DRIVING ASSISTANCE DEVICE AND METHOD FOR IMPLEMENTING AT LEAST A PARTIALLY AUTOMATIC VEHICLE FUNCTION DEPENDING ON A ROUTE TO BE EVALUATED

Description

[0001] The present invention relates to a driver assistance device and method for carrying out an at least partially automatic vehicle function as a function of a route to be evaluated.

[0002] Driver assistance systems for motor vehicles are known from the state of the art, and these are to be further developed, particularly with a view to achieving the most fully autonomous vehicle control possible. However, it is noticeable that the driver assistance systems available today cannot operate completely reliably and error-free in all traffic situations and environments. This particularly applies to traffic situations or environments that are unfamiliar to the respective driver assistance system, for which it has not been trained or optimized, for example. Therefore, in the interests of safety, there is a need for options for handling such situations safely and reliably.

[0003] US 2021 / 0165409 A1 discloses an adaptive control system for an autonomous or semi-autonomous vehicle. The system aims to mimic the decision-making process of a human driver to generate a path suitable for autonomous control of an autonomous or semi-autonomous vehicle. For this purpose, a model-based predictive control system is proposed that uses a motion model, the sequence of states, a parametric probability distribution, and information from a detection system to optimize a cost function with respect to a prediction horizon. From this, a sequence of control commands for one or more actuators of the vehicle is generated. Furthermore, a metric of the vehicle's motion is used.

[0004] US 2014 / 0136414 A1 discloses a method and system for an autonomous neighborhood vehicle traffic network in a community. It proposes calculating a set of predicted behaviors of detected objects within a boundary distance from the autonomous neighborhood vehicle and determining confidence levels for the predicted behaviors. The confidence levels are a number and / or a percentage of the probability of each predicted behavior occurring.

[0005] US 2020 / 0192393 A1 discloses an autonomous driving system that can adapt itself based on the observation of driving situations that arise after deployment in the real environment. For this purpose, it is proposed to combine problem solving with pattern recognition for unknown situations in which problems arise, thus obtaining various types of problem categories by classifying different unknown situations into groups. Furthermore, it is described how to evaluate the safety of such explorations during explorations carried out by a proposed method. For this purpose, it is described, for example, that the risk is measured based on a predefined model that checks a minimum distance to the nearest vehicle and / or pedestrian and / or road barrier, as well as limits for acceleration and speed.

[0006] EP 4 047 316 A1 discloses a method for assessing the familiarity of a route.

[0007] From US 2017 / 0102703 A1 a method and a device for providing a steering reliability map based on driven curvatures and geometry-based curvatures is known.

[0008] The present invention is based on the object of overcoming the disadvantages known from the prior art and of providing a method for carrying out an at least partially automatic vehicle function, in particular a driving function, of a vehicle as a function of a route to be evaluated, as well as a driver assistance system, the use of which offers the highest possible level of safety.

[0009] The object is achieved according to the invention by the subject matter of the independent claims. Advantageous embodiments and further developments of the invention are the subject matter of the dependent claims.

[0010] A method according to the invention (preferably computer-implemented) for carrying out an at least partially automatic and / or partially autonomous (preferably autonomous) vehicle function, in particular a driving function and / or vehicle guidance, of a vehicle as a function of a route to be evaluated by means of a particularly processor-based driver assistance device, comprises a plurality of (in particular computer-implemented) method steps, which are carried out in particular with a correspondingly configured electronic computing device (which is preferably a component of the driver assistance device or is in communication with it) and / or preferably with the driver assistance device, in particular automatically.

[0011] The driver assistance system can, for example, be or comprise a predefined or trained, and preferably trainable, model for performing the at least partially automatic or partially autonomous (preferably autonomous) vehicle function, in particular a driving function and / or vehicle guidance. This can be implemented, for example, in the form of a particularly deep, i.e., multi-layer, artificial neural network or a strategy learned using a learning algorithm.

[0012] In a (particularly computer-implemented) method step of the method according to the invention, a plurality of clusters of route data relating to at least one known route are provided and / or retrieved (for example from an internal vehicle storage device and / or an external server), wherein the clusters group the route data, in particular section-wise (relating to the known route) according to predetermined geometric parameters.

[0013] The term "grouping route data" by clusters means, in particular, that each of the plurality of clusters is assigned a plurality (or a group) of route data (relating to at least one known route). This assignment is particularly unambiguous. Each route data point is therefore assigned to exactly one cluster.

[0014] In other words, by providing the plurality of clusters of route data relating to at least one known route, a plurality of route data relating to at least one known route is provided, wherein each route data point is uniquely (exactly) assigned to one cluster of the plurality of clusters. A plurality of route data points is assigned to each cluster.

[0015] The assignment or grouping of the route data points to the respective cluster is preferably the result of an assessment of the route data point with respect to the specified geometric parameters. This preferably ensures that route data points similar in terms of the specified geometric parameters are assigned to the same cluster. In particular, the grouping is the result of subdividing the known route(s) into several (smaller) route sections, wherein the subdivision is carried out in such a way that the route data points with respect to the same route section fulfill similar properties with respect to the specified geometric parameters and / or are assigned to the same cluster.

[0016] By assigning the individual route sections on a section-by-section basis (as opposed to assigning individual route data points on the route on a point-by-point basis), causally, geometrically and / or temporally related route data points are assigned to the same cluster.

[0017] In addition to the geometric route properties that describe a causal relationship or similarity between route data points, temporal information or a minimum number of points per cluster can be determined to take into account the segment length of a group.

[0018] In particular, therefore, temporally and / or spatially and / or geometrically related route data points of the route (i.e. all route data points of a particular temporally and / or spatially and / or geometrically related route section of the route, which particularly preferably has a predetermined minimum time or minimum length), are assigned to the same cluster.

