Method for generating a training data set for training a machine learning model for controlling a vehicle, computer-implemented method for training a machine learning model, computer-implemented method for generating control signals, computer program, machine-readable storage medium, system for a vehicle and vehicle

Abstract

Method for generating a training dataset for training a machine learning model for controlling a vehicle (9), wherein the training dataset comprises environmental data of driving situations, the method comprising the following steps: • Receiving first input data generated by simulation and / or recorded by at least one first sensor (6), wherein the first input data comprises the environmental data (T100); • Analyzing properties of the first input data (T101); • Transforming the first input data into vectors based on the properties (T102); • Receiving second input data generated by simulation and / or recorded by at least one second sensor (6), wherein the second input data comprises at least one control signal, the control signal comprising at least one signal for controlling the vehicle (9), wherein the first input data and second input data are synchronized (T103);• Generating control signal class labels based on the second input data, wherein each control signal class label marks a specific value for controlling the vehicle (9) (T104); • Labeling at least some part of the first input data with the corresponding control signal class labels (T105).;

Description

The invention relates generally to the field of civil autonomous driving. More particularly, the invention relates first to a method of generating a training dataset for training a machine learning model for controlling a vehicle. Second, the invention relates to a computer-implemented method for training a machine learning model. The invention furthermore relates to a computer-implemented method for generating control signals, a computer program, a machine-readable storage medium, a system for a vehicle and a vehicle.The following definitions apply to the entire disclosure content.In the field of autonomous driving, the use of artificial intelligence, for example machine learning models such as artificial neural networks, is state of the art. In the context of machine learning, data compensation is used to compensate for under- or over-represented labels at the output. For this purpose, there are various proven methods, such as over / undersampling (also referred to as upsampling / downsampling) or re-sampling (also referred to as resampling) with different ratios. However, these methods focus on the output labels. Therefore, if the performance of a data-controlled model depends on inputs and outputs in the training data, they are not sufficient. In particular, for autonomously controlling a vehicle having an end-to-end, or E2E, driving function, it is important to train a machine learning model with an optimized training data set.Many machine learning models are trained with image data as input data. The image data may be input (input / feature) and control signal may be output (output / label). This data may be collected using a simulation or with real vehicles. For example, during frequent straight-ahead driving, it may happen that an over-representation of the steering angle 0° is contained in the data. This may result in a trained machine learning model always predicting a steering angle of 0° regardless of the input image, and the vehicle using the trained model would therefore not correctly drive in turns, for example. The solution to this problem may be to compensate for steering angles for data, so that under-represented cornering is often present in the training data. This allows the machine learning model to learn how to properly steer the vehicle in a curve. There are cases, however, that cannot be solved with data compensation, which is known from the prior art. An example is a red traffic light. Here, there is a strong accumulation of neutral steering angle data with the accelerator pedal not depressed. Now, if data compensation is used, data with red traffic lights is less sampled (since these usually correspond to a steering angle of ∼ 0°). This results in the vehicle being able to drive turns, for example, but may ignore red lights, because the machine learning model has viewed red lights too rarely as input data during training.It is therefore an object of the present invention to provide an improved method of generating a training dataset for training a machine learning model for controlling a vehicle.This object is achieved by the subject matter of the independent claims. Advantageous embodiments are given in the dependent claims.A first aspect of the invention relates to a method for generating a training dataset for training a machine learning model for controlling a vehicle, wherein the training dataset comprises environmental data of driving situations. The steering of the vehicle may include a longitudinal and / or transverse guidance of the vehicle. It may also include activation of a turn signal or horn or other activity that must be performed by the vehicle itself if the task is not performed by a human driver. Environmental data may include image data, sound data, or other data recorded by automotive environmental detection sensors, such as a camera, radar, lidar, microphone, or simulated by computer-implemented simulations based on sensor models. For example, an alarm signal of an emergency vehicle can be considered as sound data. The environmental data may be preprocessed. Driving situations may include static and / or dynamic objects. They may include, for example, traffic lights, pedestrian crossings, cornering, straight ahead driving, and many more.The method comprises the following steps:• receiving first input data generated by simulation and / or recorded by at least one first sensor, the first input data comprising the environmental data;• analysing properties of the first input data;• transforming the first input data into vectors based on the properties;• receiving second input data generated by simulation and / or recorded by at least one second sensor, the second input data comprising at least one control signal, the