A motion intelligent control method suitable for a four-wheel four-rotation mobile robot

Abstract

This invention discloses a motion intelligent control method suitable for four-wheeled, four-rotor mobile robots, belonging to the field of mobile robot motion control technology. The invention consists of three functional modules executed sequentially in the data flow direction: Module 1 generates a pseudo-measurement vector by constructing a nonlinear disturbance observer, and uses unscented Kalman filtering to perform online recursive estimation of the vehicle's equivalent mass, longitudinal and lateral centroid offsets, and equivalent moment of inertia, without requiring additional load sensors; Module 2 generates a smooth transition steering angle command using cubic B-spline curves, embedding the real-time centroid offset estimate into the B-spline intermediate control point constraint solution to ensure that the resultant force vector always passes through the actual physical centroid in the transition time domain, and suppresses torque residuals through feedforward compensation; This invention improves trajectory tracking accuracy by more than 60%, reduces transient acceleration jitter peak during mode switching by more than 80%, and reduces drive energy consumption by 20% to 35%, meeting the hard real-time requirements of industrial EtherCAT buses.

Description

Technical Field

[0001] This invention relates to the field of motion control technology for mobile robots, and specifically to a motion intelligent control method applicable to a four-wheeled, four-rotation mobile robot. Background Technology

[0002] The chassis of the four-wheeled, four-rotor mobile robot consists of four modular wheel sets, each containing a drive motor and a steering motor, forming a total of eight independent control channels. This enables movement in three generalized degrees of freedom: forward and backward translation, lateral translation, and yaw rotation. Because the number of control inputs far exceeds the number of degrees of freedom, this system is a typical overdrive system and has broad application prospects in industrial material handling and warehousing logistics scenarios.

[0003] However, the existing control methods for four-wheeled, four-rotor mobile robots have three long-standing technical problems that have not been fundamentally solved.

[0004] First, there is the problem of model mismatch and trajectory drift caused by the unknown position of the center of gravity under variable load conditions. Existing control systems generally assume that the center of gravity is fixed at the geometric center of the vehicle body. When the cargo placement position is not fixed, causing the actual physical center of gravity to shift, the distribution of the ground normal force and tangential friction force of each wheel is seriously inconsistent with the model preset value. Under high-speed cornering or rapid braking conditions, it is very easy to cause a large deviation in trajectory, or even the whole machine to overturn.

[0005] Second, there is the issue of transient torque jumps and mechanical vibrations during motion mode switching. Current technology, when switching between motion modes such as Ackermann steering, lateral translation, diagonal movement, and stationary zero-radius rotation, only issues sudden angle commands at the kinematic level, without considering the continuity of the resultant force vector of the eight motors during the switching process. This results in the chassis being subjected to strong internal counter-torques and ground reaction forces instantaneously, causing severe mechanical vibrations that endanger the transport safety of precision instruments or fragile goods.

[0006] Third, there is the bottleneck problem of real-time calculation for optimal torque allocation in an 8-motor overdrive system. The optimal allocation method based on nonlinear model predictive control involves multi-step matrix iterative calculations, and the time taken for a single solution is usually more than 50ms, which is far from meeting the hard real-time requirement of 1ms communication cycle of EtherCAT bus in industrial fields. This makes it impossible to deploy the theoretically optimal control algorithm on industrial control computers or embedded platforms. Summary of the Invention

[0007] The technical problem to be solved by this invention is: to address the model mismatch problem caused by the unknown center of mass position of a four-wheeled, four-rotor mobile robot under variable load conditions, the transient torque jump and mechanical vibration problem during motion mode switching, and the problem that the optimal torque distribution of the overdrive system cannot meet the industrial hard real-time requirements, and to provide a motion intelligent control method suitable for four-wheeled, four-rotor mobile robots.

