Method and system for training behavioral planner models using mimic learning

Abstract

The invention relates to a method for training a behavioral planner model using imitation learning, the behavioral planner model comprising at least: a backbone network for extracting potential features from input states; and a planner network for predicting an action on the basis of the potential features, the method having the following steps: providing (S1) training data which associates an input state with a corresponding action; extracting (S2) potential features from the input state of the training data by means of the backbone network; training and / or adjusting (S3) the behavioral planner model by optimizing an imitation loss function and according to at least one further predictive uncertainty loss function based on the training data and / or the behavioral planner model.

Description

Technical Field

[0001] The present invention relates to a method and system for training a behavior planner model using imitation learning. Background Art

[0002] Currently, data-based, in particular learning-based methods are used to plan the behavior of autonomous vehicles. This behavior planning is required in particular in urban areas and / or in urban situations. In this case, so-called imitation learning is usually used.

[0003] Imitation Learning (German: Nachahmungslernen) is a machine learning method in which an agent tries to imitate the behavior of an expert. Here, the agent is provided with a collection of example data, which is generated by a human expert or another well-behaved system. The agent then learns to imitate these examples by developing a model that predicts the expert's actions based on a given input state.

[0004] Imitation learning is often used in environments where it is difficult to define or formulate an explicit optimization function. Instead of designing an explicit reward function that tells the agent which actions to perform, imitation methods can directly adopt the actions of an expert. There are various imitation learning methods, including: Behavior Cloning, in which the agent simply copies the actions of the expert; and Inverse Reinforcement Learning, in which the agent attempts to reconstruct the expert's reward function.

[0005] In addition to autonomous driving, imitation learning has applications in a variety of other fields, including robotics, game development, and virtual agents. This imitation learning enables agents to learn complex tasks by mimicking the behavior of experts rather than learning the expert's behavior from scratch.

[0006] If a model trained in this way experiences situations similar to those included in the training data set when applied, for example during autonomous driving, the model will generate and / or indicate and / or initiate behavior similar to that of an expert.

[0007] Here, when using trained models, two types of uncertainty are basically distinguished. On the one hand, there is aleatory uncertainty, which is understood as uncertainty or noise in the training data set. This noise can be caused, for example, by uncertainties and / or errors in the data detection by sensors or in the upstream processing of the detected sensor data. On the other hand, there is epistemic uncertainty, which is understood as uncertainty in the learned model itself. This uncertainty can be formed, for example, if the training data is incomplete and / or erroneous. As a result, it can happen, for example, that the autonomous vehicle gets into a situation during operation that is not depicted in the training data or is depicted incompletely or incorrectly. For both types of uncertainty, there are already methods for estimating the corresponding uncertainties.

[0008] Here, the aleatoric uncertainty is usually learned directly by the model. Thus, for example, in addition to the imitation behavior, the model also generates an evaluation of the quality of the imitation of the behavior. This evaluation can be performed, for example, by variance estimation, which indicates: the reliability or correctness of the generated results or the imitated behavior. For example, such an evaluation can be achieved by minimizing an additional loss function (English: Loss, such as log-likelihood loss) during training.

[0009] For the estimation of epistemic uncertainty, so-called ensembles can be used in particular. This is known, for example, from "Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles" by Lakshminarayanan et al. In this case, the learned model is varied in a predetermined manner. The uncertainty is then estimated by the variance of the output of the model variation. If all variants generate similar outputs for the same input, this indicates a high degree of certainty. Conversely, for different outputs, this indicates a high degree of uncertainty.

[0010] Typically, uncertainty in behavior planning is not considered or is only considered implicitly, thereby increasing the risk that a trained model will incorrectly perform and / or adjust the behavior planning in unknown situations and / or situations that are underrepresented or uncertain in the training data. Summary of the invention

[0011] The object of the present invention is therefore to provide an improved method and / or system.

[0012] This object is achieved by a method for training a behavior planner model using imitation learning according to the features of patent claim 1. This object is achieved by a system for training a behavior planner model using imitation learning according to the features of patent claim 9.