[0019] The individual route data points can refer to different, known routes. Two different (known) routes can have a route segment that is essentially identical or (highly) similar with respect to the specified geometric parameters. The route data points of these two different (known) routes are given priority (as part of the grouping process) and / or are assigned to the same cluster.

[0020] The plurality of clusters of route data can preferably be the result of a previous training process of the driver assistance device, in which the driver assistance device was trained, for example, based on the route.

[0021] The vehicle function to be performed by means of the driver assistance device is preferably a driving function of the vehicle, such as longitudinal guidance of the vehicle and / or lateral guidance of the vehicle.

[0022] A route can be known (to the driver assistance device) if, for example, the driver assistance device has been trained or optimized based on this route, or if the driver assistance device has already performed its task or function along this route - in real life or in a simulated manner - i.e., for example, if it has already guided a vehicle autonomously, semi-autonomously, or at least partially automatically along the respective route. A route can also be considered known if the route is known to another driver assistance device (the vehicle or another vehicle). Likewise, a route can be considered known (to the driver assistance device) if, for example, the performance of the driver assistance device along this route has already been determined.

[0023] The route to be evaluated can be a current route or a section of route currently being traveled, based on the vehicle's current driving situation. The route to be evaluated can also be a future and / or planned or intended route, which can be specified, for example, by a user of the vehicle (the driver assistance system) and / or entered via a human-machine interface.

[0024] The route data particularly indicate the geometric properties of the at least one known route or the plurality of known routes. This route data is preferably grouped, i.e., clustered, according to the geometric properties in a property space spanned by corresponding predefined geometric parameters. The geometric properties can therefore be concrete values ​​of the predefined (geometric) parameters for the respective route, i.e., form data points, point clouds, or point groups in the property space. The parameters or the geometric properties can be or describe, for example, local curve radii or curvatures, directions of curvature, changes in curvature, roadway widths, spatial route coordinates, distances of a trajectory from a road edge, and / or the like.The route data can, for example, be provided computer-aided for simulated unknown routes and / or recorded during a journey of a measuring vehicle along the routes using appropriate sensor devices.

[0025] In other words, clusters are and / or are formed based on trajectory planning variables (e.g. reference curvature or curve curvature, speed and acceleration) as well as previously experienced driving scenarios, or clusters already formed in this way are provided.

[0026] Furthermore, (particularly in a computer-implemented method step), trend data, which indicate (at least) a course of the route to be evaluated, are provided, in particular recorded by means of (at least) one sensor device. For example, the trend data can be characteristic of a geometric course of the route to be evaluated and / or of a course (particularly at least in sections) of driving dynamics parameters (particularly of the vehicle) along the route to be evaluated. For example, the trend data can include movement data, such as speed or acceleration along the route. The movement data can be current vehicle data and / or, for example, forecast or average values ​​related to a large number of vehicles and / or (predetermined) limit values.

[0027] Preferably, history data are provided for the route to be evaluated, which are characteristic of the geometric properties of the route to be evaluated with respect to the geometric parameters specified (for the route data with respect to the at least one known route).

[0028] For the route to be evaluated, the geometric properties, i.e., in particular, the corresponding data points in the aforementioned property space, are determined, in particular from the historical data, analogously to the route data of the at least one known route. In other words, the data or properties that were also provided for the at least one known route are determined for the respective route to be evaluated.

[0029] The sensor device is preferably a sensor device of the vehicle, which records and / or determines the (in particular current) historical data and which preferably stores this historical data for further processing (in particular by the driver assistance device) in a (in particular vehicle-internal) storage device and / or transmits it to the driver assistance device. It is also conceivable that the sensor data are at least partially additionally or alternatively transmitted to a storage device external to the vehicle, for example an external server such as a backend server. It is also conceivable that at least one of the sensor devices recording the historical data is a vehicle-external sensor device.This may, for example, be the sensor device of another road user such as a vehicle ahead or an infrastructure facility, with the history data preferably being transmitted to the vehicle and in particular to the driver assistance device via a wireless communication connection.

[0030] The course data may alternatively or additionally be non-vehicle-related data, such as map data that indicate a geometric course of the route to be assessed or are characteristic of it, and / or average driving dynamics parameters (such as a speed and / or acceleration profile along the route to be assessed), which can, for example, be predicted, simulated and / or derived from the driving behavior of other vehicles on the route to be assessed.

[0031] Furthermore, the clusters are applied to the historical data (particularly in a computer-implemented method step) in order to preferably divide the route to be evaluated into route sections corresponding to the clusters and thereby assign each individual route section (preferably exactly) to a cluster. Route sections corresponding to the clusters are understood in particular to mean that the respective route sections are most similar to the clusters assigned to them in terms of their geometric properties (with respect to the predefined geometric parameters) and / or that a correlation between the geometric properties (with respect to the predefined geometric parameters) of the respective route sections and the clusters assigned to them (in the property space) is the highest.

[0032] The application of the clusters to the historical data preferably results in an assignment of the clusters, in particular section by section, to the route to be evaluated. The assignment of the clusters is preferably based on the route to be evaluated and in particular on the respective route sections (in particular also determined by applying the clusters) of the route to be evaluated. Preferably, an assignment of a cluster to the route to be evaluated or a route section is selected, for example, such that the route to be evaluated or the route section of the route to be evaluated has the highest correlation (in particular with regard to the geometric properties) with this cluster compared to the other clusters of the (entire) plurality of clusters (in the property space).