[0008] The technical solution adopted in this invention is: a motion intelligent control method suitable for a four-wheeled, four-rotor mobile robot, comprising the following three functional modules executed sequentially according to the data flow direction:

[0009] Module 1 is a multi-source fusion dynamic centroid online observation module based on soft sensing. In each control cycle, it uses the triaxial acceleration and triaxial angular velocity signals collected by the inertial measurement unit, as well as the wheel speed and steering angle signals collected by the 8-channel servo motor encoder, to construct a nonlinear disturbance observer to obtain a pseudo measurement vector. Then, it uses unscented Kalman filtering to perform online recursive estimation of four state variables: the equivalent mass of the vehicle, the longitudinal offset of the centroid, the lateral offset of the centroid, and the equivalent moment of inertia, and outputs the centroid parameter estimate value for the current control cycle.

[0010] Module 2 is a motion mode transient smoothing switching module with coupled centroid offset constraints. When a motion mode switching request is detected, it uses cubic B-spline curves to generate smooth transition commands for the steering angle of each wheel in the transition time domain. It also embeds the centroid offset estimate output by Module 1 into the constraint solution process of the B-spline intermediate control point to ensure that the resultant force vector of the four wheels always passes through the actual physical centroid in the transition time domain. At the same time, it applies feedforward compensation to the resultant force vector residual.

[0011] Module 3 is a lightweight, real-time, energy-efficient optimal control allocation module for embedded control. It incorporates the centroid offset output from Module 1 into the torque arm correction calculation of the control efficiency matrix. It constructs a convex quadratic programming problem with the weighted sum of squares of the forces of 8 actuators as the objective function and the constraints of dynamic equality and tire friction inequality as the conditions. It adopts an approximate analytical solution strategy that combines the initial solution of the generalized inverse matrix with the effective set method for iteration, compressing the single-step solution time to less than 1ms, and outputting 8 driving torque and steering angle commands.

[0012] In Module 1, the state vector of the UKF It includes four components: the actual equivalent total mass of the vehicle including load. (Unit: kg) The longitudinal offset of the actual center of gravity relative to the geometric center of the vehicle body (Unit: m, forward is positive) Lateral offset of the actual center of mass relative to the geometric center of the vehicle body (Unit: m, positive to the left) Equivalent moment of inertia of the entire vehicle, including load, about the vertical axis of the vehicle body. (Unit: kg·m²).

[0013] In Module 1, the process by which the Nonlinear Disturbance Observer (NDO) constructs the pseudo-measurement vector is as follows: In each control cycle Initially, calculate the nominal dynamic vector:

[0014]

[0015] in, The nominal weight (kg) of the empty vehicle. The nominal moment of inertia of the empty vehicle (kg·m²). , The longitudinal and lateral accelerations (m / s²) in the vehicle coordinate system measured by the IMU. For the yaw rate of two consecutive frames The yaw acceleration (rad / s²) obtained by difference.

[0016] Simultaneously, the actual resultant force vector is calculated from the current of the 8 motors and the steering angle of each wheel:

[0017]

[0018] in, , For the first The longitudinal and lateral forces (N) of each wheel in the vehicle coordinate system are estimated by the drive motor current and the tire model; , For the first The longitudinal and lateral distances (m) of each wheelset relative to the geometric center of the vehicle body.

[0019] The difference between the two yields the pseudo-measurement vector:

[0020]

[0021] This pseudo-measurement vector It is a three-dimensional vector containing longitudinal dynamic residuals, lateral dynamic residuals, and yaw dynamic residuals, which serve as the observation inputs for the UKF update step.

[0022] In Module 1, the UKF prediction step is generated according to the following formula. There are Sigma points, among which State dimension:

[0023]

[0024]

[0025]

[0026] in, Estimate the posterior mean of the state at the current time. For the corresponding error covariance matrix, For scale parameters, , , For UKF Sigma point distribution parameters, The first square root of a matrix represents the square root of the matrix. Since the centroid parameter changes much slower than the control period, a random walk model is used for the state equation, and the propagation results at each Sigma point are as follows: Only the covariance matrix increases process noise. :

[0027]

[0028] in, for Process noise covariance matrix.

