1Device and method for controlling the lateral dynamics of a vehicle

Abstract

The present invention relates to a device for controlling the lateral dynamics of a vehicle. The device comprises a trained machine learning network configured to map input data comprising a target curvature value and one or more speed values ​​to an output prediction of a corresponding wheel angle for the vehicle.

Description

[0001] The present invention relates generally to controlling the lateral dynamics of a vehicle and, more particularly, to controlling the lateral dynamics of a vehicle based on machine learning models.

[0002] Self-driving technology has been the trend for decades. The shift toward autonomous driving is being driven by recent developments in artificial intelligence, big data, and information processing techniques. Automated driving systems can comprise many modules that can work in a coordinated manner. A perception module can include multiple sensors such as cameras, GPS (Global Positioning System), LiDAR (Light Detection and Ranging), Radar (Radio Detection and Ranging), and IMU (Inertial Measurement Unit). A variety of algorithms can be used by the perception module to sense an environment and obtain relevant information. A planning and decision module can use sensor data to make decisions and plan desired speed and trajectory profiles. These decisions can be executed by a control module. Control is an important function for autonomous driving.Generally, the control module is divided into longitudinal and lateral control; longitudinal control can handle speed tracking, and lateral control can ensure precise steering. This can be achieved by controlling various actuators such as the accelerator pedal, brakes, and steering wheel.

[0003] Control targets can be obtained from a planning module and further processed by a target trajectory generator, which provides target vectors for lateral control and longitudinal control. The present invention relates to the lateral controller and, in particular, to the curvature controller. To control lateral vehicle dynamics, a cascaded control structure with an outer-loop position controller and an inner-loop curvature controller can be used. A position controller can receive a lateral control target from a trajectory generator, and its output can be a new desired target curvature for the inner loop. A purely kinematics-based position controller can use a feedforward term and a feedback term. The feedforward term can be determined by input linearization, and feedback gains can be determined by pole placement with velocity-dependent gain scheduling.

[0004] One task of the curvature controller may be to find a suitable steering reference command for the vehicle so that it closely follows the target curvature. In traditional controller design, a model-based approach can be chosen to account for vehicle dynamics. However, the model (such as the single-track model) will never be perfect, and problems with a model-based controller can arise if the assumptions made during modeling of the vehicle dynamics are invalid or if significant uncertainties exist in the vehicle parameters. Failure to adapt to changing vehicle dynamics or other environmental conditions could also lead to unsatisfactory tracking behavior of the curvature controller.

[0005] Therefore, there may be a need to improve the tracking error of conventional model-based curvature controllers.

[0006] This object is achieved by devices and methods for controlling the lateral dynamics of a vehicle according to the subclaims.

[0007] According to a first aspect, the present invention proposes an apparatus for controlling the lateral dynamics of a vehicle, such as land vehicles (e.g., cars), watercraft (e.g., ships), aircraft (e.g., airplanes), spacecraft (e.g., spaceships), or rail vehicles (e.g., trains). The proposed apparatus comprises a trained machine learning network / model configured to map input data comprising a desired curvature value and one or more speed values ​​to an output prediction of a wheel angle for the vehicle. The trained machine learning network / model can act as a feedforward controller for lateral vehicle dynamics, which relates to the movement of a vehicle in the horizontal plane and focuses primarily on steering, turning, and maintaining a lane.The feedforward controller can anticipate system behavior based on the trained machine learning model and the inputs and provide appropriate control actions. In the context of lateral vehicle dynamics, it can predict the necessary steering input to follow a target path or trajectory. Thus, a machine learning-based feedforward controller for vehicle lateral dynamics is proposed. The machine learning network takes not only the target curvature as input, but also one or more (actual or measured) speed values.

[0008] The machine learning network can be an alternative to traditional analytical methods of feedforward control. Such analytical methods must only use the road curvature to calculate the target wheel angle. For example, the kinematic or dynamic single-track model tends to saturate its wheel angle at higher speeds. By considering the vehicle speed when calculating forward power, the machine learning model can also be developed based on the vehicle's longitudinal speed, the current curvature, and past and future curvatures. The input to the neural network is selected based on the correlation of the actual wheel angle.