[0033] Preferably (particularly by applying the clusters to the historical data), the historical data (particularly data point by data point) are assigned to a cluster, preferably at most and particularly preferably exactly. This allows the route to be evaluated to be broken down into route sections, each of which is individually assigned to one of the clusters. This advantageously allows the route to be evaluated section by section, and by assigning a cluster to each section, for example, the familiarity of the respective route section (and thus of the entire route to be evaluated) can be evaluated. This offers the advantage of a very precise analysis of the route to be evaluated.

[0034] When applying the clusters to the historical data, a distance between the geometric properties of the respective route to be evaluated and the geometric properties of the at least one known route in the spanned property space is preferably determined (in particular automatically). For this purpose, a distance between the corresponding data points in the property space or, for example, a distance of the respective geometric property, i.e. of the respective data point, to a centroid, i.e. a geometric center of gravity or midpoint of groups or clusters of the geometric properties of the known routes, can be determined. This can in each case be the centroid of the cluster within whose boundaries the respective geometric property, i.e. the respective corresponding data point, of the route to be evaluated lies. The respective distance to the nearest centroid can also be determined.Depending on the number of geometric properties determined for the route to be evaluated, several individual distances can be determined in this way. These can then be further processed separately or individually. If necessary, an average distance of several data points to the centroid(s) can also be determined.

[0035] Preferably (in particular analogous to the formation of clusters from route data in relation to at least one known route described above) a route-section-by-section assignment of the individual route sections of the route to be evaluated (in contrast to a point-by-point assignment of individual route data points of the route to be evaluated) and thus in particular an assignment of causally, geometrically and / or temporally related route data points to the same cluster.

[0036] In particular, route data points of the route to be evaluated that are temporally and / or spatially and / or geometrically related (i.e., all route data points of a route section of the route to be evaluated that is, in particular, temporally and / or spatially and / or geometrically related, and which particularly preferably has a predetermined minimum time or minimum length) are assigned to the same cluster.

[0037] Based on the specified distance or distances, for example, the familiarity of the respective route to be evaluated can be specified or determined. For example, the inverse of the distance can be used directly as a measure of familiarity, so that a greater distance between the geometric properties indicates a lower familiarity, i.e., a lower degree of familiarity, of the route to be evaluated. Likewise, familiarity can be determined based on the specified distance according to a predefined calculation rule.

[0038] Preferably, a plurality of geometric properties that were recorded or determined at a plurality of measuring points along the route to be evaluated are provided and / or recorded and directly assigned to the existing clusters.

[0039] Likewise, if the route to be evaluated has a multitude of geometric properties that were determined at a multitude of measurement points along the route, these can in turn form clusters in the property space. These clusters resulting for the respective route to be evaluated can then be compared with the clusters of the known routes. The distance between the geometric properties can then be determined or given, for example, as similarity, degree of agreement, overlap, or centroid distance between the cluster(s) of the respective route to be evaluated and the cluster(s) of the known routes.

[0040] In other words, based on the geometric properties, a degree of similarity or agreement between the route being evaluated and the known routes can be determined. This can be done for the respective route as a whole, for individual sections of the route, or for individual points along the route.

[0041] In a further (particularly computer-implemented) method step, at least one uncertainty quantity is determined that is characteristic of an uncertainty regarding the assignment made between at least one of the route sections and the cluster assigned to that route section (in each case). The uncertainty or uncertainty quantity describes, in particular, an error probability for the assignment to a specific cluster. The uncertainty quantity can be characteristic of the uncertainty or confidence of the selected cluster assignment (exactly) of one route section, and preferably for the cluster assignment of several route sections of the route to be evaluated, and particularly preferably for the (route section-by-section) cluster assignment of essentially the entire route to be evaluated.

[0042] For example, the uncertainty quantity can be determined by comparing the distances of the corresponding data points (of the route section(s)) in the property space, or the distance of the respective geometric property, to the respective centroids of the plurality of clusters (of the geometric properties) of the at least one known route. It is also conceivable that, to determine the uncertainty quantity, the (approximately average) distance of the data points of the relevant route section to the cluster assigned to this route section is determined and this distance, or a characteristic value thereof, is compared with a (predetermined) threshold value.

[0043] In a further (particularly computer-implemented) method step, (at least) one control variable is determined as a function of the uncertainty variable. Preferably, the control variable is provided and preferably used to carry out the (at least one) vehicle function (by the driver assistance device). Preferably, the vehicle function is carried out as a function of the cluster assignment of at least one section of the route to be evaluated and preferably as a function of the cluster assignment of essentially the entire route to be evaluated. For each section of the route to be evaluated to which a cluster is assigned, a control variable can be determined, for example as a function of the respectively assigned cluster, and can be provided and used to carry out the vehicle function.

[0044] In other words, by assigning at least one cluster to the route to be evaluated, it is possible to determine which of the provided clusters is most comparable to the new route to be evaluated in terms of predetermined geometric parameters, for example, in terms of geometric parameters of the route and / or, for example, in terms of the speed of the vehicle traveling along the route to be evaluated, the acceleration of the vehicle along the route to be evaluated, and / or the curvature of the route. This allows unknown routes to be evaluated to be evaluated in relation to the already known routes.

[0045] The proposed method offers the advantage of not only determining the familiarity of the route to be evaluated, but also determining the uncertainty value, which allows for an uncertainty estimation to be performed for the evaluation of the route to be evaluated, which is advantageously taken into account when performing the vehicle function. To determine the uncertainty value, it is proposed to evaluate the assignment (particularly section-by-section) of the clusters provided based on known route data to a new route to be evaluated. The uncertainty estimation advantageously functions as a safety aspect for possible interventions of the vehicle function performed by the driver assistance system.If the respective uncertainty value of the route to be evaluated is greater than a predefined threshold value, a predefined safety function of the driver assistance system or a training mechanism can be automatically activated to train the learning model.

