Motion control simulation method of ground unmanned platform under cross-country maneuvering condition

Abstract

The invention discloses a motion control simulation method of a ground unmanned platform under a cross-country maneuvering condition, which comprises the following steps of: modeling an error model of the ground unmanned platform based on a simulation unmanned platform and a real ground unmanned platform, and fusing the error model with a nominal model based on vehicle dynamics, forming a corrected motion control model of the simulation ground unmanned platform; designing a finite time domain optimal controller based on the corrected semi-data driving model, and taking a quadratic objective function as a cost function and fusing an error model as a constraint condition; designing a double constraint rule to inhibit the abnormality of the predicted value of the error model; an error model training data set is constructed through sparse processing, feature space coverage optimization and a dynamic updating mechanism, and the generalization ability and the calculation efficiency of the model are improved in combination with a local subspace division strategy. According to the invention, the high-precision motion control model for ground unmanned platform simulation is constructed, so that high-fidelity motion control simulation of the ground unmanned platform is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of ground unmanned platform simulation, and particularly to a motion control simulation method for a ground unmanned platform under off-road maneuvering conditions. Background Art

[0002] The simulation test of ground unmanned platforms, with its significant advantages of low cost and high efficiency, has become a key way to verify the algorithm performance of platforms. Especially in the verification link of autonomous path planning algorithms, motion control simulation has demonstrated an irreplaceable core value.

[0003] However, the deficiencies of the existing technologies are that most of the currently widely adopted robot or autonomous driving vehicle simulation software, such as Gazebo, Carla, etc., are mostly limited to simulating standard roads or relatively simple terrains, and fail to provide a comprehensive and in-depth simulation of the profound impacts of various changing factors in complex off-road environments on vehicle motion control. The calculation models of these software often overly rely on simplified dynamic models and limited physical models, which directly leads to significant deviations between their dynamic simulation engines, such as ODE, PhysX, etc., and the behavioral characteristics of actual vehicles when simulating the ground mechanics performance of vehicles under off-road conditions. This deviation not only weakens the authenticity and credibility of the simulation test, but also makes it difficult for the ground unmanned platform simulation model to accurately reproduce the path tracking motion state of the real-world off-road maneuvering conditions under complex off-road working conditions, thus greatly limiting the practical application effect of simulation technology in improving the verification of platform algorithm performance. Summary of the Invention

[0004] The purpose of the present invention is to overcome the deficiencies of the existing technologies. To achieve the above purposes, a motion control simulation method for a ground unmanned platform under off-road maneuvering conditions is adopted to solve the problems raised in the above background art.

[0005] A motion control simulation method for a ground unmanned platform under off-road maneuvering conditions includes the following steps:

[0006] Step S1: Based on the differences in the motion data of the simulated unmanned platform and the real ground unmanned platform under off-road working conditions, an error model of the ground unmanned platform is modeled, and then the error model is fused with a nominal model based on vehicle dynamics to form a modified motion control model of the simulated ground unmanned platform;

[0007] Step S2: Design a finite-time optimal controller based on the modified semi-data-driven model, using a quadratic objective function as the cost function and fusing the error model as the constraint condition;

[0008] Step S3: Then design a double constraint rule to suppress abnormal predicted values of the error model;

[0009] Step S4: Construct an error model training data set through sparsification processing, feature space coverage optimization, and a dynamic update mechanism, and combine a local subspace partitioning strategy to improve the model generalization ability and computational efficiency.

[0010] As a further solution of the present invention: The specific steps in the step S1 include:

[0011] Step S11: Based on a classical vehicle dynamics or kinematics model, construct a nominal model of the simulated unmanned platform. The nominal model of the simulated unmanned platform is expressed as:

[0012] x(t + 1) = f(x(t), u(t));

[0013] where f is the nominal model of the ground unmanned platform, and x(t) and u(t) are the simulation system state and simulation control input of the ground unmanned platform at time t;

[0014] Step S12: Model the error of the nominal model of the simulated unmanned platform to construct a higher-precision semi-data-driven correction model;

[0015] Step S13: Integrate the error model with the nominal model to obtain a corrected motion control model of the simulated ground unmanned platform.

