmental parameters, significantly enhancing the robot's adaptability to diverse terrains and dynamic obstacles such as flat surfaces, slopes, and steps. It enables rapid identification of terrain changes and prediction of dynamic obstacle trajectories. Through dynamic switching decision-making and a three-stage coupled control strategy, it ensures the smoothness of switching between single-wheel, single-leg, and wheel-legged hybrid modes, effectively reducing attitude fluctuations and impacts during the switching process, and balancing the efficiency of wheeled motion with the terrain adaptability of legged motion. By integrating attention mechanisms and digital twin multi-source error collaborative compensation mechanisms, it accurately locates and suppresses multi-source interference such as model errors and sensor errors, significantly improving control accuracy and motion stability. At the same time, physical constraint verification ensures hardware operation safety. Attached Figure Description

[0057] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0058] Figure 1 This is a flowchart of a motion adaptive control method for a wheel-legged composite robot provided in an embodiment of the present invention;

[0059] Figure 2 This is a flowchart of constructing an adaptive control model provided in an embodiment of the present invention;

[0060] Figure 3 This is a schematic diagram of the motion adaptive control system for a wheel-legged composite robot provided in an embodiment of the present invention. Detailed Implementation

[0061] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0062] Example:

[0063] The following is combined Figures 1-3 This invention describes a motion adaptive control method and system for a wheel-legged composite robot.

[0064] like Figure 1-2 As shown in the figure, an adaptive motion control method for a wheel-legged hybrid robot provided by an embodiment of the present invention includes:

[0065] S1: Based on the structural parameters of the wheel-leg hybrid robot, a multi-mode dynamic model is constructed and sensor parameters are calibrated to obtain a unified dynamic equation and initial control parameters.

[0066] In step S1, the steps for obtaining the unified dynamic equations include:

[0067] First, the system's kinetic energy is calculated. The kinetic energy of the wheel-legged hybrid robot is the sum of the translational and rotational kinetic energies of each rigid body, including the main body, wheel assembly, and leg joint components. The formula is as follows:

[0068]

[0069] Where S is the system kinetic energy, M(q) is the inertia matrix (n×n), whose elements are related to the mass, moment of inertia, and joint angles of each component of the robot, reflecting the change in the system's motion inertia with attitude. q is the joint angle vector. Let M(q) be the joint angular velocity vector. Due to the coupling characteristics of the wheel-leg composite structure, M(q) is a symmetric positive definite matrix and includes a nonlinear term for the joint angle q. T is the transpose.

[0070] Next, the system potential energy is calculated. The system potential energy is mainly gravitational potential energy, which is determined by the height of the center of mass of each rigid body, and is expressed by the formula:

[0071]

[0072] Where V is the system potential energy, and G(q) is the gravity term vector (n×1), whose elements are related to the mass of each component, gravitational acceleration, and center of mass coordinates, and are dynamically adjusted as the joint angle q changes.

[0073] Then, the Lagrange function is defined as the difference between the system's kinetic energy and its potential energy, expressed by the formula:

[0074]

[0075]

[0076] Where, τ net Let n be the generalized external torque in the joint space, including the control torque output by the controller and external disturbance torques. Taking the partial derivatives of the Lagrangian function and substituting them into the equation for expansion, we get:

[0077]

[0078]

[0079]

[0080] in, The first derivative of the inertia matrix with respect to time reflects its dynamic change over time. Because joint angles change with time during robot motion, the robot's inertia matrix is ​​not a constant value but rather dynamically adjusts over time. t is the time variable. It is the first derivative operator with respect to time t. This represents the joint angular acceleration vector. Let be the partial derivative tensor of the inertia matrix with respect to each joint angle, representing the rate of change of the inertia matrix with respect to the joint angle q. By rearranging, the Coriolis force and centrifugal force matrices can be obtained, satisfying:

[0081]

[0082]

[0083] in, Let be the Coriolis force and centrifugal force matrices. Substituting the calculation results into the Lagrange equations, we obtain the basic dynamic equations without considering ground contact forces, expressed as:

[0084]

[0085] Where τ is the joint torque vector (n×1) output by the controller, which is the active control torque term in the basic dynamic equation.

[0086] During the movement of a wheel-legged hybrid robot, the contact force between the wheels and the ground generates additional torques on the joints, which need to be coupled into the fundamental dynamic equations. A mapping F between the operational space (ground contact point) and the joint space is established using the Jacobian matrix. e =J(q)τ e , of which F e τ is the contact force vector between the wheel leg and the ground (m×1, where m is the degree of freedom of the contact point, usually m=3, corresponding to three-dimensional force in space); J(q) is the Jacobian matrix (m×n), describing the linear mapping relationship between small changes in joint angle and small changes in contact point position; e Let be the generalized torque (n×1) generated by the contact force in the joint space. Taking the transpose of the above mapping relationship, we can obtain the generalized joint torque τ corresponding to the contact force. e =J T (q)F. The generalized torque τ generated by the contact force. e By incorporating the fundamental dynamic equations and considering the combined effect of the control torque τ and the contact force coupling torque, the unified dynamic equations for the wheel-legged hybrid robot are finally obtained, expressed as follows:

[0087]

[0088] in, This is the joint angular acceleration vector.

[0089] Sensor parameter calibration is performed, mainly for joint encoders, IMUs, and contact force sensors. The steps include:

[0090] The robot joints are fixed at multiple known angular positions. The encoder output values ​​are collected, and a linear regression model between the encoder output values ​​and the actual angles is established. The angular position zero bias and scaling factor are solved to calibrate the joint encoder. The calibration formula is expressed as:

[0091]

[0092] Where, q raw For the original output, qcal For the calibrated angle, b q For zero angular position, k q This is a scaling factor.

[0093] The IMU was fixed in a constant temperature chamber and zero-bias data was collected under static conditions. The zero-bias of the accelerometer and gyroscope were calculated. The accelerometer sensitivity was calibrated using a six-position method (the IMU was placed facing up, down, forward, backward, left, and right respectively). The gyroscope sensitivity was calibrated by setting different angular velocities using a rate turntable. The calibration formula is as follows:

[0094]

[0095]

[0096] in, The calibrated acceleration value. For accelerometer sensitivity, This is the raw output value of the accelerometer. To achieve zero bias in the accelerometer, The calibrated angular velocity value, k g For gyroscope sensitivity, b is the original output value of the gyroscope. g The gyroscope has zero bias.