[0046] By determining the uncertainty value, the uncertainty of the underlying learning and / or trained model can be advantageously analyzed and taken into account when executing vehicle functions based on the learning and / or trained model.

[0047] Geometric clustering can preferably be used to assess the familiarity of a route for the driver assistance device. In other words, it can be indicated whether or not the respective route to be assessed is similar to one or more other routes already known to the driver assistance device.

[0048] This offers the advantage that the performance of the driver assistance system can be determined particularly objectively. For example, performance can be determined or measured based on specified safety and / or performance parameters, such as compliance with traffic regulations, permitted speeds, safe limits for distances, speeds, accelerations, specified comfort limits, for example, with regard to the abruptness of maneuvers or the rate of changes in speed and / or direction, and / or the like. Likewise, performance can be evaluated in a state space or based on routes or route sections in a state space, where geometric properties of the route or route sections can be specified in the state space.As explained in more detail below, the present invention can then also be used to evaluate the performance of the driver assistance system for routes or route sections that are unknown to the driver assistance system until initial processing. Furthermore, a further or different metric can be used to determine or evaluate, i.e., assess, the performance.

[0049] Geometric clustering can thus be used as an objective quality function for assessing the generalizability of a (learning and / or trained) model underlying a driver assistance system. In particular, the performance of the driver assistance system (described in more detail below) can be determined particularly objectively using the geometric clustering method proposed here.

[0050] The assignment of new driving situations or new routes to be evaluated can then take place online in the vehicle (i.e., particularly during vehicle control). A statement regarding the amount of exploration can be made using a metric to be defined, for example, the distance to the center of the cluster, the temporal progression of driving dynamics parameters, or the agent's certainty in the respective driving situation (combination with Bayesian approaches is possible here).

[0051] In a preferred method, the predefined geometric parameters include a local curve curvature, a local curvature direction in the direction of travel, a local road width, a local distance of a vehicle trajectory along the route from a road edge, a local distance of a vehicle trajectory along the route from a lane edge, local spatial route coordinates along the route, a local speed of a vehicle along the route, and / or a local acceleration of a vehicle along the route. Quantities such as a speed and / or an acceleration or a speed and / or acceleration profile (longitudinal and / or lateral) can preferably also be used to geometrically describe a (vehicle) trajectory.

[0052] The acceleration and speed profiles also preferably provide information about previous geometries. For example, a locally geometrically identical section of road can have different desired speeds and accelerations that vary due to previous road characteristics. An example: a curve after a long straight stretch of driving compared to a combination of curves. Therefore, at least one geometric variable related to the geometric course of the route traveled (e.g., with regard to the curvature of a curve) is preferably determined from an acceleration and / or speed profile (especially along the route).

[0053] In other words, one, several, or all of these parameters are specified in order to define the property space. The geometric properties of the routes are specified or determined as values ​​of these parameters. Local here means that the corresponding parameter values, i.e. the corresponding geometric properties, are determined or apply at a single measuring point on the respective route. There can be a large number of such measuring points along the respective route. In other words, for a particular route, for example, the curvature in the direction of travel can be specified for a large number of measuring points along the route in order to define or specify the geometric properties of the route. By using the parameters proposed here, a wide variety of routes can be robustly characterized with regard to their geometric properties.

[0054] In a further possible embodiment of the present invention, the geometric properties in the property space are grouped using cluster analysis. The resulting clusters, i.e., groups of data points in the property space, are applied to the historical data of the respective route to be evaluated in order to divide the respective route to be evaluated into route sections corresponding to the clusters. The familiarity of the route to be evaluated is then determined individually for these individual route sections.If, for example, a straight section of the route follows a curve, the geometric properties determined along or in the curve can be grouped into a first cluster in the property space with respect to the curve curvature, and the geometric properties determined along or on the straight section with respect to the curvature can be grouped into a second cluster in the property space. For example, based on the spatial route coordinates of the data points or geometric properties belonging to a cluster, the corresponding route sections of the route can then be determined, in particular by assigning (preferably connected) route sections to a (particularly identical) cluster.The section-by-section assessment of familiarity can therefore provide detailed data for the respective route, thus enabling a more precise and robust assessment of the familiarity of the route. This then also enables a more precise and robust assessment of the performance of the driver assistance system.

[0055] The route to be evaluated can therefore be divided into several route sections, whereby for each route section, the geometric properties assigned to it, i.e. the data points assigned to it in the property space, are analyzed based on the clusters of the geometric properties of the known routes. Depending on this, a degree of agreement with the known routes is determined for each route section. The degree of agreement can be determined depending on an assignment certainty, which indicates the certainty or confidence with which a respective data point is or can be assigned to a specific cluster. In other words, assignment certainty describes an error probability for the assignment to a specific cluster.The assignment certainty can be determined or given depending on the respective distance of the respective point to the centroid of the corresponding cluster and / or depending on a distribution of data points of the known routes assigned to the cluster. The assignment certainty can be lower the greater the distance or the more diffuse the respective cluster is. Likewise, the distribution of the data points assigned to the respective cluster or forming the respective cluster can be characterized based on their mean distance to the centroid of the respective cluster, a maximum distance of a data point of the respective cluster to its centroid, and / or a similar metric.