[0029] The UKF update step substitutes the nine predicted Sigma points into a nonlinear measurement function to calculate the predicted measurement mean and the measurement prediction covariance. State-Measurement Cross-Covariance Then calculate the Kalman gain. It also completes the posterior state update and outputs the posterior estimate. , , , .

[0030] In Module 2, the B-spline transition matrix generation process is as follows: After detecting a switching request, the switching start time is recorded. Read the current steering angle of each wheel as the starting angle vector. The target steering angle vector is calculated by inverse kinematics of the target motion mode. The transition time is automatically estimated based on the difference between the maximum angular velocity limit of each wheel and the target angle. And determine the end time of the switchover. Construct a cubic B-spline curve independently for each steering wheel, with the number of control points... The initial and final control points are fixed at the starting and target angles, respectively, and the initial and final angular velocities are zero by making adjacent control points equal. The intermediate control points are solved by simultaneously solving the following centroid constraint conditions: at the intermediate transition time... The line of action of the resultant force vector formed by the driving forces of each wheel should pass through the actual physical center of mass. During each control cycle in the transition time domain Calculate the normalization parameters Substituting into the cubic B-spline basis function, we obtain the target steering angles for each round at the current moment:

[0031]

[0032] in, Let the row vectors be cubic B-spline basis functions. This is the control point matrix. Simultaneously, the equivalent moment residual of the resultant force vector relative to the actual physical centroid is calculated. When the residual exceeds the preset threshold, a feedforward compensation amount is applied to the driving torque of each wheel.

[0033] In Module 3, the dynamic correction method for the control performance matrix is ​​as follows: the centroid offset estimated in Module 1 is... Introducing the equivalent torque arm correction calculation for each wheelset:

[0034]

[0035] in, , The known arm length (m) is the relative geometric center of each wheel set. , Let be the corrected arm length (m) relative to the actual physical centroid. An analytical construct based on the corrected arm length is then performed. Control effectiveness matrix .

[0036] The objective function of the convex quadratic programming problem is: ,in for Decision variable vector (the first 4 components are the driving forces of each wheel, and the last 4 components are the lateral forces of each wheel). for Diagonal positive definite energy consumption weight matrix; equality constraints are ,in Let be the desired generalized force vector; the inequality constraints are the limit constraints of the friction force of each tire. The solution first utilizes... Moore-Penrose pseudoinverse matrix Calculate the initial solution of least squares If the initial solution satisfies all inequality constraints, it is output directly; otherwise, the effective set method is used for iteration. The number of iterations usually does not exceed 5, and the total calculation time does not exceed 1ms.

[0037] The beneficial effects of this invention are:

[0038] The online real-time estimation error of the three-dimensional coordinates of the dynamic center of mass and the equivalent moment of inertia of the vehicle under variable load conditions does not exceed 5%, and the trajectory tracking accuracy is improved by more than 60% compared with the fixed center of mass assumption scheme.

[0039] During the switching of motion modes, the peak values ​​of transient acceleration vibration in the chassis in both the lateral and longitudinal directions are reduced by more than 80%, effectively ensuring the transportation safety of precision goods during the mode switching process;

[0040] The optimal control allocation has a single-step calculation time of less than 1ms, meeting the hard real-time requirements of industrial EtherCAT bus. It can be directly deployed on ARM Cortex-M series microprocessors or x86 architecture industrial control computers. The total square of the drive current of the 8 motors is reduced by 20% to 35% compared with uniform distribution, and the overall energy efficiency is significantly improved. Attached Figure Description

[0041] The invention will now be further described with reference to the accompanying drawings.

[0042] Figure 1 This is a block diagram of the overall system architecture of the present invention.

[0043] Figure 2 This is the overall flowchart of the entire process of this invention.