[0016] As a further solution of the present invention: The specific steps in the step S2 include:

[0017] Step S21: For the corrected motion control model of the simulated ground unmanned platform, design a finite-horizon optimal controller, and the cost function is designed as a quadratic objective function;

[0018] Step S22: Take the corrected model containing the Gaussian process regression modeling error term as a constraint, and its discrete expression form within a finite horizon is:

[0019]

[0020] s.t. x(t + 1) = f(x(t), u(t)) + d(z(t) + ω(t)), t0 ≤ t ≤ H - 1

[0021] In the formula, J is the cost function of the optimal controller; H is the time horizon length predicted by the controller; Q(t) and S(t) are weight matrices of semi-positive definite matrices, and R(t) is a weight matrix of a positive definite matrix;

[0022] Step S23: Calculate the optimal control quantity of the controller based on the open-source optimizer qpOASES.

[0023] As a further solution of the present invention: The specific steps in the step S3 include:

[0024] Step S31, Abnormal data screening based on variance threshold: Set the confidence interval threshold in each output dimension of the error model, and achieve dynamic filtering by monitoring the estimated variance in real time;

[0025] When the variance of the predicted value exceeds the preset threshold, it is determined that the confidence level of the error estimation in this dimension is insufficient;

[0026] Step S32, Residual contribution limit control: In the stage of synthesizing the control quantity, set the upper limit of the weight of the error compensation quantity to the total control output. By dynamically monitoring the error ratio, when the correction amount contribution of the error model exceeds the set ratio of the output of the basic model, trigger the safety protection mechanism and suspend the output of the compensation quantity in this dimension.

[0027] As a further solution of the present invention: The specific steps in the said step S4 include:

[0028] Step S41, Training dataset generation, The ground unmanned platform autonomously controls its driving under the real - scene off - road working condition, and real - time collects the control input of the ground unmanned platform, characteristic state variables such as the speed and yaw angular velocity of the unmanned platform, the lateral error of the unmanned platform path, and the state variables of the heading error and the change rate of the error;

[0029] Step S42, Training dataset segmentation, Divide the entire dataset into each local space, and only the local dataset unit most relevant to the current input is used to predict and simulate the error model of the ground unmanned platform;

[0030] Step S43, Local training of the error model, Segment according to the said training dataset;

[0031] Step S44, Training dataset update, During the real - time operation of the simulation, monitor the newly generated data pairs in the dataset and judge whether the new data point meets the requirements to be inserted into the dataset;

[0032] Step S45, Training dataset regeneration, According to the above - mentioned training dataset update, since each local space will be continuously updated over time, the pre - calculated hyperparameters will gradually deviate from the optimal values of the new local space.

[0033] Compared with the prior art, the present invention has the following technical effects:

[0034] By adopting the above technical solution, by using the method of supervised learning in the motion simulation engine, and taking advantage of the differences between the virtual simulation and the real-world historical data of the ground unmanned platform in off-road conditions, the error model in the unmanned platform simulation motion is identified and corrected, so as to obtain a high-precision motion control model. This innovative method enables the simulation platform to more realistically reproduce the motion state of the actual vehicle in off-road maneuvering conditions, effectively solves the problem of insufficient fidelity in the motion control simulation of the ground unmanned platform, and significantly improves the effectiveness and credibility of the simulation results in off-road conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] The following will describe in detail the specific embodiments of the present invention with reference to the accompanying drawings:

[0036] Figure 1 Schematic diagram of the steps of the motion control simulation method for the disclosed embodiment of the present application;

[0037] Figure 2 Schematic diagram of the implementation process of the data set generation method for the error model of the disclosed embodiment of the present application. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0039] Please refer to Figure 1 and Figure 2 In the embodiments of the present invention, a motion control simulation method for a ground unmanned platform under off-road maneuvering conditions includes the following steps:

[0040] Step S1: Based on the differences in the motion data of the simulated unmanned platform and the real ground unmanned platform in off-road conditions, an error model of the ground unmanned platform is modeled, and then the error model is fused with the nominal model based on vehicle dynamics to form a corrected motion control model of the simulated ground unmanned platform. The specific steps include:

[0041] In this embodiment, a modeling method for the motion control model of the simulation platform based on supervised learning. This model is divided into two parts: a nominal model and an error model.

[0042] Step S11: Based on a classical vehicle dynamics or kinematics model, a nominal model of the simulated unmanned platform is constructed. The nominal model of the simulated unmanned platform is expressed as:

[0043] x(t + 1) = f(x(t), u(t));

[0044] Among them, \(f\) is the nominal model of the ground unmanned platform, \(x(t)\) and \(u(t)\) are the simulation system state and simulation control input of the ground unmanned platform at time \(t\);

[0045] Step S12: Model the error of the nominal model of the simulated unmanned platform to construct a semi-data-driven correction model with higher precision;

[0046] Step S13: Integrate the error model and the nominal model to obtain the motion control model of the corrected simulated ground unmanned platform.

[0047] Due to the lack of high-precision simulation of the influence of variable factors in complex off-road environments on vehicle motion control in the current vehicle motion simulation engine, there are significant differences between its dynamic simulation engine and the ground mechanics performance of actual vehicles under off-road conditions, resulting in insufficient accuracy of the basic nominal model.

[0048] In this embodiment, a supervised learning-based error model modeling method is introduced to model the error of the nominal model of the simulated unmanned platform and construct a semi-data-driven correction model with higher precision, as follows:

[0049] In this embodiment, the supervised learning algorithm of Gaussian process regression is used to model the error model from the differences in the motion state data between the simulated vehicle and the real vehicle, including the errors not modeled in the nominal model derived from vehicle dynamics / kinematics. The motion simulation model of the corrected ground unmanned platform is expressed in a discretized form as:

[0050]

[0051] In the formula, \(d\) is the error model of the constructed ground unmanned platform; \(z\) is the regression feature selected by the Gaussian process regression algorithm. In this embodiment, the regression feature \(z\) includes the longitudinal speed, lateral speed, and yaw angular velocity of the ground unmanned platform; \(\omega\) is the noise.

[0052] The training data set for constructing the error model \(d\) from the difference data of the motion states of the simulated ground unmanned platform and the real-world ground unmanned platform is expressed as:

[0053]

[0054] In the formula, is the training data set of the error model, \(y\) is the output set, and \(y\) is obtained by taking the difference between the motion state parameters of the ground unmanned platform and the output value of the nominal model \(f\):

[0055] y j =x j+1 -f(x j ,u j )=d(z j) + ω j ;

[0056] where x j and x j+1 are the simulation system states at time j and j + 1; y j is the output value at time j; u j is the simulation control input at time j; z j is the regression feature at time j; ω j is the noise at time j.

[0057] In this embodiment, the output dimension of the error model is optimized, and the output items with relatively small contributions to the error value of the final output prediction are deleted to reduce the real-time computational load.

[0058] In this embodiment, the optimized prediction vector and its corresponding data set are denoted as y s and Y s . According to the properties of the Gaussian process regression model, the data Y s follows a normal distribution, and the expression is:

[0059]

[0060] where is the prior mean function of the regression feature with an output dimension of n z . In this embodiment, it is assumed that each mean is zero; is the Gram matrix of the error model training data set,

[0061]

[0062] where is the kernel function selected by the Gaussian process regression algorithm in this embodiment, z i and z j are the regression features at time i and j. In this embodiment, the classical squared exponential kernel function can be optionally used.