[0056] In a manner analogous to the assignment certainty, the uncertainty in relation to the assignment made between at least one of the route sections and the cluster assigned to this route section can be determined. In this way, the uncertainty can be determined (in particular directly) from the assignment certainty. In particular, the uncertainty can be determined or given as a function of a respective distance of the respective point to the centroid of the corresponding cluster and / or as a function of a distribution of data points of the known routes assigned to the cluster. The uncertainty can be greater, the greater the distance or the more diffuse the respective cluster is. Likewise, the distribution of the data points assigned to the respective cluster orthe data points forming the respective cluster are characterized by their mean distance to the centroid of the respective cluster, a maximum distance of a data point of the respective cluster to its centroid and / or a similar metric.

[0057] The proposed use of cluster analysis allows for an objective and robust grouping of geometric properties. This provides an objective and robust basis for assessing familiarity, which can be a significant advantage over manual or subjective classification or familiarity assessment, especially for longer or more complex routes. This is the case because it has been shown that manual or subjective classification of routes based on their geometric similarity by eye often does not provide objectively comprehensible, robust results for assessing the actual familiarity of the respective route, i.e., the geometric similarity to one or more other routes, and thus ultimately for assessing the performance or generalizability of the driver assistance system or a learning method used for it.

[0058] In a further possible embodiment of the present invention, the geometric properties in the property space are grouped by means of an iterative cluster analysis. In this process, clusters and their cluster centroids are iteratively determined subject to a minimization of the average distances of the geometric properties, i.e. the corresponding data points in the property space, to the cluster centroids. In this process, for example, new geometric properties, i.e. data, can be taken into account piece by piece or for each route during or after a run or iteration step, which can then be put in relation to the clusters or cluster centroids already determined up to that point. In this way, a gradual refinement of the clusters or a refined determination of the cluster centroids can be achieved. This ultimately makes it possible to robustly determine the popularity of the respective clusters based on the clusters orCluster centroids can be improved for the route to be evaluated. In other words, a particularly reliable assignment of the geometric properties of the known routes to the corresponding clusters can be achieved.

[0059] The cluster analysis and / or the provision of the plurality of clusters can be carried out not only in a preceding training process or learning process but also during vehicle control of the vehicle using the driver assistance device.

[0060] In a further preferred method, to determine the control variable, a (particularly predetermined) interaction scope of the driver assistance device for performing the vehicle function is changed and / or checked depending on the uncertainty variable. The interaction scope can be the interaction scope for determining the control variable. In particular, the interaction scope can indicate the degree or amount to which the driver assistance device or a trainable model may intervene in the driving process and / or can intervene in the vehicle function. Preferably, the interaction scope can indicate the extent to which the driver assistance device (and / or an underlying trainable and / or trained model) may or can explore the environment.

[0061] Preferably, depending on the uncertainty magnitude, it can be assessed whether the scope of interaction specified for the driver assistance system is appropriate to the uncertainty regarding the route to be evaluated or whether it should be restricted. This offers the advantage that, for example, in the case of routes to be evaluated for which a greater uncertainty is assumed regarding their assignment (to clusters formed based on known routes), only a cautious intervention can be performed. Conversely, in the case of routes to be evaluated for which a very small uncertainty is assumed regarding their assignment (to clusters formed based on known routes), a greater scope of interaction can be permitted to perform the vehicle function.

[0062] The scope of interaction can in particular relate to an action space of the driver assistance device and / or a learning and / or trained model underlying the driver assistance device (for example to an action space of an agent, in particular a reinforcement learning agent, see also described below), from which the driver assistance device and / or the model (for example the reinforcement learning agent) can select or selects an action to carry out the vehicle function and / or to determine the control variable.

[0063] An agent is, in particular, a system that operates in an environment (e.g., the environment of a vehicle). With regard to the agent, the term "environment" does not necessarily refer to the physical boundaries of the agent, but rather to all components outside the agent's decision-making process. For this purpose, a so-called policy ("strategy" or "course of action" of the agent based on a specific state) and a so-called value function are approximated. If the agent learns a deterministic policy, its policy provides an explicit action. If the agent acts according to a stochastic policy, this provides a distribution across all possible actions based on the current state. The action space describes the space of all possible actions. Using a predefined function, the so-called value function, the agent evaluates its current state or the choice of an action within it.In particular, this output indicates the expected reward for the agent based on the given state. It is also conceivable that the agent evaluates how good it is to choose a specific action in a given state and then follow its policy. It is also conceivable that the agent additionally uses a model of its environment to draw conclusions about the behavior of the environment. This advantageously makes it possible to specify the subsequent state or the expected reward for an action taken. With respect to the agent, the state space refers to the set of all states of the agent.

[0064] Building on the use of geometric clustering as an objective quality function for assessing the generalizability of a trained and / or learning model (e.g., a reinforcement learning agent), the clustering method is now advantageously used to obtain a confidence measure regarding the novelty of the current situation based on geometric characteristics (or geometric parameters) and variables describing a trajectory, namely the speed and acceleration profile. Based on this, the driver assistance system and / or the model and / or the agent (e.g., the reinforcement learning agent) are then modified in their interaction scope.

[0065] In a further preferred method, the uncertainty variable with respect to a future (potential) route is assessed as the route to be evaluated. In other words, a prediction of the uncertainty or indeterminacy with respect to the future (potential) route is made. By predicting the uncertainty or indeterminacy of various future driving situations, the execution of the vehicle function can be influenced, for example, even before this situation occurs. This allows for optimized control, since an improved signal profile can be achieved.

[0066] In another possible embodiment of the present invention, an adaptive model, i.e., a machine learning device, is used as the driver assistance device or as part thereof. Such an adaptive model can be implemented, for example, by means of or in the form of an artificial neural network.