[0044] Figure 3 This is a flowchart of the dynamic centroid online observation algorithm for Module 1.

[0045] Figure 4 This is a flowchart of the transient smooth switching algorithm for motion modes in Module 2.

[0046] Figure 5 Flowchart of the real-time energy efficiency optimal control allocation algorithm for Module 3 lightweight design.

[0047] Figure 6 This is a sequence diagram showing the coordinated execution of the three modules. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] System overall architecture, such as Figure 1 As shown, the motion intelligent control method of this invention consists of modules one, two, and three executed sequentially according to the data flow direction, forming a complete sensing-estimation-decision-execution closed-loop control loop. The overall control cycle is 1ms, synchronized with the communication cycle of the industrial EtherCAT bus. The system collects raw data from IMU sensors (three-axis acceleration and three-axis angular velocity, sampling period 1ms) and 8-channel servo motor encoders (wheel speed and steering angle, sampling period 1ms). After being processed sequentially by the three modules, the 8-channel drive torque commands and steering angle commands are sent to the servo drives of each axis via the EtherCAT bus in the form of PDOs (Process Data Objects).

[0050] like Figure 2As shown, the complete real-time control main loop process of the system is as follows: system power-on initialization, loading parameters and performing bus self-test, initializing the UKF covariance matrix. The initial value of the centroid offset is set to , The initial value of the moment of inertia is set to the nominal value of the empty vehicle. After entering the 1ms real-time control main loop, data acquisition, dynamic centroid online estimation of module one, mode switching judgment, and smooth transition of module two are executed sequentially. For example, when there is a switching request or inverse kinematics solution, and when there is no switching request, the analytical convex quadratic programming control allocation of module three is performed, outputting the optimal instructions for 8 motors to complete one control closed loop.

[0051] Module 1 Specific Implementation Method: Module 1 implements online observation of dynamic centroids based on multi-source fusion using soft sensing. The algorithm flow is as follows: Figure 3 As shown.

[0052] During the system parameter configuration phase, the following known parameters are determined based on the specific model of the four-wheeled, four-rotor mobile robot: nominal weight of the empty vehicle. (kg), nominal moment of inertia of unloaded vehicle (kg·m²), longitudinal distance of each wheelset relative to the geometric center of the vehicle body and lateral distance (m, Effective radius of wheels (m). Based on actual working conditions, the UKF process noise covariance matrix is ​​set. , Reflecting the slowly varying characteristics of the centroid parameters and the measurement noise covariance matrix , This reflects the noise level of the IMU and motor current sensor, and the Sigma point distribution parameters. , , And the physical rationality test boundary of the centroid offset. , The geometric half-length and half-width of the vehicle body, in meters.

[0053] Within each 1ms control cycle, module one executes the following steps in sequence:

[0054] NDO Pseudo-Measurement Vector Construction Stage

[0055] Read the current frame IMU acceleration , (m / s²) and yaw rate (rad / s), composed of two consecutive frames Differential calculation of yaw angle acceleration (rad / s²). Read the current from 8 motors, and calculate the longitudinal force of each wheel by combining the steering angle of each wheel with the tire force estimation model (either linear tire model or Brush tire model is optional). and lateral force (N). Calculate the nominal dynamic vector according to the formula described in the "Summary of the Invention". and the actual resultant force vector The difference is used to obtain the three-dimensional pseudo-measurement vector. .

[0056] UKF Predicts Step Phase

[0057] Read the posterior mean of the output from the previous cycle. And error covariance matrix Nine Sigma points are generated according to the formula described in the "Summary of the Invention". Since the centroid parameter uses a random walk state equation, the propagation result of each Sigma point is equal to its current value, and the prediction error covariance matrix is ​​calculated according to... Update. Calculate the predicted mean. The mean weight is , ( ).