[0063] In this embodiment, in the error model modeling, it is assumed that the data can be represented as samples of a multivariate Gaussian distribution. Therefore, under the training data, the Gaussian distribution of any test point can be expressed as:

[0064]

[0065] where the first term is the mean term distribution, which is the optimal estimated value of the modeling error of the ground unmanned platform model; the second term is the variance term, which is the uncertainty of the optimal estimated value of the prediction.

[0066] In this embodiment, the prediction reliability of the error model depends on the kernel function in the Gaussian process regression model Selection of hyperparameters

[0067] In this embodiment, a hyperparameter optimization method based on probabilistic inference is adopted to construct an iterative optimization mechanism driven by gradients to achieve the convergence of the maximum value of the log marginal likelihood function, which dynamically tracks the maximum a posteriori estimation trajectory and then locks in the optimal hyperparameters of the kernel function

[0068] Aiming at the problem that single-path optimization is vulnerable to local extreme value interference, in this embodiment, through multiple optimization calculations, the risk of falling into a local suboptimal solution during the gradient descent process is reduced, and the robustness of the final hyperparameter calculation is improved

[0069] Step S2: Design a finite-horizon optimal controller based on the modified semi-data-driven model, using a quadratic objective function as the cost function and integrating the error model as a constraint condition. The specific steps include:

[0070] Step S21: Modify the motion control model of the simulated ground unmanned platform and design a finite-horizon optimal controller, with the cost function designed as a quadratic objective function

[0071] Specifically, according to the modified ground unmanned platform simulation model, design a path tracking controller for this type of semi-data-driven control model. In this embodiment, the path tracking controller of the simulation platform is designed based on the optimal controller, and the cost function is designed as a quadratic objective function

[0072] Step S22: Take the modified model containing the Gaussian process regression modeling error term as a constraint, and its discrete expression form within a finite time domain is:

[0073]

[0074] s.t. x(t + 1) = f(x(t), u(t)) + d(z(t) + ω(t)), t0 ≤ t ≤ H - 1

[0075] In the formula, J is the cost function of the optimal controller; H is the time domain length predicted by the controller; Q(t) and S(t) are weight matrices of semi-positive definite matrices, and R(t) is the weight matrix of a positive definite matrix

[0076] Step S23: Calculate the optimal control quantity of the controller based on the open-source optimizer qpOASES

[0077] Step S3: Design double constraint rules to suppress abnormal predicted values of the error model. The specific steps include:

[0078] ​Specifically, the constraint of the optimal control quantity. Due to the randomness of the Gaussian process model, when the current state is not effectively covered by the training data set, the semi-data-driven optimal controller may solve a "huge number" of control quantities, seriously affecting the accuracy of motion simulation. Therefore, the predicted value of the error model is constrained.

[0079] In this embodiment, a dual constraint mechanism is established for the compensation quantity output of the ground unmanned platform error model. The specific technology includes the following control strategies:

[0080] Step S31: Screening of abnormal data based on variance threshold: Set the confidence interval threshold in each output dimension of the error model, and realize dynamic filtering by real-time monitoring of the estimated variance;

[0081] When the variance of the predicted value exceeds the preset threshold, it is determined that the confidence level of the error estimation in this dimension is insufficient;

[0082] Specifically, the system automatically blocks the transmission channel of this parameter to the controller. This mechanism can effectively suppress the abnormal disturbance caused by model mismatch.

[0083] Step S32: Residual contribution limit control: In the control quantity synthesis stage, set the upper limit of the weight of the error compensation quantity to the total control output. By dynamically monitoring the error ratio, when the contribution of the correction quantity of the error model exceeds the set ratio of the output quantity of the basic model, trigger the safety protection mechanism and suspend the output of the compensation quantity in this dimension.

[0084] Specifically, this design not only maintains the dynamic characteristics of the basic model but also avoids the risk of system instability caused by over-correction of the error model.