[0067] If the specific familiarity of the route to be assessed is lower than a predefined threshold, a predefined safety function of the driver assistance system or a training mechanism is automatically activated to train the adaptive model. Such a safety function can, for example, automatically intensify or more intensively monitor a corresponding data-driven learning approach for routes or route sections with correspondingly lower familiarity if the distance is sufficiently large or the familiarity level is sufficiently low. In a training process, for example, corresponding routes or route sections can then be used more frequently as training data, or a respective result or route can be marked for manual review.

[0068] During operation of a vehicle equipped with the driver assistance system, for example, a navigation route or a trajectory along at least a section of a navigation route can be automatically determined for an upcoming journey of the vehicle. It can then be checked in the manner described whether and to what extent this navigation route or trajectory is known to the driver assistance system. If the level of knowledge is lower than the specified threshold, the safety function can be automatically activated, for example, to still enable safe operation of the vehicle along the navigation route.For this purpose, for example, operating parameters of the vehicle, such as its maximum speed, can be reduced or restricted; an escalation threshold for initiating predefined measures that gradually intervene in driving operations can be lowered; manual vehicle control can be requested at least in some places, and / or similar measures. This can also improve the safety of the driver assistance system's operation based on the specific awareness.

[0069] In a further possible embodiment of the present invention, a learning model, i.e. a machine learning device, is used as the driver assistance device or as part thereof, and this model is trained using reinforcement learning on the basis of driving routes specified as training data. In other words, a predetermined so-called reinforcement learning agent can be used here to train the driver assistance device, for example for automated vehicle guidance. The driving routes specified as training data can, for example, be the known driving routes or include them. A reward function can be specified for reinforcement learning, which, for example, rewards maximizing the performance of the driver assistance device. In particular, the learning model or the driver assistance device can be trained here using reinforcement learning.This is readily possible here, since the specific familiarity of the routes to be evaluated, as described, provides a robust opportunity or basis for retrospectively or temporarily verifying the generalization capability of the driver assistance system or the training method used. Thus, the design proposed here allows the driver assistance system to be provided particularly effectively and efficiently without compromising robustness or safety.

[0070] In a further preferred method, the driver assistance system uses an agent, in particular a reinforcement learning agent, to determine the control variable. The agent is suitable and intended for exploring the agent's environment, i.e., for exploring the environment through actions deviating from the best actions known to the agent. The exploration, in particular an exploration rate and / or a decision regarding further exploration, is carried out depending on the uncertainty. The agent preferably determines the control variable. This allows a greater reward to be obtained in the long term and, at the same time, the greatest possible driving safety to be achieved.

[0071] By predicting the uncertainty or indeterminacy of various future driving situations, the exploration rate can be influenced even before the situation occurs. This has particular advantages from a control engineering perspective, as it allows for improved signal progression. This allows exploration to be restricted early on in safety-critical driving maneuvers, preventing dangerous situations from arising due to additional exploration.

[0072] The agent generally prefers to explore its environment in order to be able to make a better assessment of the state space. Exploration depending on the degree of uncertainty can contribute to improvement through its informative added value. This can ensure that the agent makes its choice of action (or in particular its choice of the amount of exploration) depending on the uncertainty of the current driving situation. For example, if the route to be evaluated is assessed as uncertain, a comparatively smaller amount of exploration could initially be selected, whereas a comparatively larger amount of exploration could be selected if the route to be evaluated is assessed as less uncertain.

[0073] A further preferred method uses the uncertainty magnitude to assess when exploration is necessary. Many scientific papers on exploration strategies address the question of which action should be explored (Garcia, Francisco M., and Philip S. Thomas. "A meta-MDP approach to exploration for lifelong reinforcement learning." arXiv preprint arXiv:1902.00843 (2019)) but not when exploration should and should not occur. However, the latter is directly related to the area of ​​"live-long learning." A convergence of the exploration rate in the case of deterministic methods is also necessary for the convergence of the algorithm, so continuous exploration is not possible. However, in areas where the state space has not yet been sufficiently explored or in new driving situations, further exploration is certainly desirable.Thus, the proposed method (particularly by considering the uncertainty) provides a metric for determining when further exploration should and should not take place. This can advantageously address the convergence problem while simultaneously dealing with new driving situations, changing vehicle characteristics, or customer requests.

[0074] As a metric for determining an exploration amount or for making a statement regarding an exploration amount, for example, a distance to the center of the cluster (the respective data points of a section of the route to be evaluated) and / or the temporal course of driving dynamics parameters (such as speed and / or acceleration) and / or the certainty or uncertainty of the agent in the respective driving situation can be used.

[0075] The determination of the certainty or uncertainty of the agent (or the uncertainty quantity) can preferably be combined with Bayesian approaches.

[0076] Alternatively or additionally, the metric and / or the determined uncertainty can be used to make a decision about further exploration ("live-long learning"). For example, if uncertainty is high in certain areas and further exploration does not raise any safety concerns, a decision can be made to proceed with further exploration.

[0077] In a further preferred method, to determine the uncertainty magnitude of a specific section of the route to be evaluated, a correlation of at least part of the historical data with respect to the section and the associated cluster is determined. In other words, this makes it possible to assess, in particular, how similar the respective section is to the already known routes. Furthermore, a section-by-section assessment of the respective uncertainty and thus a locally precise prediction or assessment of the uncertainty along the route is possible.

[0078] In a further preferred method, a cluster centroid is determined for each cluster, and when determining the uncertainty value, at least the cluster centroid of the cluster assigned to the historical data to be evaluated is taken into account. This advantageously allows a metric to be formed in which the distance of the data points to be evaluated (of the historical data) to the respective cluster centroids is determined, thus determining a correlation with the already known data points (e.g., from the known routes) in a time-efficient manner. The distance to a cluster centroid advantageously represents a meaningful measure of the certainty or confidence of the assignment to the respective cluster.