[0058] In the UKF update step, the nine predicted Sigma points are substituted into the nonlinear measurement function. This function describes the nonlinear mapping between the centroid offset parameter and the dynamic residual, and its yaw residual component includes a cross-product term of the centroid offset and the resultant force. Calculate the mean of the predicted measurements. Measurement and prediction covariance matrix and state-measurement cross-covariance matrix Covariance weights are , ( Calculate the Kalman gain. Execution status post-update:

[0059]

[0060]

[0061] In the physical plausibility check and output phase, a physical plausibility check is performed on the updated estimate: if the estimated centroid offset exceeds the vehicle body geometry range ( or ), or the estimated moment of inertia is non-positive ( ), or the estimated quality is lower than the empty vehicle quality ( If the out-of-bounds component is truncated to the boundary value, then the covariance matrix will be... Reset to This prevents the filter from diverging when the sensor malfunctions. After passing the test, the posterior estimate is output. , , , This module is called by Module 2 and Module 3. The total computation time for Module 1 is approximately 0.2ms.

[0062] Module 2's specific implementation: Module 2 realizes the transient smooth switching of motion modes coupled with centroid offset constraints. The algorithm flow is as follows: Figure 4 As shown.

[0063] In each control cycle, it is determined whether the host computer issues a motion mode switching request. The target motion modes include four standard modes: Ackerman steering, lateral translation, diagonal movement, and zero-radius rotation in place.

[0064] During the initialization of transition parameters, when a handover request is detected, the current time is recorded as the handover start time. (s) Read the steering angle encoder values ​​of each wheel as the starting angle vector. (rad); Calculate the target steering angle vector using the inverse kinematics formula of the target motion pattern. (rad); based on the maximum angular velocity limit of each wheel The transition time is automatically estimated using the following formula, based on the difference in rad / s between the target angle and the target angle:

[0065]

[0066] in, The safety margin factor (ranging from 1.2 to 2.0) is used to determine the handover end time. (s).

[0067] The stage of constructing the B-spline control point matrix with centroid constraints

[0068] For each steering wheel ( Construct a cubic B-spline curve independently, with the number of control points... Let the control point be to :

[0069] Control point 1 is fixed as the starting angle, and control point 2 is set equal to control point 1 to achieve the constraint that the initial angular velocity is zero. , ;

[0070] Control point 6 is fixed at the target angle, and control point 5 is set equal to control point 6 to achieve the constraint that the end angular velocity is zero. , ;

[0071] Intermediate control points 3 and 4 are dynamically solved based on the centroid constraint conditions: at the intermediate moment of transition. The line of action of the resultant force vector, formed by the intermediate states of the steering angles of each wheel and the constant driving force of each wheel, should pass through the actual physical center of mass. This constraint establishes an algebraic equation between the centroid offset and the position of the intermediate control point, and the optimal value of the intermediate control point is obtained by solving the equation simultaneously.

[0072] The stage of generating time-varying steering angle commands and calculating the transition matrix.

[0073] In the transition time domain Within each control cycle Calculate the normalization parameters Substituting into the cubic B-spline basis function, we obtain the target steering angle for each round:

[0074]

[0075] in, Let the row vectors be cubic B-spline basis functions. For the control point matrix (dimension 1), (Based on the current steering angle of each wheel) and dynamic centroid offset, constructed analytically based on dynamic geometric relationships. Transient time-varying transition matrix The 3D expectation generalized force is mapped to the output of 8 motors.

[0076] During the feedforward compensation and resultant force vector continuity verification phase, in each transition cycle, the equivalent moment residual of the resultant force vector relative to the actual physical centroid at the current moment is calculated:

[0077]

[0078] like If the residual exceeds the preset threshold (5 N·m), a feedforward compensation is applied to the driving torque of each wheel to force the residual to be suppressed within the threshold. When the residual continuously meets the constraint, the smooth transition is considered successful, the second loop of the module is exited, and the normal control allocation process is resumed.

[0079] Module 3 provides a detailed implementation of lightweight, real-time, energy-efficient optimal control allocation for embedded control. The algorithm flow is as follows: Figure 5 As shown.