[0013] According to a first aspect, there is provided a method for training a behavior planner model using imitation learning, the behavior planner model comprising at least: a backbone network for extracting latent features from input states; and a planner network for predicting actions based on the latent features, wherein the method comprises the following steps:

[0014] - providing training data, preferably generated by experts, and associating input states with corresponding actions;

[0015] - Extract latent features from the input states of the training data with the help of the backbone network;

[0016] - Training and / or adjusting the behavior planner model by optimizing the imitation loss function and according to at least one further prediction uncertainty loss function based on the training data and / or the behavior planner model. Preferably, the at least one further prediction uncertainty loss function is optimized together with the imitation loss function.

[0017] It should be understood that the steps and other optional steps according to the present invention do not necessarily have to be performed in the order shown, but can also be performed in another order. In addition, other intermediate steps can be provided. Each step can also include one or more sub-steps, without departing from the scope of protection of the method according to the present invention.

[0018] In the present case, a variety of methods from the field of machine learning are combined with each other and used to estimate uncertainty in order to optimize behavior planning, especially behavior planning of autonomous vehicles and / or robots and / or other mobile machines. This results in the following advantages: using the method currently described for training a behavior planner model, it is possible to automatically adapt to existing and / or given uncertainties. This behavior planning can be carried out, for example, by reducing the speed of the vehicle and / or robot and / or mobile machine, as long as the uncertainty can be reduced by this speed reduction. Preferably, the given uncertainty is included in the training of the behavior planner model in the present case, so that the trained model takes into account these given uncertainties as comprehensively as possible. In particular, these uncertainties are part of the training process. Compared with previous methods, this is especially an advantage, in which uncertain situations or situations in which there are high uncertainties must be explicitly and especially manually implemented by developers. In this way, there is a risk that all situations related to the model are not taken into account. For example, trajectories are generated in a targeted manner, which appear with lower epistemic uncertainty when they are known to be within the probability distribution. In the present case, the current uncertainty can be explicitly considered. As a result, it is possible, for example, to display and / or output and / or visualize and / or output this uncertainty to the driver or operator acoustically and / or haptically and / or in another way. As a result, the interpretability and / or acceptability of the entire system is increased. In addition, the method makes it possible to optimize the training process with respect to perturbations of the training data.

[0019] For behavior planning in autonomous vehicles and / or robots and / or mobile machines, in the present case, a model is trained or made to learn by means of imitation learning, which generates a behavior plan based on an input, for example an environment model with static and / or dynamic objects and / or a map, for example in the form of road signs. During the training of the behavior planner model, the output of the model to be trained is preferably compared and evaluated with the expert behavior in a specific situation. This comparison and / or this evaluation is called imitation loss and can be described by the imitation loss function. By means of this comparison and / or this evaluation, the model to be trained is preferably optimized according to known methods, wherein in particular the error described by the imitation loss function is minimized.

[0020] The method does not rely on a model with specific inputs and outputs. In general, a model can be trained in which training is performed with the aid of a cost function to evaluate predetermined and / or sampled trajectories.

[0021] Preferably, the backbone network is an integral part of a deep learning architecture, such as a convolutional neural network (CNN) for computer vision tasks. Preferably, it is a layer or a group of layers that is used to convert the original input state (e.g., an image) into a higher, more abstract representation or so-called "latent feature". In a computer vision scenario, the backbone network may, for example, have a series of convolutional and / or pooling layers that recognize patterns and / or features in an image. These features may, for example, be edges, shapes, textures, and / or higher-level objects. The output of the backbone network is preferably forwarded to other layers in the architecture in order to perform specific tasks such as classification, object recognition, or segmentation.

[0022] The planner network is preferably designed to predict actions based on latent features and / or hidden representations of the data. Here, the planner network preferably uses machine learning techniques or artificial intelligence techniques. Such techniques can be based on neural networks using deep learning, such as recurrent neural networks (RNN) and / or long short-term memory (LSTM) networks.

[0023] In one embodiment, the behavior planner model is trained using a predictive model to: prioritize actions that translate into scenarios that are known to exist in the training data; and / or automatically adjust these actions based on the predictive model to reduce uncertainty.

[0024] Through this, in the case of increased uncertainty in the model, even if there is no more passive behavior in the training data, the behavior is also explicitly implemented. In this way, other expert knowledge about the behavior under uncertainty can also be implemented into the model.

[0025] The losses are weightedly added to form a loss, with which the model is optimized. In this case, care must be taken that the autonomous vehicle does not behave too passively and that progress is ensured.