[0079] In a further preferred method, the uncertainty quantity is determined based on a (particularly) relative distance metric based on an arrangement of at least part of the history data with respect to an arrangement of at least one cluster centroid of a cluster in a property space spanned by the predetermined geometric parameters. Preferably, a quantity characteristic of a distance between at least one data point of the history data with respect to the route section and a cluster centroid of the cluster assigned to the route section is used as the metric for determining the uncertainty quantity, wherein the distance is determined in a representation in a property space spanned by the predetermined geometric parameters. The distance metric can be a normalized distance metric.

[0080] During the training of an agent and / or a model, the training data used is used to determine the centroids of the clusters. For new routes or for the route to be evaluated, new data points are subsequently assigned to the previously determined centroids. The uncertainty of the data point can be determined based on the distance (also referred to as "spacing" above) of the respective data point to the assigned centroid, as well as the mean distance within this cluster.

[0081] Based on this uncertainty, the agent's control signal can be reduced proportionally to the uncertainty in critical driving situations (high lateral acceleration and control activity).

[0082] In a further preferred method, a driving situation of the vehicle is evaluated with regard to vehicle stability and / or driving safety, and the control variable is determined depending on this and the uncertainty variable. For example, the vehicle stability and / or driving safety is assessed for the occurrence of critical driving situations, such as high lateral acceleration and / or control activity. The control variable can be increased or decreased depending on the uncertainty variable. The control variable can be decreased in relation to the uncertainty variable (approximately proportional to it).

[0083] This advantageously results in minimal intervention in situations where no critical driving situations occur and where uncertainty is low.

[0084] Due to inaccuracies in system models and, predominantly, external, frequently and dynamically changing disturbances, as well as complex driving situations, adaptive control methods are preferred for implementing vehicle functions. This ensures consistently high control quality.

[0085] In a further preferred method, the control variable is characteristic of a manipulated variable in an adaptive control method for implementing the vehicle function. Thus, the uncertainty estimation or the determination of the uncertainty variable can advantageously be used as a safety aspect for possible interventions in the control behavior of a reinforcement learning agent in adaptive control strategies.

[0086] The reinforcement agent can take over the entire control process. However, it is also preferable for the reinforcement agent to be used as an additional actuator. This offers the advantage that, for example, a model-based control (or control component) can be combined with an additional control (or control component) implemented by the reinforcement learning agent. This allows for more robust vehicle control.

[0087] In a further preferred method, route data and / or history data are identified that cannot be clearly assigned to a cluster and / or whose correlation to an assigned cluster is only low. In particular, a further training process of the adaptive model, and in particular of the reinforcement learning agent, is carried out based on the identified route data and / or history data.

[0088] The classification of already known data points into so-called clusters is preferably used to detect so-called "white spots." "White spots" are subsets of new data clouds that either cannot be clearly assigned to a cluster or have only a very low correlation with a cluster. This information is then preferably used in a reinforcement learning framework for the directed retraining of an agent.

[0089] Thus, the exploration rate can either be adjusted based on the information, as described above, to achieve increased driving safety despite possible exploration, or the agent's positioning activity can be restricted. Furthermore, an assessment of the need for retraining can be conducted ("live-long learning").

[0090] In a further preferred method, the driver assistance system uses a learning model, and a further training process (of the learning model) is initiated or stimulated depending on the uncertainty. In particular, the uncertainty can be used to consider situations in which recorded data points cover new areas of the state space (of the agent) during the further training course or process, thereby increasing the exploration rate. This advantageously allows for a decoupling of the temporal component in the exploration rate.

[0091] Preferably, the further training process is carried out depending on the route data and / or historical data that have been identified as not clearly assignable to a cluster and / or exhibiting a low correlation with an assigned cluster. This allows for targeted and therefore more efficient follow-up training.

[0092] A well-known problem with reinforcement learning is the relationship between exploration and exploitation, in which the agent's current policy must either be evaluated or further exploration of the state space must take place. If exploration is continuous, the current policy can only be evaluated to a limited extent. If too little exploration is performed, possible locations in the state space remain unexplored, thus potentially better system states are not revealed.

[0093] Through directed retraining depending on the uncertainty of the data points regarding their assignment to a cluster, a targeted exploration of previously unknown states in the state space can be achieved.

[0094] Preferably, the learning and / or trained model and in particular the reinforcement learning agent is based on a deterministic learned strategy (policy).

[0095] Alternatively, stochastic policies can also be used for this use case, as they intrinsically incorporate uncertainty by modeling a probability distribution dependent on the state space. Stochastic policies are slower to converge in this use case and are also not suitable for using Deep Deterministic Policy Gradient, the preferred algorithm for training the agent. This algorithm requires a deterministic policy with an additional exploration functional.

[0096] The method proposed by the invention offers the advantage that by decoupling the uncertainty determination or the determination of the uncertainty variable from the actual policy, the learned strategy (in the case of stochastic policies), a prediction of the uncertainty or the uncertainty variable ("uncertainty") can be made (especially if all input variables can be determined). When planning variables are used, these are known with a certain temporal advance and can thus be used to a priori determine the uncertainty or the uncertainty variable. This enables a more continuous change of the control signal, which is advantageous for the underlying control system (for implementing the vehicle function).

[0097] Preferably, the driver assistance device is selected from a group comprising (at least) a longitudinal guidance system, for example ACC (abbreviation for automatic cruise control), a lateral guidance system, a lane centering system, an ESC system (electronic stability control system), ABS (anti-lock braking system), an evasive assistance system, an emergency braking assistance system, a lane departure warning system, a fatigue warning system and the like, as well as combinations thereof.