[0009] "Curvature" refers to the measure of the rate of change in direction of a curve (e.g., road) or surface relative to a particular point. In simpler terms, it describes how sharply a curve bends or turns. Mathematically, curvature can be defined in several ways, depending on the context and dimensionality of the space in which the curve or surface exists. For a curve in two dimensions, the curvature k at a given point is defined as the rate of change of the tangent angle with respect to the arc length. If θ(s) is the angle tangent to the curve at a point a distance s along the curve from a fixed point, then the curvature k is given by: k = d θ / ds. For a circle of radius r, the curvature is constant and equal to 2 π / 2 π r = 1 / r.

[0010] In some embodiments, the input data of the machine learning network includes a current curvature value indicating the current curvature of a target trajectory, a current speed value indicating the current longitudinal speed of the vehicle, at least one past speed value indicating the past longitudinal speed of the vehicle, and at least one future curvature value indicating the future curvature of the target trajectory. The current longitudinal speed and the at least one past speed value indirectly provide information about the longitudinal acceleration of the vehicle. It is also conceivable to use the (longitudinal) acceleration of the vehicle directly as input for the machine learning network. By taking these additional inputs into account, the performance of the model can be further improved compared to conventional analysis methods.

[0011] In some embodiments, the input data of the machine learning network includes a current speed value indicating the current longitudinal speed of the vehicle, a first past speed value indicating the longitudinal speed of the vehicle before a first time, and a second past speed value indicating the longitudinal speed of the vehicle before a second time. By considering these additional inputs, the performance of the model can be further improved compared to conventional analysis methods.

[0012] In some embodiments, the input data of the machine learning network includes a current curvature value indicating the current curvature of a target trajectory, a past curvature value indicating the past curvature of the trajectory, a first future curvature value indicating the future curvature of the trajectory at a first look-ahead time, and a second future curvature value indicating the future curvature of the trajectory at a second look-ahead time. By considering these additional inputs, the performance of the model can be further improved compared to conventional analysis methods.

[0013] In some embodiments, the machine learning network comprises a fully connected neural network with an input layer, at least two hidden layers, and an output layer. In a fully connected neural network, each neuron in one layer is connected to every neuron in the next layer. Fully connected networks have a simple and straightforward architecture. This universality makes them applicable to problems where the relationship between input and output is not spatial or temporal. Due to their dense connections, these networks are capable of capturing complex patterns and relationships in the data. They can learn complicated functions that map inputs to outputs, which can be important for tasks involving classification, regression, or feature extraction. Fully connected networks are relatively easy to implement and train.Fully connected networks can handle inputs of different sizes as long as the input is flattened (converted into a one-dimensional vector) before being fed into the network.

[0014] In some embodiments, the machine learning network comprises a Long Short-Term Memory (LSTM) network with an input layer, at least two LSTM layers, and an output layer. Long Short-Term Memory Networks (LSTMs) can offer advantages for tasks with sequential data. For example, LSTMs can be specifically designed to address the challenge of long-term dependencies in sequential data. They can remember information for longer periods of time. LSTMs can mitigate gradients becoming too small for effective learning through their gating mechanisms, enabling more effective and stable training over many time steps. The gates in LSTMs (forget gate, input gate, and output gate) allow the network to selectively pass information across the network. This mechanism can help control the flow of information and make the network more efficient at learning from data.LSTMs can process various types of sequential data and are not limited to fixed-length input-output mappings. They can handle variable-length sequences. LSTMs represent an improvement in capturing temporal dynamics, making them suitable for complex sequential tasks. LSTMs can maintain a state that captures historical information. This feature is particularly useful in applications where the context and order of events are critical for making predictions.

[0015] In some embodiments, the device further comprises a motion planning circuit upstream of the machine learning network. The motion planning circuit is configured to generate the vehicle's target trajectory and / or the target curvature value based on environmental sensor data (e.g., from cameras, GPS, LiDAR, radar, and / or IMU). The motion planner can generate the target trajectory based on the vehicle's dynamics, road geometry, and the environment. This target path can be smooth, safe, and efficient, taking into account the vehicle's speed, heading, and possible obstacles.