[0097] A force of known magnitude and direction is applied to the sensor using a high-precision force gauge. The sensor's original output voltage is collected, and a linear mapping model between force and voltage is established. The sensitivity and zero bias of the contact force sensor are then calculated. The calibration formula is expressed as follows:

[0098]

[0099] in, To determine the contact force value after calibration, k F For the sensitivity of the contact force sensor, b is the original output voltage of the contact force sensor. F This is for zero bias of the contact force sensor.

[0100] After calibration, the calibration parameters are stored in the controller as initialization control parameters for subsequent data acquisition and control calculations. Zero bias refers to the sensor's core system error parameter, which is the non-zero inherent deviation value output by the sensor when there is no target input; it is a fixed error caused by the sensor's hardware characteristics.

[0101] S2: Based on a cross-modal fusion perception framework, it collects and processes synchronous data from multiple sensors to accurately perceive terrain, obstacles, and ground adhesion characteristics, predict dynamic obstacle trajectories, and perceive terrain abrupt changes. It outputs perception results containing multi-dimensional ground characteristic parameters. The cross-modal fusion perception framework is a hardware-software integrated framework that integrates perception information from different modalities such as vision, laser, inertial, and force sensors. Through data-level, feature-level, or decision-level fusion, it improves the accuracy and robustness of environmental perception. Multi-dimensional ground characteristic parameters include a comprehensive set of parameters such as terrain type, terrain geometry, ground adhesion characteristics, and obstacle parameters.

[0102] Step S2, which outputs the sensing results containing multi-dimensional ground characteristic parameters, includes:

[0103] S21: Based on the cross-modal fusion sensing framework, data from lidar, vision sensors, IMU, and contact force sensors are synchronously acquired through time synchronization to obtain raw sensing data.

[0104] S22: Preprocess the raw sensory data to obtain processed sensory data. The preprocessing steps include:

[0105] Statistical filtering algorithms are used to remove outliers; voxel mesh filtering is used to reduce the number of point clouds; and point cloud data is transformed from the lidar coordinate system to the robot body coordinate system based on the extrinsic parameters of the lidar and the robot body.

[0106] Gaussian filtering is used to remove noise, and histogram equalization is performed to enhance contrast. A hole-filling algorithm is used to fill in missing depth values ​​and remove invalid depth points that are outside the range of the lidar. Based on extrinsic parameters, the depth image is aligned with the RGB image to generate a color point cloud.

[0107] The calibrated accelerometer and gyroscope biases are subtracted; high-frequency noise is removed using Kalman filtering; and the Mahony complementary filtering algorithm based on quaternions is used to fuse acceleration and angular velocity data to calculate the robot's real-time attitude. The Mahony complementary filtering algorithm is a lightweight attitude calculation filtering algorithm, an improved form of complementary filtering. Its core function is to fuse IMU acceleration and angular velocity data to accurately calculate the real-time roll, pitch, and yaw angles of the wheel-legged hybrid robot, while suppressing sensor noise and drift errors.

[0108] Subtract the calibrated contact force sensor zero bias; use a Butterworth low-pass filter to remove high-frequency vibration noise; map the force data to the range [-1,1]. The Butterworth low-pass filter is a linear low-pass filter with the largest flat amplitude-frequency response. Its core function is to allow low-frequency signals below the cutoff frequency to pass through, attenuate high-frequency signals above the cutoff frequency, and make the signal amplitude response as flat as possible within the passband without ripple distortion, effectively suppressing high-frequency noise.

[0109] S23: A multimodal fusion sensing algorithm is employed to fuse visual image features, LiDAR 3D terrain features, and IMU attitude features from the processed sensing data. Terrain type classification and recognition are performed through semantic segmentation, and terrain geometric parameters are obtained through 3D reconstruction. Visual image features are extracted from deep semantic features of RGB images using a ResNet-50 network; LiDAR features are extracted from downsampled point clouds, including geometric features such as normal vectors, curvature, and roughness; and IMU attitude features are extracted from roll angle, pitch angle, and their rate of change.

[0110] The three features are fused using an attention mechanism, as expressed in the formula:

[0111]

[0112]

[0113] Among them, F fusion The fused feature is F v For visual features, F l For lidar characteristics, F i The IMU pose features are represented by α, β, and γ, which are the attention weights of the three features, obtained through training. They are usually set to α=0.6, β=0.3, and γ=0.1.

[0114] The fused features are input into the Softmax classifier, which outputs the terrain type probability. The type with the highest probability is taken as the classification result.

[0115] The steps to obtain terrain geometry parameters include:

[0116] The angle θ between the normal vector of the lidar point cloud and the vertical direction of the ground is calculated using the following formula:

[0117]

[0118] Where n is the point cloud normal vector and z0 is the vertical ground unit vector.

[0119] The formula for calculating the maximum height difference of point clouds within a sliding window is as follows:

[0120]

[0121] Among them, z i Let i be the vertical coordinate of point i within the window, i.e., the height coordinate.

[0122] The RANSAC algorithm is used to fit the step plane, and the height difference and horizontal distance between the upper and lower planes are calculated. The steps include:

[0123] Based on the terrain change areas detected by the sliding window, the corresponding range of lidar point cloud data is extracted as candidate point clouds, and obvious outliers are removed.

[0124] Set the number of iterations (1000 times), the distance threshold from the point to the plane (0.02m), and the interior point ratio threshold (0.6).

[0125] Randomly select 3 non-collinear points from the candidate point cloud to construct an initial plane. Calculate the distance from all candidate points to this plane and count the number of interior points whose distance is less than the distance threshold from the point to the plane. Repeat the iterative process and retain the plane with the most interior points as the lower plane of the step.

[0126] Remove the interior points that have been matched to the lower plane, and repeat the RANSAC fitting process on the remaining point cloud to obtain the upper plane of the step.

[0127] Based on the upper and lower planes, calculate the angle between the normal vectors of the two planes to verify whether they are parallel planes. The upper and lower planes of the step are approximately parallel. Calculate the vertical distance between the two planes through spatial geometric relationships, which is the height of the step. On the horizontal projection plane, calculate the shortest horizontal distance between the projected edges of the two planes, which is the width of the step.