control signal comprising at least one signal for controlling the vehicle, the first input data and second input data being synchronized;• generating control signal class labels based on the second input data, each control signal class label labeling a specific value for controlling the vehicle;• Label at least a portion of the first input data with the corresponding control signal class labels.Machine learning models may include artificial neural networks.The first input data may be generated by a GAN (generative adventaric network), for example. They may also be recorded by at least one sensor, such as a camera, radar sensor, lidar sensor, acoustic sensor, or other type of sensor. The first input data can be stored in a storage medium. According to an aspect, transforming the first input data may include encoding the first input data. Using a decoder, the original first input data can be reconstructed from the coding. Storing the first input data in the storage medium is particularly advantageous in cases where the coding is non-linear. In one aspect, the first input data is embedded in a latent space that may have lower dimensionality than the input space.The properties of the first input data can comprise, for example, driving situations, such as traffic lights, pedestrian transitions, driving straight ahead, cornering, bus stops or other aspects.The transformation of the first input data into vectors based on the characteristics may be performed, for example, by a machine learning model, in particular by a trained neural network such as YOLO, ResNet, or others. Each vector corresponds to assigned first input data. The vectors may have as many dimensions as necessary.The vectors may be placed in an embedding space to facilitate comparison of the vectors. For example, a large Euclidean distance between two vectors may indicate different characteristics of the first input data.The second input data may be, for example, a trajectory generated by simulation. The second input data may also be recorded by at least one sensor, such as a steering angle sensor, an accelerometer, or others.The control signal may include a braking value and / or an acceleration value and / or a steering angle of the vehicle.Advantageously, the first input data and the second input data are synchronized in order to assign the correct second input data to the first input data.The control signal class identifiers may, for example, mark a specific braking value and / or an acceleration value and / or a steering angle or others relevant to the control of the vehicle on the basis of the assigned environmental data of driving situations.Depending on the destination of the generated training dataset, only selected first input data or all first input data may be labeled with the corresponding control signal class labels. In some cases where a specific training goal is to be achieved, it may be useful to identify only some of the first input data and use only this identified data for training a machine learning model. For example, a goal may be training and / or retraining a machine learning model only at specific driving situations such as cornering, or a goal may be training the model only at first input data assigned to balanced vectors, for example. There may be several objectives depending on the case of use.A second aspect of the invention relates to a computer-implemented method for training a machine learning model on control signals for controlling a vehicle. The method comprises the following steps:• Receiving a training dataset, wherein the training dataset comprises environment data of driving situations and control signal class labels, wherein the environment data has been generated by simulation and / or recorded by at least one sensor, wherein the control signal class labels mark a braking value and / or an acceleration value and / or a steering angle of the vehicle;• forward propagation of the environmental data;• generating output data comprising the control signal class labels predicted by the machine learning model;• comparing the predicted control signals with the control signal class labels of the training dataset;• back propagation of the machine learning model with deviations between the predicted control signals and the control signal class labels of the training dataset;• Learning control signals by gradient-based optimization of weights of the machine learning model.Computer-implemented means that the steps of the method are carried out by a data processing device, for example a computer, a computer system, a computer network, for example a cloud system, a hardware of a control unit or parts thereof. According to one aspect, the training dataset may have been generated according to the previously disclosed method for generating a training dataset for training a machine learning model for controlling a vehicle. As mentioned above, environmental data may include image data, sound data, and other data. For example, an alarm signal of an emergency service vehicle can be regarded as sound data. The environmental data may be preprocessed. The environmental data may be generated by, for example, a GAN (generative adventaric network). They may also be recorded by at least one sensor, such as a camera, radar sensor, lidar sensor, acoustic sensor, or other type of sensor. The training procedure is a supervised learning procedure. That is, the machine learning model further optimizes the predicted control signals based on predetermined control signal class labels by optimizing the weighting factors of the machine learning model such that the control signals further align with the predetermined control signal class labels.The machine learning model thus trained may be deployed in a control unit of a vehicle.A third aspect of the invention relates to a computer-implemented method for generating control signals for controlling a vehicle. The method comprises the following steps:• receiving sensor data of a driving situation;• processing the sensor data using a machine learning model trained to determine control signals for controlling a vehicle;• generating a control signal according to the control signal determined by the machine learning model;• Outputting the