[0016] In some embodiments, the device further comprises a feedback control circuit configured to determine a feedback wheel angle based on a difference between the target curvature and an actual (measured) curvature of an actual trajectory and based on an actual (measured) speed of the vehicle. The control circuit may complement the machine learning-based feedforward control circuit. The device is configured to determine a wheel angle for the vehicle based on the output prediction of the trained machine learning network for the wheel angle and the feedback wheel angle. The controller may adjust the control action based on the error between the target curvature and the actual curvature. When lateral controlling the vehicle, it may continuously adjust the steering based on the difference between the actual vehicle path and the target path.The feedforward control part can provide a basic control action based on the machine learning model and the target curvature, while the feedback control part can continuously adjust this control to correct deviations to ensure accurate and stable vehicle handling.

[0017] According to another aspect, the present invention proposes a method for controlling the lateral dynamics of a vehicle. The method comprises using a trained machine learning network configured to map input data comprising a desired curvature value and one or more speed values ​​to an output prediction of a corresponding wheel angle for the vehicle.

[0018] In some embodiments, the input data of the machine learning network includes a current curvature value indicating the current curvature of a target trajectory, a current speed value indicating the current longitudinal speed of the vehicle, at least one past speed value indicating the past longitudinal speed of the vehicle, and at least one future curvature value indicating the future curvature of the target trajectory.

[0019] According to a further aspect, the present invention proposes a method for supervised training of a machine learning network for controlling the lateral dynamics of a vehicle. The training method comprises simulating target curvature values ​​corresponding to a predefined speed profile and a predefined wheel angle profile. The training method further comprises using the simulated target curvature values, the predefined speed profile, and the predefined wheel angle profile as ground-truth data for training the machine learning network. The simulated target curvature values ​​and the predefined speed profile serve as ground-truth input data, and the predefined wheel angle profile serves as the corresponding ground-truth output data.

[0020] In some embodiments, the simulation of the desired curvature values ​​is performed based on a kinematic or dynamic linear single-track model of the vehicle. Those skilled in the art, benefiting from the present disclosure, will recognize that more complex models can be used to simulate ground-truth data, such as the full-vehicle dynamics model considering all four wheels of the vehicle and without neglecting vertical dynamics. Actual real-world (measurement) data can also be used as ground-truth data. However, this may not be feasible in some scenarios.

[0021] Some examples of devices and / or methods are described below by way of example only and with reference to the accompanying figures, in which Fig. 1 shows a block diagram of an apparatus for controlling the lateral dynamics of a vehicle according to an embodiment; Fig. 2 shows an example of a machine learning network configured as a fully connected artificial neural network according to one embodiment; Fig. 3 shows an example of a machine learning network configured as an LSTM according to one embodiment; Fig. 4 shows a flowchart of a method for controlling the lateral dynamics of a vehicle according to an embodiment; Fig. 5 shows the linear single-track model; Fig. Figure 6 illustrates a concept for generating training / ground truth data; Fig. 7 shows a block diagram of an apparatus for controlling the lateral dynamics of a vehicle using feedforward and feedback control according to one embodiment; Fig. Figure 8 shows performance results for controlling lateral dynamics in a first target trajectory scenario (U-turn); Fig. Figure 9 shows performance results for controlling lateral dynamics in a second target trajectory scenario (infinite); Fig. 10 shows performance results for controlling lateral dynamics in a third target trajectory scenario (Fast Track); and Fig. Figure 11 shows performance results for controlling lateral dynamics in a fourth target trajectory scenario (random track).

[0022] Fig. Figure 1 shows a schematic block diagram of a device 10 according to the invention for controlling the lateral dynamics of a vehicle. The device 10 can be implemented on special hardware designed for automotive applications. Such hardware components should be highly reliable, durable, and capable of real-time processing. For example, the device 10 can be implemented in an electronic control unit (ECU). An ECU dedicated to lateral dynamics can process inputs from various sensors (such as camera, LiDAR, radar, and wheel speed sensors), execute control algorithms (feedforward and / or feedback controllers), and send commands to actuators (such as the steering system). Modern vehicles can have multiple ECUs for different functions.The device 10 may, for example, be implemented in one or more microcontrollers, in one or more FPGAs (Field-Programmable Gate Arrays), in one or more automotive-grade processors, or in one or more DSPs (Digital Signal Processors).