[0085] Step S4: Construct the error model training data set through sparsification processing, feature space coverage optimization, and dynamic update mechanism, and combine the local subspace division strategy to improve the generalization ability and calculation efficiency of the model. Its specific steps include:

[0086] In this embodiment, "high-quality" data points can be collected from the online data stream and the data set can be continuously updated to better cover the feature space within a limited data scale.

[0087] As Figure 2 shown, the generation and optimization of the error model training data set mainly include: data set generation, training data set segmentation, local training of the error model, training data set update, and training data set regeneration. The method disclosed in the present invention is compatible with any simulation platform developed by a physics engine, such as UE, Gazebo, etc. This implementation example will take UE5 as the simulation platform as an example to show the construction method of the error model data set.

[0088] Step S41: Generate a training dataset. The ground unmanned platform autonomously controls its movement under real - world off - road conditions, and real - time collects characteristic state variables such as the control input of the ground unmanned platform, the speed and yaw rate of the unmanned platform, the lateral error of the unmanned platform path, as well as the state variables of the heading error and the change rate of the error.

[0089] Specifically, there are multiple types of off - road conditions. At the same time, the training dataset will also construct corresponding datasets for different road surface types. When the dataset is called, different datasets will also be used for training the error model under this condition according to the simulated task type.

[0090] During the collection of the dataset, gradient - style planning is used to target the speed and yaw rate of the ground unmanned platform and other target states, so that the movement trajectory of the unmanned platform covers the characteristic states as widely as possible.

[0091] In the simulation environment of UE5, configure the ground unmanned platform model. The key settings such as the driving form of the ground unmanned platform, the overall vehicle parameters, the tire - type parameters, and the power output curve are kept consistent with the parameters of the real unmanned platform. The parameters of these physical simulation models of the ground unmanned platform are stored in a josn file and are called when the system starts.

[0092] Use the WheeledVehiclePawn class in UE5 to implement the basic motion control function of the ground unmanned platform simulation model.

[0093] Based on the above - mentioned simulation environment and the simulation model of the unmanned platform, set the control model of the ground unmanned simulation vehicle to the same basic model as that used for real - world ground unmanned platform path tracking.

[0094] Use the same reference trajectory input value collected from real - world data in the ground unmanned platform simulation, and collect characteristic state variables such as the control input of the simulated unmanned platform, the speed and yaw rate of the simulated unmanned platform, the lateral error of the simulated unmanned platform path, the heading error, and the change rate of the error.

[0095] Compare the motion state data of the real vehicle and the simulated vehicle, and use the above - mentioned motion state parameters and the formula of the output value of the nominal model f to calculate the tracking error between the motion trajectories of the simulated ground unmanned platform and the real ground unmanned platform, and then synthesize the original dataset for model error training.

[0096] Implement sparsification processing on the original dataset to screen out a data sample combination with higher information content under limited storage capacity. This embodiment adopts a two - stage optimization strategy:

[0097] First, perform sparsification on the data points generated by oversampling. Specifically, based on proximity analysis, redundant data points with too close spatial distances are removed through iterative calculation. The formula can be expressed as:

[0098] z i -z j ≤η d

[0099] In the formula, η d is a preset distance metric threshold, and this parameter retains the key samples with significant discrimination for model training; z i and z j are the regression features at the i-th and j-th moments.

[0100] Then, implement feature space coverage optimization to construct a sample set with multi-dimensional representation capabilities. In this embodiment, data points that effectively cover a wider feature space are used. Specifically, by setting the independence scale parameter,

[0101]

[0102] In the formula, β i is the independence scale parameter. In this embodiment, with the independence scale parameter as the reference index, data points with a lower independence scale are discarded, and finally a sparsified error model training data set is generated.

[0103] Step S42: Split the training data set. Divide the entire data set into individual local spaces, and only the local data set unit that is most relevant to the current input is used to predict the error model of the simulated ground unmanned platform;

[0104] Specifically, the division of the local subspace of the training data set uses the distance scale γ based on the Gaussian kernel i :

[0105]

[0106] In the formula, represents the center of the local subspace of the j-th training data set, and M a is the Gaussian kernel width.