[0128] S24: Calculate the ground normal and tangential forces based on contact force sensor data, and obtain ground hardness parameters through pressure distribution detection. The formulas for the ground normal and tangential forces are expressed as follows:

[0129]

[0130]

[0131]

[0132] Among them, F x F represents the component of the contact force vector along the x-axis in the robot's body coordinate system. y Let F be the component of the contact force vector along the y-axis in the robot's body coordinate system. z F represents the component of the contact force vector along the z-axis in the robot's body coordinate system. n For the ground normal force, F t This is the tangential force on the ground.

[0133] Based on the structural dimensions of the wheels / feet, the ground contact area is estimated. The contact area A during wheeled motion is represented as... Where r is the wheel radius and b is the wheel width. The contact area during leg movements is represented as... , where S is the preset contact area of ​​the foot end.

[0134] Then, the ground pressure is calculated, expressed by the formula:

[0135]

[0136] Where σ is the ground pressure.

[0137] The relationship between ground pressure and ground deformation δ was established through experiments. , where k h k is the ground hardness coefficient. h The larger the value, the harder the ground. Ground deformation δ is obtained by measuring the change in distance between the wheel / foot and the ground using lidar, and the formula is: Where d0 is the initial distance, d t This represents the real-time distance.

[0138] S25: An LSTM neural network is used to learn from historical motion trajectory data of dynamic obstacles to predict their trajectory. The steps include:

[0139] Dynamic obstacles are detected from visual images, their center coordinates are extracted, and a sequence of their historical motion trajectories is generated. Where B is the number of historical time steps, i.e. the number of trajectory points, and we take B=20.

[0140] An LSTM network is constructed, with a 2-dimensional input layer, a 64-dimensional hidden layer, and a 2-dimensional output layer. The historical trajectory sequence is used as input, and the coordinates of the next time step are used as output. The model is trained using the MSE loss function. The loss function is expressed as:

[0141]

[0142] Where Loss is the loss function, and i is the sample index. , To predict coordinates, , is the true coordinate, and N is the number of training samples.

[0143] The latest B trajectory points collected in real time are input into the trained LSTM model, which outputs the obstacle coordinates for the next 5 time steps, i.e., the predicted trajectory. .

[0144] Based on terrain geometry parameters and ground hardness parameters, a sliding window detection algorithm is used to monitor abrupt changes in terrain parameters in real time, and to identify areas and degrees of abrupt changes in terrain.

[0145] The steps involved in terrain change sensing include:

[0146] Set the sliding window length to 5 and the window sliding step size to 1.

[0147] The rate of change of terrain geometric parameters within the calculated window is used as an indicator of abrupt change, and is expressed as follows:

[0148]

[0149]

[0150] Where Δθ is the rate of change of slope, θ t Let θ be the slope value of the terrain at the current moment. t-4 Here, Δt represents the terrain slope value at four sampling times, 4Δt represents the sampling time interval, and 4Δt represents the total time span corresponding to the sliding window, i.e., the sum of the time intervals between the five sampling points. Δh represents the rate of change of undulation, h var,t h represents the terrain relief value at the current moment. var,t-4 The topographic relief values ​​are the values ​​from four sampling times ago. or At that time, it was determined to be a sudden change in terrain.

[0151] The degree of mutation is quantified by the ratio of mutation index to a threshold, expressed by the formula:

[0152]

[0153] Where D is a quantitative index of the degree of terrain abruptness, Δθ th Δh is the threshold for the rate of change of slope. th The threshold for the rate of change of fluctuation. It is a mild mutation. It is a moderate mutation. It is a severe mutation.

[0154] S26: Based on obstacle trajectory and terrain abrupt change areas and degrees of change, a state-environment coupled perception model is established. This model correlates and analyzes robot posture fluctuations, wheel-leg contact force data, terrain texture features, and motion response parameters to correct robot state errors and terrain assessment errors. The output includes perception results containing multi-dimensional ground characteristic parameters. A BP neural network is used to construct the state-environment coupled perception model. The input layer has 24 dimensions (2D for posture fluctuations, 3D for contact force fluctuations, 16D for texture features, and 3D for motion response), the hidden layer has 64 dimensions, and the output layer contains state error and terrain assessment error. The model corrects the robot state estimate and terrain parameter evaluation values ​​using the output state error and terrain assessment error, ultimately outputting multi-dimensional ground characteristic parameters, including: terrain type, slope, undulation, step height, ground hardness coefficient, obstacle prediction trajectory, terrain abrupt change areas, and degrees of change.

[0155] S3: Based on structural parameters and perception results, preset multi-parameter fusion switching thresholds including three motion modes: single wheel, single leg, and wheel-leg combination.

[0156] Step S3, the step of presetting the multi-parameter fusion switching threshold includes:

[0157] S31: Physical constraint parameters for determining motion modes based on robot structural parameters. These physical constraint parameters are motion limit parameters determined by the robot's structural design, including maximum joint angle, maximum angular velocity, maximum torque, and mechanical limits of the wheel-leg modules. They are the hardware constraints for motion mode switching. Specific parameters are shown in Table 1.

[0158] Table 1:

[0159]

[0160] S32: Combining multi-dimensional ground characteristic parameters, extract key perception parameters that affect motion mode switching, including terrain slope θ and ground hardness coefficient k. h Step height h s Obstacle speed v o The degree of terrain abruptness (D) affects the stability of wheeled motion. If the slope exceeds a certain threshold, a switch to legged motion is necessary. Ground hardness coefficient affects wheel-to-ground traction; at lower hardness, wheeled motion is prone to slipping, requiring a switch to legged motion. Step height determines whether wheeled motion can traverse steps exceeding the wheel radius, thus determining whether a switch to legged or a combination of wheeled and legged motion is needed. Obstacle speed determines whether a more flexible motion mode is needed to avoid obstacles. When the degree of terrain abruptness is high, a switch to a more adaptable motion mode is required.

[0161] S33: Determine the weights of physical constraint parameters and key perception parameters using the Analytic Hierarchy Process (AHP), and construct a multi-parameter fusion evaluation function. The Analytic Hierarchy Process (AHP) is a multi-criteria decision analysis method that combines qualitative and quantitative analysis by constructing a hierarchical structure model to determine the weights of each evaluation index. The steps for constructing the hierarchical structure model are as follows:

[0162] Goal layer: Evaluation of the adaptability of motion mode switching.

[0163] Criterion layer: Physical constraint criterion (C1), terrain characteristic criterion (C2), obstacle characteristic criterion (C3).

[0164] Index layer: C1 includes ; C2 includes ; C3 includes v o .