[0023] The device 10 comprises a trained machine learning network 11 which is configured to apply input data 12 comprising a target curvature value k0 and one or more speed values ​​v0 to an output prediction 13 of a wheel angle δ ff for the vehicle. The trained machine learning network 11 can act as a feedforward controller for lateral vehicle dynamics, which refers to the movement of a vehicle in the horizontal plane and focuses primarily on steering, turning, and maintaining a lane. The trained machine learning model 11 can anticipate the vehicle's behavior based on the inputs 12 and calculate the wheel angle δff , which matches the inputs 12, as control parameters for a steering actuator. The trained machine learning model 11 can predict the necessary steering input to follow a target path or trajectory.

[0024] The machine learning network 11 can be an alternative to conventional analytical methods of feedforward control. In such analytical methods, only the road curvature is used to calculate the desired wheel angle. The kinematic or dynamic single-track model, for example, tends to saturate its wheel angle at higher speeds. By taking the vehicle speed into account when calculating the feedrate, the machine learning model can also be developed based on the (actual) longitudinal speed of the vehicle, the current desired curvature, and past and future (desired) curvatures. The input to the neural networks is selected based on the correlation of the actual wheel angle.

[0025] The machine learning network 11 can be implemented as one or more neural networks capable of learning complex patterns from large amounts of data. For lateral guidance, such models can be trained using data capturing various driving scenarios, road conditions, and vehicle responses. The machine learning network 11 can be implemented as a fully connected neural network, a convolutional neural network (CNN), and / or a recurrent neural network (RNN), including LSTMs (long short-term memory networks). While CNNs are effective at processing spatial information (such as camera images), RNNs and LSTMs excel at handling sequential data (such as time-series sensor data). The machine learning network 11 can be based on reinforcement learning.In this approach, the model 11 can learn the optimal control policy by interacting with (or simulating) the environment and receiving feedback in the form of rewards or punishments based on its performance.

[0026] As in the example of the Fig. 2, the input data 12 of the machine learning network 11 may include a current curvature value k0, which indicates the current curvature of a target trajectory, and a current speed value v0, which indicates the current vehicle current. (Actual) longitudinal speed, wherein at least one past speed value v p_t indicates the past longitudinal speed of the vehicle, and at least one future curvature value k p_t indicates the future curvature of the target trajectory. In the Fig. In the example shown in Figure 2, the input data 12 comprise two past (actual) speed values ​​v p_100 and vp_200 , where v p_100 the past (actual) longitudinal speed of the vehicle 100 ms ago and v p_200 indicates the past (actual) longitudinal speed of the vehicle 200 ms ago. The speed values ​​v0, v p_100 v p_200 may, for example, originate from one or more wheel speed sensors. Those skilled in the art having benefited from the present disclosure will recognize that other past speed values ​​are also possible and depend on the application, the vehicle, etc. The speed values ​​v0, v p_100 , v p_200 can be used directly or independently to determine one or more acceleration values. Thus, the input data 12 of the machine learning network can additionally or alternatively include one or more acceleration values ​​indicating the current and past acceleration of the vehicle.

[0027] In the Fig. 2, the input data 12 further comprises a past curvature value k p_100 which indicates the past curvature of the target trajectory (e.g. 100 ms ago) k f_100 a first future curvature value indicating the future curvature of the target trajectory at a first look-ahead time (e.g., 100 ms) and a second future curvature value k f_200 which indicates the future curvature of the target trajectory at a second look-ahead time (e.g., 200 ms). Those skilled in the art having benefited from the present disclosure will recognize that other past and future curvature values ​​are possible and depend on the application, vehicle, etc.

[0028] In the Fig. In the example shown in Figure 2, the input data 12 provided to the input layer 21 of the fully connected network 11 thus consists of a vector of seven values ​​v p_200 , v p_100 , v0, kp_10o , k0, k f_100 and k f_200The number of neurons in this layer corresponds to the number of features in your input data 12 (here: seven). The number of neurons in the first hidden layer 22 is a parameter that can be chosen based on the complexity of the task. For example, 512 neurons could be used. Each neuron in the first hidden layer 22 is connected to every neuron in the input layer 21. The neurons in the first hidden layer 22 typically have an activation function, such as ReLU (Rectified Linear Unit), sigmoid, or tanh. This activation function introduces nonlinearity into the model, allowing it to learn more complex patterns. The second hidden layer 23 also contains a number of neurons (this can be the same or different from the first hidden layer 22) and is fully connected to the first hidden layer 22.Neurons in the second hidden layer 23 also have an activation function. The final layer is the output layer 24. The number of neurons here depends on the task. For example, in a classification task with 10 classes, you have 10 neurons, each representing a class. The activation function in the output layer 23 can be chosen based on the type of task: softmax for multi-class classification, sigmoid for binary classification, or no activation (or linear activation) for a regression task. Each connection between neurons has an associated weight. These weights are adjusted during the training process. The network 11 learns by adjusting these weights to minimize the difference between the actual output (ground truth) and the predicted output.During training, ground truth data is fed to the input layer 21, which then propagates through the hidden layers 22, 23, and finally the output is generated at the output layer 24. The network 11 can use algorithms such as backpropagation and an optimization technique (such as gradient descent) to update the weights and reduce the error during training.