[0080] Dynamic centroid correction control efficiency matrix calculation stage

[0081] At the start of each control cycle, read the current steering angle of each wheel. Centroid offset output by module 1 , Calculate the corrected torque arm , ( ). Parsing and building Control effectiveness matrix : Matrix number Column and number The columns correspond to the first one. The contributions of the driving force components and lateral force components of each wheel to the three generalized degrees of freedom (longitudinal force, lateral force, and yaw moment), each element is determined by... , , , The result is obtained through a combination of analytical calculations. This step involves pure matrix element assignment, and the calculation time is approximately 0.02 ms.

[0082] In the stage of finding the initial feasible point of least squares using the generalized inverse matrix, the already constructed... Calculate its Moore-Penrose pseudo-inverse matrix ( ), to obtain the initial solution of least squares ,in Let be the desired generalized force vector. If Satisfying all inequality constraints (i.e., all components are in) (within the range), then directly use The optimal solution is output, skipping the effective set method iteration. This step takes approximately 0.05 ms to compute.

[0083] in, For the first The tire friction limit (N) of each wheel is determined by the product of the tire friction coefficient and the current estimated value of the normal force; The corresponding actuator force lower limit (N) forms a two-way symmetric constraint.

[0084] In the stage of iteratively solving the optimal solution of constrained quadratic programming using the effective set method, if the initial solution... If some inequality constraints are violated, the iterative solution process using the effective set method is initiated. The effective set is initialized. Given the set of indices for all default constraints, saturate the corresponding components to the boundary values. In the... In this iteration, let the current effective set be... The component values ​​corresponding to the effective constraints are fixed as boundary values, and their contributions are subtracted from the right-hand side of the equality constraint to obtain the right-hand side of the dimensionality reduction constraint. In the dimension-reduced system composed of the remaining free components, the generalized inverse matrix is ​​applied again to solve for the optimal solution. The system checks whether the dimension-reduced solution satisfies all constraints and whether the corresponding Lagrange multipliers are non-negative. If there are valid constraints with negative Lagrange multipliers, they are removed from the valid set, and the next iteration begins. If the multipliers of all valid set constraints are non-negative and all invalid constraints are not violated, the iteration converges, and the current solution is output as the optimal solution. Since the dimension of the decision variables is fixed at 8, the number of iterations of the valid set method in actual industrial applications typically does not exceed 3 to 5 times. With the pre-compiled Eigen matrix library, the entire process can be completed within 0.3ms to 0.8ms, and the total computation time is always less than 1ms.

[0085] In the stage of converting the optimal solution into motor control commands and outputting them, the optimal actuator force vector is... Converted to motor control commands: The output torque command of each drive motor is obtained by multiplying the driving force by the effective radius of the wheel. ( (Unit: N·m); Target steering angle commands for each steering motor. During the transition phase, the B-spline interpolation steering angle output from module two is used, while during the normal tracking phase, it is determined by the inverse kinematics of the target motion mode. and Commands are sent to each axis servo drive via the EtherCAT bus in PDO form within a 1ms communication cycle, completing a full control allocation cycle.

[0086] The timing sequence of the three modules working together, such as Figure 6 As shown, the collaborative execution timing of the three modules within each 1ms control cycle (0 to 1000 microseconds) is arranged as follows: The data acquisition layer completes uplink frame parsing of IMU and encoder data within the first 50 microseconds; Module 1 sequentially completes NDO pseudo-measurement construction, UKF prediction step, and UKF update step within 50 to 450 microseconds, with a total time consumption of approximately 0.2ms; Module 2 completes B-spline interpolation evaluation within 450 to 600 microseconds when a switching request is detected (skipped if there is no switching, the time consumption is negligible); Module 3 completes control performance matrix calculation and effective set method solution within 600 to 900 microseconds; finally, EtherCAT downlink frame encapsulation and command issuance are completed within 900 to 1000 microseconds. The three modules execute sequentially, and the total computation time always meets the 1ms hard real-time constraint.