[0165] The steps for determining the weights of each evaluation index are as follows:

[0166] Determine the relative importance of each criterion / index through expert scoring, and construct a judgment matrix, which is expressed by the formula:

[0167]

[0168] Among them, the judgment matrix P is a 3×3 matrix.

[0169] Calculate the maximum eigenvalue of the judgment matrix and its corresponding eigenvector, and obtain the weight vector after normalization.

[0170] Calculate the consistency index CI, which is expressed by the formula:

[0171]

[0172] Among them, m is the order of the judgment matrix, 1 < m ≤ 3. When CI < 0.1, the consistency of the judgment matrix is acceptable.

[0173] Construct a multi-parameter fusion evaluation function using linear weighted sum f(x) , which is expressed by the formula:

[0174]

[0175] Among them, x s is the original value of each index, is the normalized index value, mapped to [0,1], w s is the weight of each index, and s is the index index. f(x) The larger the value, the stronger the adaptability of the motion mode to the current environment.

[0176] S34: Using a robot simulation platform, conduct multiple sets of motion tests in typical terrain scenarios to obtain optimal motion performance data for three motion modes under different scenarios. Based on the robot operating system and Gazebo simulation software, build a simulation model of a wheel-legged hybrid robot, and import typical terrain scenarios, including flat roads, 30° slope grass, 20cm steps, gravel ground, and dynamic obstacle areas.

[0177] In each terrain scenario, the robot was controlled to move in single-wheel, single-leg, and wheel-leg hybrid modes, and motion performance data was collected, including average speed, energy consumption, posture stability, and obstacle avoidance success rate. Each mode was tested 10 times, and the average value was taken as the performance data for that scenario.

[0178] For each terrain scenario, the evaluation function values ​​of the three motion modes are calculated, and the mode with the largest evaluation function value is selected as the optimal motion mode in that scenario. The corresponding performance data is recorded as the optimal motion performance data.

[0179] S35: Based on the multi-parameter fusion evaluation function and optimal motion performance data, a bisection method is used to determine the multi-parameter fusion switching thresholds for different mode switching scenarios, forming a threshold matrix. For example, in a slope scenario, the switching threshold between single-wheel mode and wheel-leg combined mode is determined. ,when The single-round mode is optimal. The optimal wheel-leg combination mode is achieved; the switching threshold between wheel-leg combination mode and single-leg mode. ,when The single-leg mode is optimal.

[0180] The multi-parameter fusion evaluation function integrates physical constraint parameters and perception parameters to quantify the adaptability of different motion modes, determining the optimal motion mode and switching threshold. The threshold matrix stores the threshold parameters for switching between three motion modes under different terrain scenarios, with dimensions of 3×3×K, representing 3 modes, 3 switching directions, and K key parameters. An example is shown below:

[0181]

[0182] in, To start from the current mode h Switch to the next sports mode l The threshold vector contains .in, This is the terrain slope switching threshold, a critical slope value used to determine whether to switch motion modes. The threshold for switching the ground hardness coefficient. The threshold for switching step height. The dynamic obstacle speed switching threshold, This sets the threshold for adjusting the degree of terrain abruptness. For example, Q. 12 This is the threshold vector for switching from the current single-wheel mode to the single-leg mode. If... h and l If the values ​​are the same, then the motion mode will not be switched, i.e., Q. 11 This indicates that the single-wheel mode will continue to be used without switching. (Q) 22 Q indicates continued use of single-leg mode. 33 This indicates that the wheel-leg composite mode will continue to be used.

[0183] S4: Based on the perception results and the multi-parameter fusion switching threshold, formulate dynamic switching decisions based on reinforcement learning, and output the mode switching timing and dynamic weight coefficients.

[0184] Step S4, which involves making dynamic switching decisions based on reinforcement learning, includes the following steps:

[0185] S41: Construct a reinforcement learning agent, define the action space, and set the reward function.

[0186] The agent architecture is built around the learning objective of "optimal switching of movement modes," constructing a reinforcement learning agent architecture that includes a state input layer, a decision hidden layer, and an action output layer. The action space is the set of actions the agent can execute, including maintaining the current mode, switching to single-wheel mode, switching to single-leg mode, and switching to a combined wheel-leg mode. The reward function is the core function used to evaluate the quality of the agent's actions and is a key basis for guiding the agent to learn the optimal strategy in reinforcement learning. It outputs a reward value by quantifying the "goodness" of the state the agent is in after performing an action. The design of the reward function needs to be linked to the robot's motion performance indicators, specifically: positive rewards are given when the agent's chosen movement mode enables the robot to maintain high average speed, low energy consumption, stable posture, or successfully avoid obstacles; negative rewards are given when the movement mode causes the robot to become unstable, consume excessive energy, collide with obstacles, or fail to adapt to sudden terrain changes. Normalization is used to map the reward value to the range [-1,1] to ensure a balanced reward weight for different performance indicators, ultimately guiding the agent to learn the switching strategy of "selecting the optimal movement mode in different terrain scenarios."

[0187] S42: A deep deterministic policy gradient algorithm is used to train a reinforcement learning agent. Combined with a multi-parameter fusion switching threshold, the action decision-making strategy is optimized through joint training in a robot simulation environment and a physical testing environment, resulting in a mode-switching agent. Specifically, this includes:

[0188] By combining the ROS-Gazebo simulation environment with the physical testing environment, a training scenario library containing various typical terrains and dynamic obstacle scenarios is constructed.

[0189] By combining the preset multi-parameter fusion switching threshold, the threshold condition is used as a training constraint, and historical state-action-reward samples are input for iterative training.

[0190] The strategy parameters are initially trained and optimized in a simulation environment, and then transferred to a physical robot for fine-tuning to reduce the difference between simulation and physical operation, ultimately resulting in a mode-switching agent with strong generalization ability.

[0191] S43: Inputs current state space data in real time, outputs mode switching timing judgment results through the mode switching agent, and outputs motion mode switching instructions and dynamic weight coefficients when the state variable meets the switching threshold. Specifically, it includes:

[0192] Real-time acquisition of multi-dimensional ground characteristic parameters and robot motion state parameters constitutes the state space input.

[0193] Input real-time state space data into the mode switching agent, and output the mode switching timing judgment result when the state variable meets the preset switching threshold.

[0194] The motion mode switching command and dynamic weight coefficient are generated and passed to the subsequent control module.