control signal to a control unit for controlling at least one longitudinal and / or transverse guidance of the vehicle.In one aspect, the machine learning model has been trained according to the previously disclosed computer-implemented method for training a machine learning model on control signals for controlling a vehicle. The control unit may control the longitudinal and / or transverse guidance of the vehicle. The longitudinal guide may include acceleration and / or braking. The transverse guide may comprise steering the vehicle. The control unit may further control components such as a horn, a direction indicator, a warning indicator, a backup lamp, a stop lamp, or others.Another aspect of the invention relates to a computer program for generating control signals for controlling a vehicle. The computer program comprises instructions which, when executed by a computer, cause the computer to carry out the steps of the method for generating control signals for controlling a vehicle. The instructions of the computer program according to the invention may comprise machine instructions, source code or object code written in assembly language, an object oriented programming language, for example C++, or in a procedural programming language, for example C. The computer program is executed in whole or in part on an integrated circuit or in a remote system comprising a cloud. The integrated circuit may include at least one GPU (graphics processing unit), a CPU (central processing unit), and / or an FPGA (field programmable gateway array).Another aspect of the invention relates to a computer readable storage medium comprising instructions which, when executed by a computer, cause the computer to carry out the steps of the method of generating control signals for controlling a vehicle. The storage medium may be, for example, a non-volatile permanent memory or a volatile memory. A storage medium may be, for example, a USB stick, a hard disk, a memory card, or a CD-ROM. However, the storage medium may also be, for example, a cloud in which the computer program is stored.Another aspect of the invention relates to a data carrier signal which carries the computer program for classifying control signals for controlling a vehicle. By means of a data carrier signal, the computer program for classifying control signals for controlling a vehicle can be transmitted, for example, from a cloud to a control unit of an automated vehicle, for example, via a so-called update by radio.Another aspect of the invention relates to a system for a vehicle, wherein the system comprises at least one sensor for receiving environmental data of driving situations, at least one hardware module which processes the data according to the method for generating control signals for controlling a vehicle, and a control unit which controls the vehicle autonomously on the basis of the control signals. The sensor may be a camera, lidar sensor, radar sensor, acoustic sensor, or other sensor. The hardware module may include a trained machine learning model. As already described, the control unit may control the longitudinal and / or transverse guidance of the vehicle. The longitudinal guide may include acceleration and / or braking. The transverse guide may comprise steering the vehicle. The control unit may further control components such as a horn, a turn signal indicator, a hazard warning light, a backup lamp, a brake light, or others.Another aspect of the invention relates to a vehicle, wherein the vehicle comprises a system comprising at least one sensor for receiving environmental data of driving situations, at least one hardware module processing the data according to the method for generating control signals for controlling a vehicle, and a control unit autonomously controlling the vehicle on the basis of the control signals. The vehicle may be an automated vehicle, for example, a road vehicle such as a car, bus, truck, shuttle, passenger transport, cargo transport, or others. In one aspect, the vehicle is an autonomous driving vehicle.An embodiment relates to a method for generating a training dataset for training a machine learning model for controlling a vehicle, the method additionally comprising the following steps for optimizing the energy consumption of the training:• determining a representation of the control signal class labels after generating control signal labels;• Balancing the representation of the control signal class labels;• filtering first input data on the basis of the balanced representation of each control signal class label, wherein only the first input data whose vectors have a maximum distance from one another are filtered out.In order to obtain an optimized training data set, it is useful to optimize both the first input data and the control signal class labels. Depending on the representation of the control signal class labels, balancing may include, for example, oversampling under-represented control signal class labels and / or subsampling over-represented control signal labels and / or adjusting control signal class weights inversely proportional to control signal class frequencies in the input data. Other compensation methods may also be carried out. The advantage of additionally filtering the first input data after compensating for the representation of the control signal class labels is that only the first input data which are as different as possible is used. This means that, for example for a steering angle of 0°, not only red traffic lights are considered in the first input data, but also straight-ahead driving situations and others. According to the embodiment, only the filtered first input data is marked with the corresponding control signal class labels.Another embodiment relates to another method for generating a training dataset for training a machine learning model for controlling a vehicle, the method additionally comprising the following steps for optimizing the energy consumption of the training:• determining a representation of each generated control signal class label and determining a representation of the vectors after generating control signal class labels;• Balancing the representation of the control signal class labels and the