[0026] In the present case, "scenario" refers to the situation, environment and / or state in which the system that uses the behavior planner model to work is located. The scenario can include various aspects, such as the location and movement of objects, the presence of obstacles, the characteristics of the environment, and other relevant features. In this context, the behavior planner model is used to select actions that are converted into known scenarios. In other words, the behavior planner model selects actions based on trained predictions, which involve how these actions convert the current situation of the system into the situation detected in the training data. This means that the model is trained to give priority to the following actions, which cause the system to enter a situation that the system is familiar with based on the training data. Through this method, the behavior planner model can learn to select appropriate actions for specific scenarios and therefore improve decision-making in specific contexts. This can be used in various application fields such as robots, autonomous driving, or other systems that need to react and act based on observations and training data.

[0027] In the context of the present application, a "prediction model" refers to a model or prediction function that is set up to predict future states or events based on current conditions and, if necessary, historical data. The prediction model is preferably used in conjunction with a behavior planner model. Preferably, the prediction model is intended to predict the future state of a scene or system (the situation in which the system is located). For example, the prediction model may attempt to predict the position and movement of objects present in the environment of the system, or the expected development of the scene over a certain period of time. Preferably, the prediction model may be included in the behavior planner model, or provided as an independent model. The prediction model is preferably pre-trained.

[0028] In one embodiment, the additional prediction uncertainty loss function describes aleatory uncertainty and / or epistemic uncertainty. Therefore, in the present method, preferably, for behavior planning, in addition to the imitation loss function, other loss functions or losses are considered, through which these uncertainties can be estimated, and preferably also these uncertainties can be evaluated.

[0029] In one embodiment, the at least one further prediction uncertainty loss function of the aleatory uncertainty has a variance function, in particular a variance matrix. In order to learn the aleatory uncertainty, preferably, the output of the model to be trained is expanded with a variance function. Next, the landmarks are preferably defined by 2D (i.e., two-dimensional) or 3D (i.e., three-dimensional) positions and a variance matrix. In order to evaluate the uncertainty estimate, for example, a loss function is used, which in particular compares probabilities. Such a loss function can be, for example, a log-likelihood loss.

[0030] In one embodiment, the at least one further prediction uncertainty loss function has an epistemic uncertainty about the scenario, which is conditioned on the action prediction of the behavior planner model, which action prediction depends on the prediction of the planner network. Preferably, the assessment of epistemic uncertainty is achieved by the prediction of the actions, in particular the scenario, of the behavior planner model and an estimation of the degree of presence of the predicted scenario in the training data.

[0031] A simulation prediction model can also be provided, which is set up to evaluate the degree of representation of the future scene (implemented by the action) in the data set for the scene. In particular, when the model is subjected to insufficient representation in the training data or is incorrectly represented or not represented, cognitive uncertainty occurs in the model architecture. Preferably, this can be reduced by an additional training loss or another cognitive loss function. Preferably, the training loss or loss function is more inclined to lead to the behavior of the predetermined situation. For this reason, the following example is preferred, which evaluates the degree of representation of the situation predicted by the model in the training data. This can be achieved, for example, by simulating the prediction, especially the planning of the trained model at a predetermined time or a predetermined time period in the future. Subsequently, preferably, based on the simulation prediction, the corresponding state of the model is evaluated. This can be achieved, for example, by the integration of the model and / or by an estimation model. Such an estimation model can, for example, have a "novelty module" to estimate the novelty of the situation in the data set. The prerequisite here is that the model learns to represent the (training) data set. At the beginning of training, preferably, the evaluation of the integration of the untrained model is poor or low. Therefore, to speed up the training process, previously trained or pre-trained models are advantageous.

[0032] In another embodiment, preferably, other losses or loss functions for other uncertainties may be included for the uncertainty estimation. In this way, for example, the speed reduction may be evaluated with a reduced uncertainty. By this, in case of increased uncertainty of the model, a more passive behavior is explicitly implemented even if this behavior is not present in the training data. In this way, other expert knowledge about the behavior under uncertainty may also be implemented into the model.

[0033] In one embodiment, the simulation loss function and the at least one other prediction uncertainty loss function are weighted and added to form a total loss function, wherein the total loss function is optimized. Therefore, preferably, the individual losses or loss functions are weightedly added to form a total loss or total loss function, which is used to optimize the model. In this case, it should be noted that the autonomous vehicle and / or robot and / or mobile machinery does not behave too passively, so that the progress of the training is ensured by the rapid optimization of the total loss function.