[0195] S5: A three-stage coupled control strategy is adopted, which combines mode switching timing, dynamic weight coefficients and unified dynamic equations to construct an adaptive control model, and performs detailed control of three motion modes and transition control between modes to ensure smooth switching process and stable motion.

[0196] Step S5, the steps for constructing the adaptive control model include:

[0197] S51: Determine the core stages of the three-stage coupled control strategy, including the pre-switching preparation stage, the mode switching execution stage, and the post-switching stabilization stage. The staged control strategy designed to achieve smooth motion mode switching reduces attitude fluctuations and shocks during mode switching through coordinated control of the three consecutive stages of pre-switching preparation, switching execution, and post-switching stabilization, ensuring motion continuity.

[0198] S52: In the pre-switching preparation stage, based on the mode switching timing judgment result and dynamic weight coefficient, the control parameters of the current motion mode are pre-adjusted, and the initial control quantity of the target motion mode is calculated according to the unified dynamic equation.

[0199] Step S52, the step of pre-adjusting the control parameters of the current motion mode, includes:

[0200] S521: Calculate the adjustment amount of the control parameters based on the dynamic weighting coefficient, and use the exponential smoothing method to gradually adjust the control parameters to avoid motion shocks caused by sudden parameter changes.

[0201] The formula for adjusting the control parameter is expressed as follows:

[0202]

[0203] Among them, y t The adjustment amount is the control parameter, α1 is the smoothing coefficient, and x t y is the current parameter value. t-1 These are the historical adjusted parameter values.

[0204] S522: Based on the unified dynamic equation, the dynamic parameters of the target motion mode are substituted into the unified dynamic equation, that is, the inertia matrix, Coriolis force matrix, etc. are substituted into the unified dynamic equation to calculate the initial joint angle, angular velocity and torque command of the target motion mode as the initial control quantity.

[0205] S523: Perform physical constraint verification on the initial control quantity to ensure that the initial control quantity does not exceed the torque, speed and angle limits of the robot joint. If it does, perform trimming and adjustment.

[0206] S53: Based on the initial control input, during the mode switching execution phase, a dynamic weighted coupling control law is constructed, integrating the dynamic weight coefficients into the control model. The dynamic weighted coupling control law is a control law that integrates dynamic weight coefficients. By adjusting the control proportion of different motion modes through weight allocation, it achieves smooth transition and coordinated control between modes.

[0207] In step S53, the steps for constructing the dynamic weighted coupling control law include:

[0208] S531: Set the general form of the dynamic weighted coupling control law to obtain the general control law, expressed by the formula:

[0209]

[0210] in, For dynamic weighting coefficients, For the control torque of the corresponding mode, For the final control torque output, i1 is the index of the motion mode. When i1=1, it corresponds to the single wheel motion mode; when i1=2, it corresponds to the single leg motion mode; when i1=3, it corresponds to the wheel-leg combined motion mode.

[0211] S532: For single-wheel motion mode, a model predictive control algorithm is used for speed and steering control. With the goal of minimizing speed tracking error, a predictive model is constructed and the optimal control quantity is solved to achieve precise control of speed and steering.

[0212] S533: For single-leg movement mode, a fuzzy PID algorithm is used to dynamically adjust PID parameters through fuzzy rules to perform joint angle tracking control.

[0213] S534: For the wheel-leg combined motion mode, a coupled PID algorithm is used to compensate for the mutual interference of wheel-leg motion through the coupling term, so as to carry out wheel-leg coordinated motion control.

[0214] S535: Substitute the control torque of the three modes into the general control law, and combine it with the dynamic weighting coefficient to obtain the final control torque output. That is, substitute the control torque of the three modes of single wheel, single leg, and wheel-leg combination into the general control law, and combine it with the dynamic weighting coefficient to calculate the final control torque.

[0215] S54: During the stabilization phase after switching, the robot's motion state data is collected in real time and compared with the predicted values ​​of the unified dynamic equation. The control parameters are adjusted until the state error converges to the allowable range, ensuring motion stability after switching.

[0216] S55: Based on a three-stage coupled control strategy, an adaptive control model is constructed using unified dynamic equations. This adaptive control model dynamically adjusts control parameters according to changes in the robot's motion state and environment. Combined with unified dynamic equations and mode switching commands, it achieves precise control and smooth transitions between different modes. The steps for constructing the adaptive control model include:

[0217] The core algorithm modules of each stage of the three-stage coupled control strategy are identified, the input / output parameter formats of each module are clarified, and the data interaction interface is unified. At the same time, the core parameters of the unified dynamic equation and the parameters of the dynamic weighted coupled control law are identified to ensure that the parameter dimensions are matched.

[0218] The model architecture is built around the framework of "multi-source input - core fusion - control output". The input layer receives three types of data, including multi-dimensional ground characteristic parameters, mode switching commands and dynamic weight coefficients, and robot motion state data collected in real time by sensors. The core fusion layer integrates the three-stage control algorithm and unified dynamic equation, embeds dynamic weight coupling control law, and constructs the core logic of "state analysis - control quantity calculation - weight allocation". The output layer outputs the final control torque and parameter adjustment commands.

[0219] Substituting the dynamic weighted coupling control law into the unified dynamic equation and replacing the control torque in the original equation, we obtain the fused closed-loop control equation, which is expressed as:

[0220]

[0221] This achieves a deep integration between control strategies and dynamic characteristics.

[0222] Initialize control parameters and physical constraint parameters The initial values ​​of the dynamic weight coefficients are substituted into the model to complete the parameter initialization; at the same time, the inherent parameters of the core algorithm are set.

[0223] Import various typical terrain scenarios into the simulation environment, run the adaptive control model, collect the control torque and robot motion state data output by the model, and compare them with the preset optimal performance data; if there is a deviation, fine-tune the core algorithm parameters and repeat the verification until the model output meets the performance requirements.

[0224] The model that has passed simulation verification is deployed to the physical robot controller for real vehicle testing; the motion data and environmental interaction data of the physical robot are collected in real time and fed back to the model through the CAN bus to build a closed-loop optimization mechanism of "real vehicle operation - data feedback - parameter correction". Finally, a complete adaptive control model adapted to the physical robot is formed, realizing detailed control of three motion modes and transition control between modes.

[0225] S6: Based on the attention mechanism, and combined with real-time digital twin mapping data and initial control parameters, multi-source error collaborative compensation is performed on errors in three motion modes. Error compensation commands are output and fed back to the adaptive control model. Multi-source error collaborative compensation generates comprehensive compensation commands by fusing the compensation amounts of each error source to construct a collaborative optimization model. This avoids secondary interference caused by single error compensation and achieves closed-loop optimization of "error detection - feature extraction - weight allocation - compensation correction".