representation of the vectors.Another possibility for obtaining an optimized training dataset is compensating the representation of the control signal class labels according to the determined representation, for example by oversampling underrepresented control signal class labels and / or subsampling overrepresented control signal labels and / or adjusting control signal class weights inversely proportional to control signal class frequencies in the input data. Other compensation methods may also be carried out. In addition, the vectors are balanced. Not only the same vectors can be considered as a representation unit, but also vectors having a maximum predefined Euclidean distance, for example. The balancing of the representation of the vectors can be effected, for example, by oversampling underrepresented first input data and / or down sampling overrepresented first input data. Other compensation methods may also be carried out. According to the embodiment, only the first input data assigned to the balanced vectors are marked with the corresponding control signal class labels.Another embodiment relates to another method for generating a training dataset for training a machine learning model for controlling a vehicle, the method additionally comprising the following steps for optimizing the energy consumption of the training:• clustering the vectors into different clusters after generating control signal class labels;• sampling first input data in each cluster.Clustering identifies groups in embeddings and therefore allows, for example, to identify which first input data contains, for example, the same driving situations. Sampling first input data in each cluster has the effect of balancing the representation of the first input data. This means that different driving situations, such as traffic lights, pedestrian crossings, cornering and others, are sensed equally. According to the embodiment, only the sampled first input data is marked with the corresponding control signal class labels.The mentioned embodiments have the advantage that the automated comparison of the input and output enables to obtain the most diverse input data of driving situations and to control signal class labels to have the greatest possible diversity in the data. This reduces redundancy in the data and allows much faster training because iteration over less data is required. Moreover, it is possible, for example, to store only the most varied first input data and / or combination of first input data and control signal class labels which are possible, which leads to a great reduction in the required storage space for further use.Another embodiment relates to another method for generating a training dataset for training a machine learning model for controlling a vehicle, the method additionally comprising the following steps for retraining a machine learning model for controlling a vehicle that has failed or at least generated inaccurate control signals in a real world application:• receiving third input data after generating control signal class labels, the input data comprising environmental data of driving situations that caused a trained machine learning model to fail or at least generate inaccurate control signals;• analysing the properties of the third input data;• transforming the third input data into vectors based on the properties;• comparing the vectors of the first input data with the vectors of the third input data;• Determining the vectors of the first input data which have the closest distance to the vectors of the third input data.Normally, if a trained machine learning model does not adequately avoid some driving situations, for example pedestrian crossings, it is necessary to retrain the model for these situations. For this purpose, new data of these driving situations must be recorded, which corresponds to a very great effort. It is therefore an advantage to apply the method to such problems. Therefore, third input data of these driving situations are transformed into vectors. The third input data may include image data, sound data, or others. The first input data closest to the third input data according to the assigned vectors is marked with the corresponding control signal class labels and used to retraine the model.The invention described above is reproduced with reference to specific embodiments which are illustrated in the appended figures. The following are shown: FIG. 1 is a flowchart of a method for generating a training dataset for training a machine learning model for controlling a vehicle; FIG. 2 is a flowchart of an embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle; FIG. 3 is a flowchart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle; FIG. 4 is a flowchart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle; FIG. 5 is a flowchart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle; FIG. 6 is a flowchart of a method for training a machine learning model on control signals for controlling a vehicle; FIG. 7 is a flow chart of a method for generating control signals for controlling a vehicle; FIG. 8 is a schematic drawing of a computer program for classifying control signals for controlling a vehicle; FIG. 9 is a schematic drawing of a computer readable storage medium; FIG. 10 is a schematic drawing of a data carrier signal; FIG. 11 is a schematic drawing of a vehicle including a system.The reference symbols used in the drawings and their meanings are given in summary form in the following list of reference symbols. In principle, identical parts in the figures are provided with the same reference symbols.FIG. 1 shows a flow chart of a method for generating a training dataset for training a machine learning model for controlling a vehicle 9. the method comprises a step T 100 of receiving first input data generated by simulation and / or recorded by at least one first sensor 6, wherein the first input data comprises the environment data. In addition, the method comprises a step T 101 of analyzing properties of the first input data. Referring to FIG. 1, the method further includes a step T 102 of transforming the first input data into vectors based on the characteristics. Furthermore, FIG. 1 shows a step T 103 of receiving