[0034] In one embodiment, during training and / or adjustment, input states and / or actions are perturbed or at least partially masked, taking into account cognitive and aleatory uncertainties. Additionally, preferably during training, the inputs of the algorithm or model to be trained are perturbed and / or at least partially masked. For example, masking can be performed by adding additional noise. As a result, the robustness of the model trained in this way with respect to such changes in the training data or input is increased, in particular when the change has a relevant influence on the behavior of the model. With the aid of uncertainty estimation, the method is easy to expand in that changes with a high influence on the model, in particular due to high uncertainty, are performed more frequently than changes with a low influence, in particular due to low uncertainty. As a result, advantages are generated relative to standard perturbations of the training data, in which the changes are performed equally regardless of their relevance. Additionally, the developer obtains information about the vulnerabilities of the model. Such vulnerabilities can be represented, for example, by perturbations with a high influence on the uncertainty, which can already be included during the training process.

[0035] In one embodiment, the adjustment of the behavior planner model by optimizing and / or minimizing the imitation loss function using the prediction uncertainty loss function has:

[0036] - comparing and / or evaluating the output of the behavior planner model to be trained with the actions of the expert for a specific input state, especially for a specific input state in a specific situation; and / or

[0037] - at least one model parameter and / or at least one hyperparameter of the behavior planner model is adapted, in particular iteratively, based on the training data.

[0038] In the present case, "hyperparameters" refer to adjustable parameters of the behavior planner model that are not learned during the training process, but must be specified prior to training. Unlike model parameters that are learned during the training based on the training data, hyperparameters are preferably external settings that affect the learning process and control the performance and / or behavior of the model. Preferably, hyperparameters affect: the way the learning algorithm works, the convergence speed of the behavior planner model, and the degree of adaptation of the behavior planner model to new (training) data. Examples of hyperparameters might be learning rate, number of hidden layers and neurons, batch size, regularization factor, and / or number of training iterations (epochs).

[0039] In one embodiment, the method further comprises at least one of the following steps:

[0040] - Validate the performance of the trained behavior planner model against validation data; and / or

[0041] - if the performance of the trained behavior planner model meets specified criteria, storing the trained behavior planner model, in particular for using the model in a technical application; and / or

[0042] - if the performance of the trained behavior planner model does not meet the specified criteria, repeatedly adjusting the behavior planner model by optimizing model parameters, particularly with the use of additional and / or extended training data; and / or

[0043] - Repeating this step at least until at least one predetermined criterion is met or an interruption criterion is reached.

[0044] According to a second aspect, a system for training a behavior planner model using imitation learning is provided, the behavior planner model comprising at least: a backbone network for extracting latent features from input states; and a planner network for predicting actions based on the latent features. The system has: a providing device, the providing device is designed to: provide training data, the training data are preferably generated by an expert and associate input states with corresponding actions; and a computing device, the computing device is set up to: extract latent features from the input states of the training data by means of the backbone network, and train and / or adjust the behavior planner model by optimizing an imitation loss function and according to at least one other prediction uncertainty loss function based on the training data and / or the behavior planner model.

[0045] It should be understood that the providing device and the computing device can be implemented in a single component or in different components. It should also be understood that the embodiments and examples mentioned for the method claimed in accordance with the first aspect relate to the claimed system in the same or at least equivalent manner, rather than claiming protection for the system alone.

[0046] According to the present invention, a computer program is also claimed, which has a program code for executing at least part of the method according to the present invention in one of its embodiments when the computer program is executed on a computer. In other words, according to the present invention, a computer program (product) is claimed, which includes instructions, which, when the program is executed by a computer, cause the computer to execute the method / steps of the method according to the present invention in one of its embodiments.

[0047] According to the invention, a computer-readable data carrier is also proposed, which has the program code of a computer program for executing at least part of the method according to the invention in one of its embodiments when the computer program is executed on a computer. In other words, the invention relates to a computer-readable (storage) medium, which includes instructions which, when executed by a computer, cause the computer to execute the method / steps of the method according to the invention in one of its embodiments.

[0048] The described embodiments and refinements can be combined with one another as desired.

[0049] Further possible embodiments, developments and implementations of the invention also include combinations of features of the invention described above or below with reference to the exemplary embodiments that are not explicitly mentioned. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] The accompanying drawings are intended to provide a further understanding of the embodiments of the present invention. These drawings illustrate the embodiments and are used to describe and explain the principles and designs of the present invention in the context.