[0029] Fig. 3 shows an embodiment in which the machine learning network 11 is configured as an LSTM.

[0030] The example LSTM 11 from Fig. Figure 3 includes three LSTM layers / cells 31-0, 31-1, 31-2. Note that a different number of LSTM layers / cells can also be used. Each LSTM cell 31-0, 31-1, 31-2 consists of three gates—an input gate, a forget gate, and an output gate (not shown). These gates control the flow of information into and out of the respective cell and decide what is retained or discarded in the cell state. The input gate determines how much of the new information is stored in the cell state. The forget gate decides which information from the cell state is discarded. The output gate determines which information from the cell state is used to calculate the cell's output. The cell state is the key feature of LSTM cells. It acts as a kind of conveyor belt that transports relevant information throughout the data sequence. Changes in the cell state are carefully regulated by the gates.The hidden state is the output portion of each LSTM cell, which is used for calculations in the current cell and is passed to both the cell state and the next LSTM cell in the sequence. The LSTM 11 example from [source: LSTM 11]. Fig. 3 comprises one or more dense layers (fully connected layers) 32 between the LSTM layers / cells and a final output layer 33 for further processing of the data.

[0031] Fig. 4 shows a flowchart of a method 40 for supervised training of the machine learning network 11 for controlling the lateral dynamics of a vehicle.

[0032] The training method 40 includes an act 41 for simulating target curvature values ​​corresponding to a predefined speed profile and a predefined wheel angle profile. The training method 40 further includes an act 42 in which the simulated target curvature values, the predefined speed profile, and the predefined wheel angle profile are used as ground truth data for training the machine learning network 11. The simulated target curvature values ​​and the predefined speed profile serve as ground truth input data, and the predefined wheel angle profile serves as the corresponding ground truth output data.

[0033] The simulation of the 41 (ground truth) desired curvature values ​​can be performed based on a kinematic or dynamic linear single-track model of the vehicle. Mathematical models of a vehicle's dynamics exist at varying levels of complexity, ranging from simple linear single-track models to three-dimensional high-fidelity multi-body models. The required modeling depth depends on the application. Thus, those skilled in the art, benefiting from the present disclosure, will recognize that more complex models can be used to simulate ground-truth data for input and output, such as full-vehicle dynamic models that consider all four wheels of the vehicle and do not neglect vertical dynamics. Actual real-world (measurement) data can also be used as ground-truth data. However, this may not be feasible in some scenarios.

[0034] The kinematic linear single-track model enables a physically plausible approximation of vehicle lateral dynamics without major modeling effort. This simple model results from a number of assumptions made in the foresight of the modeling process. First, all heave, roll, and pitch movements of the vehicle are neglected. Second, it is assumed that the mass is located at the center of gravity S of the vehicle. Third, two imaginary tires centered between the actual front and rear tires are used, as in Fig. introduced, and the tire forces act at these imaginary contact points V and H. It is also assumed that the load distribution between the front and rear axles is constant. The speed v at the center of gravity S is assumed to be constant. Furthermore, it is assumed that the vehicle rolls freely so that the longitudinal forces F v , x and Fh, x on the tires can be neglected.