second input data generated by simulation and / or recorded by at least one second sensor 6, the second input data comprising at least one control signal, the control signal comprising at least one signal for controlling the vehicle 9, the first input data and second input data being synchronized. It also shows a step T104 of generating control signal class labels on the basis of the second input data, each control signal class label marking a specific value for controlling the vehicle 9. Finally, the method shown in FIG. 1 comprises a step T 105 of identifying at least some of the first input data with the corresponding control signal class identifiers.FIG. 2 shows a flowchart of an embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle 9. FIG. 2 shows, after step T 104 of generating control signal labels, a step T 106 of determining a representation of the control signal class labels, in addition to the steps described in FIG. 1. Furthermore, the method shown in FIG. 2 comprises a step T 107 of compensating the representation of the control signal class labels and a step T 108 of filtering first input data on the basis of the representation of each control signal class label, wherein only the first input data whose vectors have a maximum distance from one another in the embedding space are filtered out. Finally, the method shown in FIG. 2 also comprises step T 105 as described in the above FIG. 1.FIG. 3 shows a flow chart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle 9. FIG. 3 shows, after step T 104 of generating control signal labels, a step T 109 of determining a representation of each generated control signal class label and determining a representation of the same. Furthermore, the method shown in FIG. 3 comprises a step T 110 of compensating the representation of the control signal class labels and the representation of the assigned first input data. Finally, the method shown in FIG. 3 also includes step T 105 as described above in FIG. 1.FIG. 4 is a flowchart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle 9. FIG. 4 shows, after step T 104 of generating control signal labels, a step T 111 of clustering the vectors in different clusters and a step T 112 of sampling first input data in each cluster, in addition to the steps described in FIG. 1. Finally, the method shown in FIG. 4 also comprises step T 105 as described above in FIG. 1.FIG. 5 shows a flow chart of another embodiment of a method for generating a training dataset for training a machine learning model for controlling a vehicle 9. FIG. 5 shows, after step T 104 of generating control signal labels, a step T 113 of receiving third input data, wherein the input data comprises environmental data of driving situations that have caused a trained machine learning model to fail or at least generate inaccurate control signals, in addition to the steps described in FIG. 1. Further, the method shown in FIG. 5 includes a step T 114 of analyzing the characteristics of the third input data and a step T 115 of transforming the third input data into vectors based on the characteristics. Referring to FIG. 5, the method further includes a step T 116 of comparing the vectors of the first input data with the vectors of the third input data. The method shown in FIG. 5 also comprises a step T 117 of determining the vectors of the first input data that have the closest distance to the vectors of the third input data. Finally, the method shown in FIG. 5 also comprises step T 105 as described in the above FIG. 1.FIG. 6 shows a flow diagram of a method for training a machine learning model on control signals for controlling a vehicle 9. the method comprises a step C 100 of receiving a training dataset, wherein the training dataset comprises environment data of driving situations and control signal class labels, wherein the environment data has been generated by simulation and / or recorded by at least one sensor 6, wherein the control signal class labels mark a braking value and / or an acceleration value and / or a steering angle of the vehicle. The method also comprises a step C 101 of forward propagation of the environment data. Referring to FIG. 6, the method further includes a step C 102 of generating output data including the control signal class labels predicted by the machine learning model, and a step C 103 of comparing the predicted control signals with the control signal class labels of the training dataset. Further, FIG. 6 shows a step C 104 of back propagating the machine learning model with deviations between the predicted control signals and the control signal class labels of the training dataset. Finally, the method shown in FIG. 6 comprises a step C 105 of learning control signals by gradient-based optimization of weights of the machine learning model.FIG. 7 shows a flow chart of a method for generating control signals for controlling a vehicle 9. Referring to FIG. 7, the method further includes a step CV 2 of processing the sensor data using a machine learning model trained to determine control signals for controlling a vehicle 9. Further, FIG. 7 shows a step CV 3 of generating a control signal according to the control signal determined by the machine learning model. Finally, the method shown in FIG. 7 comprises a step CV 4 of outputting the control signal to a control unit 8 for controlling at least one longitudinal and / or transverse guidance of the vehicle 9.FIG. 8 shows a schematic drawing of a computer program 1 for classifying control signals for controlling a vehicle 9.FIG. 9 is a schematic drawing of a computer readable storage medium 3. the computer readable storage medium 3 of the embodiment in FIG. 9 is, for example, a USB stick.FIG. 10 shows a schematic drawing of a data carrier signal 4. according to the exemplary embodiment of FIG. 10, the computer program 1 is transmitted from a cloud using the data carrier signal 4.FIG. 11 shows a schematic drawing of a vehicle 9 including a system 5. The embodiment of the system 5 shown in FIG. 5 comprises a sensor 6 for receiving environmental data of driving situations, a hardware module 7 and a control unit 8 which controls the vehicle 9 autonomously on the basis of the control signals. According to the embodiment of FIG. 11, the vehicle 9 is, for example, a car.List of reference symbolsT 100 - T 117 Method steps C 100 - C 105 Method steps CV 1 - CV 4 Method steps 1 Computer program 2 Computer 3 Computer-readable storage medium 4 Data carrier signal 5 System 6 Sensor 7 Hardware module 8 Control unit 9 Vehicle