[0051] Further embodiments and a number of the advantages mentioned can be derived with reference to these drawings. The elements represented in these drawings are not necessarily shown to scale with respect to one another.

[0052] in:

[0053] Figure 1 A flow chart of the present method for training a behavior planner model is shown;

[0054] Figure 2 A schematic diagram of various uncertainties is shown graphically;

[0055] Figure 3 A block diagram showing various loss functions associated with uncertainty; and

[0056] Figure 4 A block diagram of the present method for training a behavior planner model is shown.

[0057] In the figures of these drawings, unless otherwise indicated, the same reference numerals denote the same or functionally identical elements, components or assemblies. DETAILED DESCRIPTION

[0058] Figure 1 A schematic flow chart of a method for training a behavior planner model using imitation learning is shown, the behavior planner model comprising at least: a backbone network for extracting latent features from input states; and a planner network for predicting actions based on the latent features.

[0059] In any embodiment, the method may be at least partially performed by a system 1, for which the system may include a plurality of components not shown in more detail, such as one or more providing devices and / or at least one computing device. It should be understood that the providing device may be designed together with the computing device, or may be different from the computing device. The system may also include a storage device and / or an output device and / or a display device and / or an input device.

[0060] In particular, the computer-implemented method comprises at least the following steps:

[0061] In step S1 , training data are provided, which are preferably generated by an expert and associate input states with corresponding actions.

[0062] In step S2, latent features are extracted from the input states of the training data with the help of the backbone network.

[0063] In step S3, the behavior planner model is trained and / or adapted by optimizing an imitation loss function and according to at least one further prediction uncertainty loss function based on the training data and / or the behavior planner model.

[0064] Figure 2 An exemplary representation of aleatoric and epistemic uncertainty in a data-based, learned behavior planner model is shown. In this case, which quantities and / or units are assigned to the abscissa and / or ordinate is currently irrelevant. The depicted data has a sinusoidal curve around the ordinate. For example, the ordinate can depict or present a curve of a ground truth. The presented points represent data points of a signal with noise around the sinusoidal curve of the ground truth.

[0065] The distribution of data points around the sinusoidal curve of the ground truth represents the aleatory uncertainty. The aleatory uncertainty is denoted by reference numeral 200. The epistemic uncertainty is denoted by reference numeral 202 and is plotted in areas of the ground truth curve where no data points are present. In such areas where no data points are present, the trained model can only make estimates with insufficient accuracy because, for example, the model has not been provided with any training data for such data points during training. The ground truth curve has reference numeral 204.

[0066] The behavior at intersections can vary greatly, for example, due to occlusions (e.g. caused by freight traffic). A standardly trained behavior planner or a standardly trained model would probably ignore occlusions if this situation was not specifically depicted in the training data. In contrast, a model trained in the current manner would generate greater uncertainty in the case of unknown occlusions and thus, for example, cause the speed of the vehicle controlled by the model to be reduced. For example, a currently trained model could cause a slow approach to an intersection due to the high uncertainty detected, even though this maneuver behavior was not depicted in the original training data used to train the model.

[0067] Furthermore, in similar cases, very different behaviors can be included in a data set based on human drivers, especially depending on the passivity or aggressiveness of the driver. In this case or for this case, the present model is preferably trained to have high uncertainty and therefore behave passively. In contrast, a model learned by standard may marginalize the behavior of the data set, or focus on a specific behavior. Alternatively, a known model must launch multiple plans for possible model behaviors, which are evaluated by means of an additional cost function.

[0068] Figure 3 An exemplary calculation of a total loss function 300 for the current training of a behavior planner model is shown. In this case, additional prediction uncertainty loss functions 304, 306, 308, 310 are added to the imitation loss function 302 in order to form the total loss function 300. In this case, the individual prediction uncertainty loss functions 304, 306, 308, 310 are weighted with each other by weights β1, β2, β3, β4. The prediction uncertainty loss functions 304, 306 describe aleatoric uncertainty. The prediction uncertainty loss functions 304, 308 describe epistemic uncertainty. The prediction uncertainty loss functions 304, 310 describe at least one additional prediction uncertainty loss function of an additional uncertainty.