[0107] In this embodiment, if the distance measurement value is a relatively large value, indicating that z i is relatively relevant to , then several points closest to the center point are inserted into the local subspace of j.

[0108] The segmentation of the dataset reduces the amount of data that the supervised learning algorithm of the error model needs to process, reduces the dimension of the covariance matrix in the Gaussian process regression model, and lowers the computational complexity. On the other hand, it alleviates the problem of unbalanced data point distribution caused by the aggressive oversampling of the original data.

[0109] Step S43: Locally train the error model and segment it according to the training dataset;

[0110] In this embodiment, during the learning process of the error model, new regression features are sampled in real time, and the subspace data closest to the input features is found from each local subspace.

[0111] The selection of the local subspace of the training dataset inherits from the data sparsification process of the above-mentioned training dataset, and uses the independence scale parameter β i , which represents the approximation degree between the query point and the feature space of the tested local subset.

[0112] In this embodiment, a larger β i indicates that the query point is not suitable for using the current local subspace for Gaussian process regression. Therefore, the local subspace with the smallest independence from the query point is selected for the prediction of the error model.

[0113] Step S44: Update the training dataset. During the real-time operation of the simulation, the dataset monitors newly generated data pairs and determines whether the new data point meets the requirements to be inserted into the dataset;

[0114] This step is to more efficiently utilize the limited data points in the continuous data stream and cover the feature space as much as possible.

[0115] During the real-time operation of the simulation, the dataset monitors newly generated data pairs and determines whether the new data point needs to be inserted into the dataset. The judgment principle is as follows:

[0116] In this embodiment, the dataset update uses two thresholds of the data independence scale [η lower , η upper to determine whether to update and control the update frequency. Specifically, the design rules are as follows:

[0117] If the independence scale of the new input point is small, that is, less than the threshold η lower , it means that it can be effectively covered by the local area, and there is no need to update the dataset;

[0118] If the independence scale value is between the two thresholds, the input data point will replace the original point with the smallest scale;

[0119] If the independence scale value is greater than the threshold η upper, the new data point is added to the corresponding local subspace to obtain a broader feature space.

[0120] After inserting the new data point, the Gram matrix of the Gaussian process regression model is synchronized. In this embodiment, the number of allowed points slightly exceeds the maximum limit of the local area. If the number exceeding the limit reaches a certain critical value, the newly inserted points will be stored in a temporary space, and the local subspace of this data set will not participate in the real-time calculation of the error model temporarily.

[0121] Step S45, the training data set is regenerated and updated according to the above training data set. Since each local space will be continuously updated over time, the pre-computed hyperparameters will gradually deviate from the optimal values of the new local space.

[0122] In this embodiment, the data points in the temporary space are added to the data set, the data set is regenerated and the new data set is segmented.

[0123] The processes of data set regeneration and re-segmentation are both automatically executed during the initialization process of the experiment to prevent affecting the real-time performance during the simulation test;

[0124] As the joint debugging experiment between the simulation and the real vehicle progresses, since the data set gradually covers the feature space more comprehensively, the update frequency of the data set and the growth rate of the number of local subspaces will also gradually tend to be stable.

[0125] After completing all the above processes, the training data set of the error model of the ground unmanned platform under off-road conditions is optimized, and this data set will be input into step S1 to realize the prediction of the control model error of the unmanned platform in the simulation system.

[0126] Although the embodiments of the present invention have been shown and described, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents and should be included within the protection scope of the present invention.