[0087] The method of this invention is applicable to various four-wheeled, four-rotor mobile robot models with different wheelbases and track widths, by parametrically configuring the wheel assembly geometric parameters ( , , ) and nominal dynamic parameters of empty vehicle ( , It can be adapted to different models without modifying the control algorithm structure.

[0088] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A motion intelligent control method applicable to a four-wheeled, four-rotor mobile robot, wherein the chassis of the four-wheeled, four-rotor mobile robot is composed of four wheel sets, each wheel set includes a drive motor and a steering motor, forming a total of 8 independent control channels, driving three generalized degrees of freedom: forward and backward translation, lateral translation, and yaw rotation; Its features are, The method comprises the following three functional modules, executed sequentially in the data flow direction within each control cycle: Module 1 utilizes the triaxial acceleration and triaxial angular velocity signals acquired by the inertial measurement unit and the wheel speed and steering angle signals acquired by the 8-channel servo motor encoder to obtain pseudo-measurement vectors by constructing a nonlinear disturbance observer. Then, it uses unscented Kalman filtering to perform online recursive estimation of four state variables: equivalent mass of the vehicle, longitudinal offset of the center of gravity, lateral offset of the center of gravity, and equivalent moment of inertia, and outputs the estimated value of the center of gravity parameters for the current control cycle. The method for constructing the pseudo-measurement vector by the nonlinear disturbance observer is as follows: the nominal dynamic vector calculated from the nominal dynamic parameters of the empty vehicle and the data from the inertial measurement unit is subtracted from the actual resultant force vector calculated from the current of the 8 motors and the steering angle of each wheel to obtain a three-dimensional pseudo-measurement vector containing longitudinal dynamic residual, lateral dynamic residual and yaw dynamic residual. Module 2: When a motion mode switching request is detected, a cubic B-spline curve is used to independently generate a smooth transition steering angle command for each steering wheel in the transition time domain. The intermediate control point of the cubic B-spline curve is solved by solving the following centroid constraints: at the intermediate moment of the transition, the line of action of the resultant force vector formed by the driving forces of each wheel passes through the actual physical centroid position estimated by Module 1; in each transition control cycle, the equivalent torque residual of the resultant force vector relative to the actual physical centroid is calculated, and when the residual exceeds a preset threshold, feedforward compensation is applied to the driving torque of each wheel. Module 3 incorporates the centroid offset output from Module 1 into the calculation of the equivalent torque arm correction for each wheel group, and analyzes and constructs a dynamically updated control performance matrix. Using the weighted sum of squares of the forces of 8 actuators as the objective function and the constraints of dynamic equality and tire friction inequality as the conditions, a convex quadratic programming problem is constructed. An approximate analytical solution strategy combining the initial solution of the generalized inverse matrix with the effective set method is adopted to compress the single-step solution time to less than 1ms and output 8 driving torque commands and steering angle commands.

2. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The prediction step of the unscented Kalman filter described in Module 1 generates the Sigma point set as follows: Let the posterior mean of the current state be... The corresponding error covariance matrix is State dimension ,generate Sigma points: in, For scale parameters, , , For the Sigma point distribution parameters, The first square root of a matrix represents the square root of the matrix. The propagation of each Sigma point uses a random walk state equation, and the prediction error covariance matrix is ​​calculated according to... Update, in which for Process noise covariance matrix.

3. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, Module 1 also includes physical rationality verification and protection steps: the estimated value of each output of the unscented Kalman filter is verified. If the estimated value of the centroid offset exceeds the geometric range of the vehicle body, or the estimated value of the equivalent moment of inertia is non-positive, or the estimated value of the equivalent mass of the whole vehicle is lower than the nominal mass of the empty vehicle, the out-of-bounds component is truncated to the corresponding physical boundary value, and the error covariance matrix is ​​reset to the initial covariance matrix to prevent the filter from diverging when the sensor is abnormal.

4. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The cubic B-spline curve described in Module 2 has 6 control points. The B-spline curves of each steering wheel satisfy the following boundary constraints: the initial control point is fixed at the steering angle of each wheel at the start of the switch, and the initial angular velocity is zeroed by setting control point 2 equal to control point 1; the final control point is fixed at the steering angle of each wheel corresponding to the target motion mode, and the final angular velocity is zeroed by setting control point 5 equal to control point 6; the transition time is automatically estimated by the difference between the maximum steering angular velocity limit of each wheel and the target angle using the following formula: in, For the target motion mode, the first The target steering angle (rad) for each steering wheel. To switch the start time The current steering angle (rad) of each steering wheel. This is the maximum permissible angular velocity (rad / s) for the steering motor. This is the safety margin factor, with a value ranging from 1.2 to 2.

0.

5. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The equivalent moment arm correction method for each wheel set described in Module 3 is as follows: The longitudinal offset of the center of mass estimated in Module 1 is... and lateral offset The following correction calculation is introduced: in, , For the first The known longitudinal and lateral distances (m) of each wheel set relative to the geometric center of the vehicle body. , The corrected arm length (m) is relative to the actual physical center of mass; based on the corrected arm length, an analytical construct is built. Control effectiveness matrix Each element is obtained by analytical calculation of the combination of the cosine and sine values ​​of the current steering angle of each wheel and the modified arm length.

6. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The iterative solution steps of the effective set method described in Module 3 are as follows: If the initial solution of the generalized inverse matrix violates some inequality constraints, the effective set is initialized as the set of indices of all violated constraints, and the corresponding components are saturated to the boundary values; in each iteration, the component values ​​corresponding to the effective constraints are fixed as boundary values, and the generalized inverse matrix is ​​applied again to solve for the optimal solution in the dimension reduction system composed of the remaining free components, and the signs of the Lagrange multipliers corresponding to all effective constraints are checked; if there are effective constraints with negative Lagrange multipliers, they are removed from the effective set and the next iteration is performed; if the multipliers of all effective set constraints are non-negative and all ineffective constraints are not violated, the iteration converges and the optimal solution is output; the number of iterations does not exceed 5, and the total computation time does not exceed 0.8ms.

7. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The three functional modules execute in the following sequence within each 1ms control cycle: data acquisition is completed within the first 50 microseconds of the control cycle; Module 1 sequentially completes the construction of the nonlinear disturbance observer, the unscented Kalman filter prediction step, and the unscented Kalman filter update step from 50 to 450 microseconds, with a total time not exceeding 0.2ms; Module 2 completes the B-spline interpolation evaluation from 450 to 600 microseconds when a switching request is detected, and skips it when there is no switching request; Module 3 completes the calculation of the control performance matrix and the solution of the effective set method from 600 to 900 microseconds; and completes the encapsulation of the EtherCAT bus downlink frame and the issuance of commands from 900 to 1000 microseconds. The three modules are executed sequentially, and the total calculation time for a single control cycle does not exceed 1ms.

8. The intelligent motion control method for a four-wheeled, four-rotor mobile robot according to claim 1, characterized in that, The target motion modes include four standard modes: Ackerman turn, lateral translation, diagonal movement, and stationary zero-radius rotation.

9. A motion intelligent control method for a four-wheeled, four-rotor mobile robot according to claim 8, characterized in that, When there is no motion mode switching request, module 2 is not activated, and the expected generalized force vector is calculated directly through the inverse kinematics solution of the current motion mode. And transmit it to Module 3; the final output torque command of each drive motor. Optimal driving force output by module three With the effective radius of the wheel The product is determined. ( The target steering angle command for each steering motor is sent to the servo driver of each axis in the form of process data object via the EtherCAT bus within a 1ms communication cycle.

Images