[0226] Step S6, the steps for multi-source error collaborative compensation include:

[0227] S61: Analyze the sources of multi-source errors and output the error analysis results. Multi-source errors are a set of various errors generated throughout the entire robot motion control process. The core sources include dynamic model errors, sensor measurement errors, mode switching transition errors, and environmental interference errors. The error source decomposition method is as follows: Based on the robot motion control link, decompose the error generation mechanism of each link and clarify the time-domain / frequency-domain characteristics of the errors. The original error calculation method is as follows: Calculate the original error through "physical-virtual" data comparison and "measured value-theoretical value" comparison to obtain the error set, represented as:

[0228]

[0229] in, For the error set, For model error, For sensor error, To account for the switching transition error, This is due to environmental interference errors.

[0230] S62: Based on the error analysis results, the real-time mapping data of the digital twin is compared with the predicted output data of the adaptive control model to extract the error features of the error sources. Joint time-domain and frequency-domain feature extraction is performed on each error component, including time-domain features and frequency-domain features, ultimately obtaining a 12-dimensional feature vector for each error source, represented as:

[0231]

[0232] Where μ is the mean, σ 2 For variance, S represents the peak value, RMSE represents the root mean square error, and K represents the kurtosis, a statistical measure describing the "peaking degree" of the error signal probability distribution curve, reflecting the proportion of extreme values ​​in the error sequence. factor is the waveform factor, the ratio of the root mean square value to the mean absolute value of the error sequence, characterizing the "smoothness" of the error waveform. I is the impulse factor, the ratio of the peak error to the mean absolute value, characterizing the intensity of "pulse-like impulses" in the error sequence. c The characteristic frequency is the frequency corresponding to the largest amplitude value in the amplitude spectrum of the error signal, representing the dominant fluctuation frequency of the error. A c The peak value of the amplitude spectrum is f, and the characteristic frequency is f. c The corresponding amplitude spectrum value characterizes the fluctuation intensity of the error at the dominant frequency. g The frequency of the centroid in the frequency domain is the "frequency centroid" of the error signal amplitude spectrum, which represents the center of the main frequency band where the error energy is distributed. is the mean square frequency in the frequency domain, and is the root mean square value of the frequency, which characterizes the degree of dispersion of the error frequency distribution. Where is the frequency domain bandwidth, and is the amplitude drop to the peak value A in the error signal amplitude spectrum. c of The frequency range at which the error is characterized is the effective frequency bandwidth.

[0233] Variance σ 2 The calculation method is as follows:

[0234]

[0235] Where, N long Let be the length of the error sequence, and ΔE(t) be the error value at time t.

[0236] peak The calculation method is as follows:

[0237]

[0238] The root mean square error (RMSE) is calculated as follows:

[0239]

[0240] It reflects the overall magnitude of error better than the mean and is the core time-domain indicator for error assessment.

[0241] The kurtosis K is calculated as follows:

[0242]

[0243] K>0 indicates a peaked distribution, where the error contains many extreme values, such as environmental interference errors; K<0 indicates a flat-peaked distribution, where the error fluctuates gently, such as model errors. The kurtosis of the normal distribution is 3, and we subtract 3 here to make K=0 for the normal distribution.

[0244] Waveform factor S factor The calculation method is as follows:

[0245]

[0246] S factor The larger the value, the more irregular the error waveform; S factor The closer it is to 1, the smoother the error waveform.

[0247] The pulse factor I is calculated as follows:

[0248]

[0249] The larger I is, the more significant the pulse-like impacts in the error, such as the torque sudden change error during mode switching and the impact error from ground protrusion.

[0250] The frequency domain centroid frequency fg is calculated as follows:

[0251]

[0252] Where A(f) is the amplitude spectrum, f s Here, f is the sampling frequency, and f is the frequency variable. This indicates that the product of "frequency × amplitude spectrum" and "amplitude spectrum" corresponding to all discrete frequency points in the range of 0 to fs / 2 are summed to calculate the weighted average frequency.

[0253] Frequency domain mean square frequency The calculation method is as follows:

[0254]

[0255] The larger the value, the more dispersed the error frequency distribution, such as environmental interference errors in complex terrain; conversely, the frequency distribution is concentrated.

[0256] The 12-dimensional feature vector is mapped to the range [0,1] using min-max normalization, as expressed by the formula:

[0257]

[0258] in, The normalized 12-dimensional error feature vector serves to eliminate the dimensional differences between different error features, ensuring that each feature has equal weight when allocating weights in the subsequent attention mechanism, and improving the training and computational stability of the compensation model. This is the minimum value in the original 12-dimensional error eigenvector. This is the maximum value in the original 12-dimensional error eigenvector.

[0259] S63: Weights are assigned to the error features of the error sources using an attention mechanism. The normalized error feature vector is then... The input feature matrix of the attention mechanism is obtained by concatenating the vectors of each error source identifier. Where T1 is the time step of the error sequence. The input feature matrix X is mapped to a query vector through a linear transformation. Key vector Value vector The formula is expressed as ,in This is a learnable linear transformation matrix. The single-head attention score is then calculated and scaled, as expressed by the formula:

[0260]

[0261] in, For single-head attention scores, To query the product of a vector and the transpose of the key vector, d k The dimension of the query / key vector is given by _____. softmax(·) is the softmax activation function that normalizes the relevance score matrix so that the sum of the elements in each row is 1. The output is the attention weight of the features at each time step.

[0262] With 8 heads, the 8 single-head attention outputs are concatenated, and then a linear transformation is applied to obtain the multi-head attention output, which is the weighted result of the error features. The formula is as follows:

[0263]

[0264] in, This is the weighted result of the error characteristics. The output is a linear transformation matrix with the same dimension as the input feature matrix. This represents the output of 8 single-head attention functions, each with a dimension of T×64. Concat(·) is the concatenation function, which concatenates the 8 single-head outputs column-wise to obtain a feature matrix with a dimension of T×512.

[0265] The weight coefficients of each error source are separated from the multi-head attention output and normalized using the Softmax function to ensure that the sum of the weights is 1.

[0266] S64: Combine the initial control parameters to establish an error prediction model, predict the error change trend based on historical error data and error characteristics of error sources, and generate multi-source error collaborative compensation quantity.