[0069] Figure 4 A block diagram of the present method for training a behavior planner model 400 is shown. The behavior planner model 400 is trained based on training data 402 by means of imitation learning 404. The training data 402 or scenario associates input states with corresponding actions. Potential features can be extracted from the input states of the training data by means of a backbone network, which is part of the behavior planner model 400. The behavior planner model 400 is trained and / or adjusted by optimizing an imitation loss function 406 and according to at least one prediction uncertainty loss function 408 based on the training data 402 and / or the behavior planner model 400.

[0070] In the present case, the prediction uncertainty loss function 408 depicts aleatoric uncertainty 410 and epistemic uncertainty 412. In the present case, the prediction uncertainty loss function 408 for the aleatoric uncertainty 410 has a variance function, in particular a variance matrix. Based on these actions, the future state of the scenario or situation in which the system is located can be predicted, preferably with the aid of a prediction model 414. The epistemic uncertainty 412 can be determined by an estimation model 416 based on the predictions of the prediction model 414.

[0071] The estimation model 416 evaluates how well the scene predicted by the prediction model 414 is depicted in the training data. The loss values ​​for the various uncertainties may preferably be added together to form a cumulative loss value.

Claims

1. A method for training a behavior planner model using imitation learning, the behavior planner model comprising at least: The backbone network is used to extract latent features from the input state; and a planner network for predicting actions based on latent features, wherein the method comprises the following steps: - providing (S1) training data, said training data associating input states with corresponding actions; - extracting (S2) latent features from the input states of the training data with the aid of the backbone network; - Training and / or adapting (S3) the behavior planner model by optimizing an imitation loss function and according to at least one further prediction uncertainty loss function based on the training data and / or the behavior planner model.

2. The method according to claim 1, wherein: The behavior planner model is trained using a predictive model to: prioritize actions that translate into known scenarios present in the training data; and / or automatically adjust the actions to reduce uncertainty.

3. The method according to claim 1 or 2, wherein: The additional prediction uncertainty loss function describes aleatoric uncertainty and / or epistemic uncertainty.

4. The method according to claim 3, wherein: At least one further prediction uncertainty loss function has an aleatory uncertainty as a function of a variance function, in particular a variance matrix.

5. The method according to claim 3 or 4, wherein: At least one additional prediction uncertainty loss function has an epistemic uncertainty about the scenario that depends on the actions of the behavior planner model.

6. A method according to any one of the preceding claims, wherein: The imitation loss function and the at least one further prediction uncertainty loss function are weighted and added to form a total loss function, wherein the total loss function is optimized.

7. The method according to any one of claims 3 to 5, wherein: During the training and / or the adapting ( S4 ), the input state and / or the action is perturbed or at least partially masked taking into account cognitive and aleatoric uncertainties.

8. A method according to any one of the preceding claims, wherein: The adjustment of the behavior planner model by optimizing and / or minimizing the imitation loss function while using the prediction uncertainty loss function has: - comparing and / or evaluating the output of the behavior planner model to be trained with the actions of the expert for a particular input state; and / or -Adjusting at least one model parameter and / or at least one hyperparameter of the behavior planner model based on the training data.

9. A method according to any one of the preceding claims, wherein: The method further comprises at least one of the following steps: - Validate the performance of the trained behavior planner model against validation data; and / or - if the performance of the trained behavior planner model meets specified criteria, storing the trained behavior planner model, in particular for use of the behavior planner model in a technical application; and / or - if the performance of the trained behavior planner model does not meet the specified criteria, repeatedly adapting the model by optimizing model parameters, in particular using additional and / or extended training data; and / or - Repeating this step (S4) at least until at least one predetermined criterion is met or an interruption criterion is reached.

10. A system (1) for training a behavior planner model using imitation learning, the behavior planner model comprising at least: The backbone network is used to extract latent features from the input state; and a planner network for predicting actions based on the latent features, wherein the system has: - a providing device, the providing device being designed to: provide training data, the training data associating input states with corresponding actions; and - A computing device configured to extract latent features from input states of the training data by means of the backbone network, and to train and / or adjust the behavior planner model by optimizing an imitation loss function and based on at least one additional prediction uncertainty loss function based on the training data and / or the behavior planner model. 11 . A computer program having a program code for executing at least part of the method according to claim 1 when the computer program is executed on a computer. 12 . A computer-readable data carrier having a program code of a computer program for executing at least parts of the method according to claim 1 when the computer program is executed on a computer.

Images