[0035] As in Fig. As shown in Figure 6, a functional mock-up unit (FMU) of the single-track model and the speed control can be used to generate ground truth data. Based on a speed requirement v p_200 , v p_100 , v0 from the ground truth speed profile, the speed control unit (FMU) can provide braking and acceleration torque. The single-track model assumes a ground truth target wheel angle δ ff from the ground truth wheel angle profile, the ground truth acceleration and the braking torque value from the speed control unit (FMU). The ground truth target curvature values ​​k0, k f_100 and k f_200can be generated based on the brute-force method of all combinations of wheel angle and speed, and based on the random wheel angle and random speed profile. The generated data is preprocessed and trained.

[0036] As in the embodiment of Fig. 7, the device 10 may further comprise a motion / position planning circuit 71 upstream of the machine learning-based feedforward controller 11. The motion / position planning circuit 71 is configured to determine the target trajectory of the vehicle and / or the target curvature values ​​k0 (k f_100 , k f_200 ) based on environmental sensor data (e.g., from cameras, GPS, LIDAR, RADAR, and / or IMU) that is input to the motion / position planning circuit 71. The motion / position planning circuit 71 can generate the desired trajectory and thus the desired curvature values ​​k0 (k f_100 , k f_200) based on vehicle dynamics, road geometry, and the environment. This target trajectory can be smooth, safe, and efficient, taking into account the vehicle's speed, heading, and potential obstacles.

[0037] Furthermore, the device 10 may include a feedback control circuit 72 configured to determine a feedback wheel angle feedback δ based on a difference / error between the target curvature k0. (k f_100 , k f_200 ) from the movement / position planning circuit 71 and an actual curvature k0, actual ( kf_100,actua l, kf_200,actua l) an actual trajectory detected by environmental sensors and based on an actual speed v0 ( vp_10 0 and vp_200) of the vehicle. The feedback control circuit 72 may complement the machine learning-based feedforward control circuit 71. The device 10 may be configured to calculate a wheel angle for the vehicle based on the 71 output prediction δ f f of the trained machine learning-based feedforward control for the wheel angle and the feedback wheel angle feedback δ from the feedback controller 72. The device 10 may be configured to determine a wheel angle for the vehicle based on a sum of δ f f and δ feedback to be determined.

[0038] The Fig. show the performance of the proposed machine learning-based feedforward controller 11 compared to a conventional feedforward controller based on the bicycle (single track) kinematics model. Fig. 8 shows a target track in the form of U-turns, Fig. 9 shows a target track in the form of an infinity symbol, Fig. 10 shows a target track in the form of a fast track and Fig. Figure 11 shows a target trajectory in the form of a random track. It can be seen that the proposed machine learning-based feedforward control 11 achieves good performance for all target trajectories, close to or even superior to the performance of the conventional feedforward control.

[0039] The described aspects and features relating to a particular one of the preceding examples may also be combined with one or more of the further examples in order to replace an identical or similar feature of that further example or to additionally introduce the features into the further example.

[0040] Examples may further be or refer to a (computer) program containing program code to perform one or more of the above methods when the program is executed on a computer, processor, or other programmable hardware component. Thus, steps, operations, or processes of various of the methods described above may also be performed by programmed computers, processors, or other programmable hardware components. Examples may also include program storage devices, such as digital data storage media, that are machine-, processor-, or computer-readable and encode and / or contain machine-executable, processor-executable, or computer-executable programs and instructions.Program storage devices may include, for example, digital storage devices, magnetic storage media such as magnetic disks and magnetic tapes, hard disk drives, or optically readable digital data storage media. Other examples may also include computers, processors, control units, field-programmable logic arrays (FPLAs), field-programmable gate arrays (FPGAs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), integrated circuits (ICs), or system-on-chip (SoCs) systems programmed to perform the steps of the methods described above.

[0041] It is further understood that the disclosure of multiple steps, processes, operations, or functions disclosed in the description or claims should not be interpreted as necessarily requiring these operations to be performed in the described order, unless expressly stated in the specific case or required for technical reasons. Therefore, the foregoing description does not limit the performance of multiple steps or functions to any particular order. Furthermore, in further examples, a single step, function, process, or operation may comprise and / or be divided into multiple sub-steps, functions, processes, or operations.

[0042] If some aspects have been described with reference to a device or system, these aspects should also be understood as a description of the corresponding method. For example, a block, device, or functional aspect of the device or system may correspond to a feature, such as a method step, of the corresponding method. Methodically described aspects are therefore also understood to include the description of a corresponding block, element, property, or functional feature of a corresponding device or system.