[0267] S65: Converts the multi-source error collaborative compensation quantity into error compensation command and feeds it back to the adaptive control model. Based on the input requirements of the adaptive control model, the collaborative compensation quantity... It is decomposed into joint parameter correction commands and control torque correction commands. The joint parameter correction commands are mapped to joint angle correction amounts. Joint angular velocity correction Specifically, the mapping is based on the Jacobian matrix of the unified dynamic equation, and the formula is as follows:

[0268]

[0269] in, It is the pseudo-inverse of the Jacobian matrix.

[0270] Control torque correction commands are directly mapped to joint torque compensation amounts. The formula is:

[0271]

[0272] in, This is the torque compensation coefficient, calibrated by the motor torque characteristics.

[0273] The converted standardized compensation instructions With physical constraint parameters Perform verification. If the compensated joint parameters... If the maximum angle limit is exceeded, a saturation function is used to trim the compensation amount, expressed as:

[0274]

[0275] in, This represents the joint angle correction amount after trimming, and sat(·) is the saturation function. This represents the maximum permissible angle of the joint. This represents the minimum permissible angle of the joint.

[0276] like Figure 3As shown, this invention also provides a motion adaptive control system for a wheel-legged hybrid robot, comprising: a dynamic modeling module, a parameter initialization module, a cross-modal fusion sensing module, a mode switching threshold preset module, a dynamic decision-making module, a three-stage coupled control module, and a multi-source error compensation module; the dynamic modeling module is used to construct a multi-mode dynamic model based on the structural parameters of the wheel-legged hybrid robot; the parameter initialization module is used to calibrate sensor parameters based on the structural parameters of the wheel-legged hybrid robot; the cross-modal fusion sensing module is used to collect and process synchronous data from multiple sensors based on the cross-modal fusion sensing framework, to accurately perceive terrain, obstacles, and ground adhesion characteristics, and to predict dynamic obstacle trajectories and perceive terrain abrupt changes, outputting perception results containing multi-dimensional ground characteristic parameters; the mode switching threshold preset module... The module is used to preset multi-parameter fusion switching thresholds for three motion modes—single wheel, single leg, and wheel-leg combination—based on structural parameters and perception results. The dynamic decision module is used to formulate dynamic switching decisions based on reinforcement learning according to perception results and multi-parameter fusion switching thresholds, and output the mode switching timing and dynamic weight coefficients. The three-stage coupled control module is used to construct an adaptive control model by adopting a three-stage coupled control strategy, combining mode switching timing, dynamic weight coefficients, and unified dynamic equations, to perform detailed control of the three motion modes and transition control between modes. The multi-source error compensation module is used to perform multi-source error collaborative compensation for errors in the three motion modes based on an attention mechanism, combined with real-time digital twin mapping data and initial control parameters, outputting error compensation commands and feeding them back to the adaptive control model.

[0277] In summary, the present invention provides a motion adaptive control method and system for a wheel-legged hybrid robot. By combining multi-mode dynamic modeling with cross-modal fusion perception, it accurately captures the robot's dynamic characteristics and complex environmental parameters, significantly enhancing the robot's adaptability to diverse terrains and dynamic obstacles such as flat surfaces, slopes, and steps. This enables rapid identification of terrain changes and prediction of dynamic obstacle trajectories. Through dynamic switching decision-making and a three-stage coupled control strategy, it ensures the smoothness of switching between single-wheel, single-leg, and wheel-legged hybrid modes, effectively reducing attitude fluctuations and impacts during the switching process, and balancing the efficiency of wheeled motion with the terrain adaptability of legged motion. By integrating attention mechanisms and digital twin multi-source error collaborative compensation mechanisms, it accurately locates and suppresses multi-source interference such as model errors and sensor errors, significantly improving control accuracy and motion stability. At the same time, physical constraint verification ensures hardware operational safety.

[0278] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0279] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A motion adaptive control method for a wheel-legged hybrid robot, characterized in that, include: S1: Based on the structural parameters of the wheel-leg hybrid robot, a multi-mode dynamic model is constructed and sensor parameters are calibrated to obtain a unified dynamic equation and initial control parameters; S2: Based on a cross-modal fusion perception framework, it collects and processes synchronous data from multiple sensors to accurately perceive terrain, obstacles, and ground adhesion characteristics, predict dynamic obstacle trajectories, and perceive terrain abrupt changes, outputting perception results containing multi-dimensional ground characteristic parameters. S3: Based on the structural parameters and the perception results, a multi-parameter fusion switching threshold is preset, which includes three motion modes: single wheel, single leg, and wheel-leg combination. S4: Based on the perception results and the multi-parameter fusion switching threshold, formulate a dynamic switching decision based on reinforcement learning, and output the mode switching timing and dynamic weight coefficients; S5: A three-stage coupled control strategy is adopted, and an adaptive control model is constructed by combining the mode switching timing, dynamic weight coefficient and the unified dynamic equation to perform detailed control of the three motion modes and transition control between modes. S6: Based on the attention mechanism, combined with the real-time mapping data of the digital twin and the initialization control parameters, multi-source error collaborative compensation is performed on the errors in the three motion modes, and the error compensation command is output and fed back to the adaptive control model.

2. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, In step S1, the unified dynamic equation is expressed as: ； Where M(q) is the inertia matrix, Let G(q) be the Coriolis force and centrifugal force matrix, G(q) be the gravity term vector, and q be the joint angle vector. The joint angular velocity vector. Let τ be the joint angular acceleration vector, τ be the joint torque vector output by the controller, J(q) be the Jacobian matrix, and F be the joint torque vector. e Let T be the contact force vector between the wheel leg and the ground, and T is the transpose.

3. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, Step S2, which outputs the sensing results containing multi-dimensional ground characteristic parameters, includes: S21: Based on the cross-modal fusion sensing framework, data from lidar, vision sensor, IMU and contact force sensor are synchronously acquired through time synchronization to obtain raw sensing data; S22: Preprocess the original sensing data to obtain processed sensing data; S23: A multimodal fusion perception algorithm is adopted to fuse visual image features, lidar 3D terrain features and IMU attitude features in the processed perception data, and to classify and identify terrain types through semantic segmentation and to obtain terrain geometric parameters through 3D reconstruction. S24: Calculate the normal and tangential forces on the ground based on contact force sensor data, and obtain ground hardness parameters through pressure distribution detection; S25: An LSTM neural network is used to learn the historical trajectory data of dynamic obstacles and predict their trajectory; based on the terrain geometry parameters and the ground hardness parameters, a sliding window detection algorithm is used to monitor changes in terrain parameters in real time and identify areas and degrees of terrain change. S26: Based on the obstacle movement trajectory and the terrain change area and degree of change, establish a state-environment coupled perception model, correlate and analyze robot posture fluctuations, wheel-leg contact force data and terrain texture features, motion response parameters, correct robot state error and terrain assessment error, and output the perception results containing multi-dimensional ground characteristic parameters.

4. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, Step S3, the step of presetting the multi-parameter fusion switching threshold includes: S31: Physical constraint parameters for determining motion modes based on robot structural parameters; S32: Combining the multi-dimensional ground characteristic parameters, extract the key sensing parameters that affect motion mode switching; S33: The weights of the physical constraint parameters and the key sensing parameters are determined by the analytic hierarchy process (AHP), and a multi-parameter fusion evaluation function is constructed. S34: Through a robot simulation platform, conduct multiple sets of motion tests in typical terrain scenarios to obtain the optimal motion performance data of three motion modes under different scenarios; S35: Based on the multi-parameter fusion evaluation function and the optimal motion performance data, determine the multi-parameter fusion switching threshold under different mode switching scenarios, and form a threshold matrix.

5. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, Step S4, which involves making dynamic switching decisions based on reinforcement learning, includes the following steps: S41: Construct a reinforcement learning agent, define the action space, and set the reward function; S42: The reinforcement learning agent is trained using a deep deterministic policy gradient algorithm. Combined with the multi-parameter fusion switching threshold, the action decision-making strategy is optimized through joint training in a robot simulation environment and a physical testing environment to obtain a mode-switching agent. S43: Input the current state space data in real time, and output the mode switching timing judgment result through the mode switching agent. When the state variable meets the switching threshold, output the motion mode switching command and dynamic weight coefficient.

6. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, Step S5, the steps for constructing the adaptive control model include: S51: Determine the core stages of the three-stage coupling control strategy, including the pre-switching preparation stage, the mode switching execution stage, and the post-switching stabilization stage; S52: In the pre-switching preparation stage, based on the mode switching timing judgment result and dynamic weight coefficient, the control parameters of the current motion mode are pre-adjusted, and the initial control quantity of the target motion mode is calculated according to the unified dynamic equation. S53: Based on the initial control quantity, during the mode switching execution phase, a dynamic weighted coupling control law is constructed, and the dynamic weight coefficients are integrated into the control model; S54: During the stabilization phase after the switch, the robot's motion state data is collected in real time and compared with the predicted value of the unified dynamic equation to adjust the control parameters. S55: Based on the three-stage coupled control strategy, the adaptive control model is constructed by combining the unified dynamic equation.

7. The adaptive motion control method for a wheel-legged hybrid robot according to claim 6, characterized in that, Step S52, the step of pre-adjusting the control parameters of the current motion mode, includes: S521: Calculate the control parameter adjustment amount based on the dynamic weighting coefficient, and perform gradual adjustment of the control parameter using the exponential smoothing method; S523: Based on the unified dynamic equation, substitute the dynamic parameters of the target motion mode to calculate the initial joint angle, angular velocity and torque command of the target motion mode as the initial control quantity; S524: Perform physical constraint verification on the initial control quantity to ensure that the initial control quantity does not exceed the torque, speed and angle limits of the robot joint. If it does, perform trimming and adjustment.

8. The adaptive motion control method for a wheel-legged hybrid robot according to claim 6, characterized in that, In step S53, the steps for constructing the dynamic weighted coupling control law include: S531: Set the general form of the dynamic weighted coupling control law to obtain the general control law; S532: For single-wheel motion mode, a model predictive control algorithm is used for speed and steering control; S533: For single-leg movement mode, a fuzzy PID algorithm is used for joint angle tracking control; S534: For the wheel-leg combined motion mode, a coupled PID algorithm is used for wheel-leg coordinated motion control; S535: Substitute the control torque of the three modes into the general control law, and combine it with the dynamic weighting coefficient to obtain the final control torque output.

9. The adaptive motion control method for a wheel-legged hybrid robot according to claim 1, characterized in that, Step S6, the steps for multi-source error collaborative compensation include: S61: Analyze the sources of multi-source errors and output the error analysis results; S62: Based on the error analysis results, compare the real-time mapping data of the digital twin with the predicted output data of the adaptive control model, and extract the error characteristics of the error source; S63: Weight the error features of the error sources using an attention mechanism; S64: Combine the initialization control parameters to establish an error prediction model, predict the error change trend based on historical error data and the error characteristics of the error sources, and generate multi-source error collaborative compensation amount; S65: Convert the multi-source error collaborative compensation amount into an error compensation command and feed it back to the adaptive control model.

10. A motion adaptive control system for a wheel-legged hybrid robot, comprising the motion adaptive control method for a wheel-legged hybrid robot as described in any one of claims 1 to 9, characterized in that, include: The dynamics modeling module is used to construct multi-mode dynamics models based on the structural parameters of the wheel-legged hybrid robot. The parameter initialization module is used to calibrate sensor parameters based on the structural parameters of the wheel-legged hybrid robot. The cross-modal fusion perception module is used to collect and process multi-sensor synchronous data based on the cross-modal fusion perception framework, to accurately perceive terrain, obstacles and ground adhesion characteristics, predict dynamic obstacle trajectories and perceive terrain changes, and output perception results containing multi-dimensional ground characteristic parameters. The mode switching threshold preset module is used to preset a multi-parameter fusion switching threshold including three motion modes: single wheel, single leg, and wheel-leg combination, based on the structural parameters and the perception results. The dynamic decision-making module is used to formulate dynamic switching decisions based on reinforcement learning according to the perception results and the multi-parameter fusion switching threshold, and output the mode switching timing and dynamic weight coefficients. The three-stage coupling control module is used to construct an adaptive control model by adopting a three-stage coupling control strategy, combining the mode switching timing, dynamic weight coefficients and the unified dynamic equation, and to perform detailed control of the three motion modes and transition control between modes. The multi-source error compensation module is used to perform multi-source error collaborative compensation for errors in three motion modes based on the attention mechanism, combined with real-time digital twin mapping data and the initialization control parameters, output error compensation commands and feed them back to the adaptive control model.

Images