Claims

1. A motion control simulation method for a ground unmanned platform under cross-country maneuvering conditions, characterized in that It includes the following steps: Step S1: Based on the differences in the motion data of the simulated unmanned platform and the real ground unmanned platform under off-road conditions, an error model of the ground unmanned platform is built, and then the error model is fused with the nominal model based on vehicle dynamics to form a modified motion control model of the simulated ground unmanned platform; Step S2: Design a finite-time optimal controller based on the modified semi-data-driven model, with a quadratic objective function as the cost function and fusing the error model as the constraint condition; Step S3: Then design a double constraint rule to suppress abnormal predicted values of the error model; Step S4: Construct an error model training data set through sparsification processing, feature space coverage optimization, and dynamic update mechanism, and combine the local subspace division strategy to improve the model generalization ability and computational efficiency.

2. The motion control simulation method of a ground unmanned platform under off-road maneuvering conditions according to claim 1, wherein The specific steps in the above Step S1 include: Step S11: Based on the classical vehicle dynamics or kinematics model, construct the nominal model of the simulated unmanned platform. The nominal model of the simulated unmanned platform is expressed as: x(t + 1) = f(x(t), u(t)); where f is the nominal model of the ground unmanned platform, and x(t) and u(t) are the simulation system state and simulation control input of the ground unmanned platform at time t; Step S12: Model the error of the nominal model of the simulated unmanned platform to construct a higher-precision semi-data-driven modified model; Step S13: Fuse the error model with the nominal model to obtain the modified motion control model of the simulated ground unmanned platform.

3. The motion control simulation method of an unmanned ground platform under off-road maneuvering conditions according to claim 1, characterized in that The specific steps in the above Step S2 include: Step S21: For the modified motion control model of the simulated ground unmanned platform, design a finite-time optimal controller, and the cost function is designed as a quadratic objective function; Step S22: Take the modified model containing the Gaussian process regression modeling error term as the constraint, and its discrete expression form within the finite time domain is: s.t. x(t + 1) = f(x(t), u(t)) + d(z(t) + ω(t)), t0 ≤ t ≤ H - 1 In the formula, J is the cost function of the optimal controller; H is the time domain length predicted by the controller; Q(t) and S(t) are weight matrices of positive semi-definite matrices, and R(t) is the weight matrix of positive definite matrices; Step S23: Calculate the optimal control quantity of the controller based on the open-source optimizer qpOASES.

4. The motion control simulation method of a ground unmanned platform under off-road maneuvering conditions according to claim 1, characterized in that, The specific steps in the above Step S3 include: Step S31: Screening of abnormal data based on variance threshold: Set the confidence interval threshold in each output dimension of the error model, and realize dynamic filtering by real-time monitoring of the estimated variance; When the variance of the predicted value exceeds the preset threshold, it is determined that the confidence level of the error estimation in this dimension is insufficient; Step S32: Residual contribution limit control: In the control quantity synthesis stage, set the upper limit of the weight of the error compensation quantity to the total control output. By dynamically monitoring the error ratio, when the contribution of the correction quantity of the error model exceeds the set ratio of the output quantity of the basic model, trigger the safety protection mechanism and suspend the output of the compensation quantity in this dimension.

5. The motion control simulation method of a ground unmanned platform under off-road maneuvering conditions according to claim 1, characterized in that, The specific steps in the above Step S4 include: Step S41, generation of training dataset: The unmanned ground platform autonomously controls its driving under real - scene off - road conditions, and in real - time collects the control input of the unmanned ground platform, characteristic state variables such as the speed and yaw angular velocity of the unmanned platform, the lateral error of the path of the unmanned platform, as well as the state variables of the heading error and the change rate of the error. Step S42, segmentation of training dataset: The entire dataset is divided into individual local spaces, and only the local dataset unit most relevant to the current input is used to predict and simulate the error model of the unmanned ground platform. Step S43, local training of the error model: Segment according to the training dataset. Step S44, update of training dataset: During the real - time operation of the simulation, the dataset monitors newly generated data pairs and determines whether the new data points meet the requirements for insertion into the dataset. Step S45, regeneration of training dataset: According to the above - mentioned update of the training dataset, since each local space will be continuously updated over time, the pre - calculated hyperparameters will gradually deviate from the optimal values of the new local space.

Images