[0043] The following claims are hereby incorporated into the Detailed Description, each claim being understood to stand on its own as a separate example. Although a dependent claim in the claims refers to a particular combination with one or more other claims, other examples may include a combination of the dependent claim with the subject matter of another dependent or independent claim. Such combinations are hereby expressly contemplated unless it is stated in a particular case that a particular combination is not intended. Furthermore, features of one claim should also be included for any other independent claim, even if that claim is not directly defined as dependent on that other independent claim. Reference symbol 10 Device for controlling the lateral dynamics of a vehicle 11 machine learning network 12 Input data 13 Initial forecast 21 Input layer 22 hidden layer 23 hidden layer 24 Output layer 31 LSTM cell 32 fully connected layer 33 Output layer 41 Simulating target curvature values 42 Using the simulated target curvature values ​​as ground truth data 71 positioner 72 controllers

Claims

[1] Device (10) for controlling the lateral dynamics of a vehicle, the device (10) comprising: a trained machine learning network (11) configured to map input data (12) comprising a target curvature value and one or more speed values ​​to an output prediction (13) of a corresponding wheel angle for the vehicle. [2] The device (10) of claim 1, wherein the input data (12) of the machine learning network (11) comprises a current curvature value indicating the current curvature of a target trajectory, a current speed value indicating the current longitudinal speed of the vehicle, at least one past speed value indicating the past longitudinal speed of the vehicle, and at least one future curvature value indicating the future curvature of the target trajectory. [3] The apparatus (10) of claim 1 or 2, wherein the input data (12) of the machine learning network (11) comprises a current speed value indicative of the current longitudinal speed of the vehicle, a first past speed value indicative of the longitudinal speed of the vehicle before a first time, and a second past speed value indicative of the longitudinal speed of the vehicle before a second time. [4] Apparatus (10) according to any one of the preceding claims, wherein the input data (12) of the machine learning network (11) comprises a current curvature value indicating the current curvature of a target trajectory, a past curvature value indicating the past curvature of the trajectory, a first future curvature value indicating the future curvature of the trajectory at a first look-ahead time, and a second future curvature value indicating the future curvature of the trajectory at a second look-ahead time. [5] Device (10) according to one of the preceding claims, wherein the machine learning network (11) comprises a fully connected neural network with an input layer (21), at least two hidden layers (22; 23) and an output layer (24). [6] Device (10) according to one of claims 1 to 4, wherein the machine learning network (11) comprises a LSTM (Long Short-Term Memory) network with an input layer, at least two LSTM layers (31) and an output layer (33). [7] The apparatus (10) of any preceding claim, further comprising a motion planning circuit (71) configured to generate the target trajectory of the vehicle and the target curvature value based on environmental sensor data. [8] Device (10) according to one of the preceding claims, further comprising a feedback control circuit (72) configured to determine a feedback wheel angle based on a difference between the desired curvature and an actual curvature of an actual trajectory and an actual speed of the vehicle; and wherein the device (10) is configured to determine a wheel angle for the vehicle based on the output prediction of the trained machine learning network (11) for the wheel angle and the feedback wheel angle. [9] A method for controlling the lateral dynamics of a vehicle, the method comprising: Using a trained machine learning network (11) configured to map input data (12) comprising a target curvature value and one or more speed values ​​to an output projection of a corresponding wheel angle for the vehicle. [10] The method of claim 9, wherein the input data 812) of the machine learning network (11) comprises a current curvature value indicating the current curvature of a target trajectory, a current speed value indicating the current longitudinal speed of the vehicle, at least one past speed value indicating the past longitudinal speed of the vehicle, and at least one future curvature value indicating the future curvature of the target trajectory. [11] A method (40) for supervised training of a machine learning network (11) for controlling the lateral dynamics of a vehicle, the method comprising: Simulating (41) target curvature values ​​corresponding to a predefined speed profile and a predefined wheel angle profile; and Using (42) the simulated target curvature values, the predetermined speed profile and the predetermined wheel angle profile as ground truth data for training the machine learning network, where the simulated target curvature values ​​and the predefined speed profile serve as ground truth input data and the predefined wheel angle profile serve as corresponding ground truth output data. [12] The method (40) of claim 11, wherein simulating the desired curvature values ​​is performed based on a linear single-track model of the vehicle.

Images