Claims

A method for generating a training dataset for training a machine learning model for controlling a vehicle (9), the training dataset comprising environment data of driving situations, the method comprising the steps of: • receiving first input data generated by simulation and / or recorded by at least one first sensor (6), the first input data comprising the environment data (T100); • analyzing characteristics of the first input data (T101); • transforming the first input data into vectors based on the characteristics (T102); • receiving second input data generated by simulation and / or recorded by at least one second sensor (6), the second input data comprising at least one control signal, the control signal comprising at least one signal for controlling the vehicle (9), the first input data and second input data being synchronized (T103); • Generating control signal class labels based on the second input data, each control signal class label marking (T 104) a specific value for controlling the vehicle (9); • Labeling at least a portion of the first input data with the corresponding control signal class labels (T 105).The method according to claim 1, wherein the method additionally comprises the following steps for optimizing the energy consumption of the training: • determining a representation of the control signal class labels after generating control signal labels (T106); • compensating the representation of the control signal class labels (T107); • filtering first input data on the basis of the representation of each control signal class label, wherein only the first input data whose vectors have a maximum distance from each other are filtered out (T108).The method of claim 1, wherein the method additionally comprises the steps of optimizing the energy consumption of the training: • determining a representation of each generated control signal class label and determining a representation of the vectors after generating control signal class labels (T109); • balancing the representation of the control signal class labels and the representation of the assigned first input data (T110).The method of claim 1, wherein the method additionally comprises the steps of optimizing the energy consumption of the training: • clustering the vectors into different clusters after generating control signal class labels (T111); • sampling first input data in each cluster (T112).The method according to claim 1, wherein the method additionally comprises the steps of retraining a machine learning model for controlling a vehicle (9) that has failed or at least generated inaccurate control signals in a real world application: • receiving third input data after generating control signal class labels, wherein the input data comprises environment data of driving situations that have caused a trained machine learning model to fail or at least generate inaccurate control signals (T113); • analyzing the characteristics of the third input data (T114); • transforming the third input data into vectors based on the characteristics (T115); • comparing the vectors of the first input data with the vectors of the third input data (T116); • Determining the vectors of the first input data which have the closest distance to the vectors of the third input data (T117).Computer-implemented method for training a machine learning model on control signals for controlling a vehicle (9), the method comprising the following steps: • receiving a training dataset, wherein the training dataset comprises environment data of driving situations and control signal class labels, wherein the environment data has been generated by simulation and / or recorded by at least one sensor (6), wherein the control signal class labels mark (C100) a braking value and / or an acceleration value and / or a steering angle of the vehicle; • forward propagation of the environment data (C101); • generating output data comprising the control signal class labels predicted by the machine learning model (C102); • comparing the predicted control signals with the control signal class labels of the training dataset (C103); • back propagation of the machine learning model with deviations between the predicted control signals and the control signal class labels of the training dataset (C 104); • learning control signals by gradient-based optimization of weights of the machine learning model (C 105).Computer-implemented method for generating control signals for controlling a vehicle (9), the method comprising the following steps: • receiving sensor data of a driving situation (CV1); • processing the sensor data by means of a machine learning model trained according to claim 6 to determine (CV2) control signals for controlling a vehicle (9); • generating a control signal according to the control signal determined (CV3) by the machine learning model; • outputting the control signal to a control unit (8) for controlling at least a longitudinal and / or transverse guidance of the vehicle (9) (CV4).A computer program (1) for generating control signals for controlling a vehicle (9), the computer program (1) comprising instructions which, when the program (1) is executed by a computer (2), cause the computer (2) to carry out the steps of the method according to claim 7.A computer readable storage medium (3) comprising instructions which, when executed by a computer (2), cause the computer (2) to carry out the steps of the method of claim 7.A data carrier signal (4) carrying the computer program (1) of claim 8.A system (5) for a vehicle (9), the system (5) comprising at least one sensor (6) for receiving environmental data of driving situations, at least one hardware module (7) processing the data according to the method of claim 7, and a control unit (8) autonomously controlling the vehicle (9) on the basis of the control signals.A vehicle (9), wherein the vehicle (9) comprises a system (5) according to claim 11.

Images