[0098] The driver assistance system can (in each case) be a system that is at least partially automated and ready to take over driving. It is also conceivable that it is a driver assistance system that is at least temporarily suitable (and intended) to completely take over (in particular independent) vehicle guidance (in particular within defined limits, for example, as a traffic jam assistant for taking over longitudinal guidance with respect to the vehicle in front as well as lateral guidance within the vehicle's own lane in a traffic jam). It is also conceivable that the driver assistance system is suitable and intended to take over the entire driving task, in particular longitudinal and lateral guidance, (over an extended period of time) – in particular independently of the current driving scenario.

[0099] The vehicle function can be a function of a vehicle component of the vehicle, for example, a perception function for at least one driver assistance system. Preferably, the vehicle component is selected from a group including a system for (in particular automated and / or partially automated) driving and / or controlling the vehicle, a driver assistance system, a navigation system, a warning system, a vehicle safety system (e.g., a braking system), or the like, as well as combinations thereof.

[0100] In a further possible embodiment of the present invention, a vehicle is guided at least partially autonomously or automatically along the respective route to be evaluated by means of the driver assistance device. The performance of the driver assistance device achieved in this process is then preferably combined with the familiarity determined for this route, i.e. for example, weighted, in order to determine or assess the generalisability or generalization capability of the driver assistance device. In other words, this determines how well the driver assistance device copes with unknown, i.e. new, routes or situations. The familiarity of the respective route to be evaluated for the driver assistance device can therefore serve or be used as a confidence measure for the driver assistance device.The confidence measure and the performance of the driver assistance system along the route to be evaluated can then be used to create a measure in a predetermined manner to assess the robustness or reliability of the driver assistance system with regard to automated vehicle guidance on unknown routes, i.e. to assess the generalizability or generalization capability of the driver assistance system.

[0101] The performance can be determined for the entire route to be evaluated, or for the route sections mentioned elsewhere that correspond to the clusters, or for corresponding or individual driving maneuvers of the vehicle guided by the driver assistance system along the respective route to be evaluated. The combination of familiarity and achieved performance proposed here allows for the fact that, on a route that is new but relatively similar to familiar routes, a higher performance of the driver assistance system can generally be expected than on new routes that are less familiar, i.e., differ more significantly from the routes already familiar to the driver assistance system.Such a higher performance on a relatively well-known, new route accordingly provides less reliable or meaningful information about the generalisability of the driver assistance system. Accordingly, the performance on routes with greater familiarity can be relatively underweighted for the assessment of generalisability. On the other hand, a possibly even relatively poor performance on a certain route can nevertheless indicate a relatively good generalisability of the driver assistance system if this route has very little familiarity for the driver assistance system. Accordingly, the performance on routes with less familiarity can be relatively overweighted for the assessment of generalisability. The embodiment of the present invention proposed here thus makes it possible to determine the generalisability orThe generalization capability of the driver assistance system can be determined.

[0102] In a possible development of the present invention, the performance of the driver assistance device is determined for several routes to be evaluated and combined to determine the generalization capability of the driver assistance device. The performance is weighted depending on the determined familiarity of the respective route. The performance of the driver assistance device for a route with lower familiarity, i.e., the performance of the driver assistance device when driving the vehicle on routes whose geometric properties are significantly different from the geometric properties of the known routes, is overweighted. In contrast, the performance for routes with a smaller distance, i.e., greater familiarity and thus greater objective geometric similarity to the known routes, is underweighted.By assessing the generalization capability based on the performance of the driver assistance system on several potentially unknown routes to be evaluated, and applying the weighting described above, an even more precise, robust, and reliable determination of the driver assistance system's generalization capability is possible. This may be due in particular to the larger data set, as multiple routes offer a greater probability that the driver assistance system will be confronted with truly unknown features, geometric properties, or route layouts.

[0103] The present invention is further directed to an electronic computing device for a driver assistance device and / or a, in particular processor-based, driver assistance device for a vehicle for carrying out an at least partially automatic vehicle function, in particular a driving function, of a vehicle depending on a route to be evaluated, which is suitable and determined and / or configured to carry out the following computer-implemented method steps: Providing a plurality of clusters of route data relating to at least one known route, wherein the clusters group the route data section by section according to predetermined geometric parameters; determining course data, in particular acquired by means of a sensor device, which indicate a course of the route to be evaluated and applying the clusters to the course data to be evaluated in order to divide the route to be evaluated into route sections corresponding to the clusters and thereby assign a cluster to each individual route section; determining at least one uncertainty variable which is characteristic of an uncertainty with regard to the assignment made between at least one of the route sections and the cluster assigned to this route section; determining a control variable as a function of the uncertainty variable and providing the control variable for carrying out the vehicle function.

[0104] It is therefore also proposed within the framework of the driver assistance device according to the invention that an uncertainty variable for analyzing the uncertainty of the driver assistance device when evaluating the (unknown or new) route to be evaluated is taken into account when determining the control variable.

[0105] An external server (in particular those listed above or below) is understood to mean, in particular, a server external to the vehicle, in particular a backend server. The external server is, for example, a backend of a vehicle manufacturer or a service provider. The functions of the backend or the external server can be performed on (external) server farms. The (external) server can be a distributed system. The external server and / or the backend can be cloud-based.

[0106] Preferably, the driver assistance device, in particular a processor-based one, comprises the electronic computing device. However, it is also conceivable that the driver assistance device is in communication (for data exchange) with the electronic computing device and that the electronic computing device is provided in the form of an external server (and / or a distributed system), for example, as a backend server.

[0107] The present invention is further directed to a vehicle, in particular a motor vehicle, comprising a driver assistance device for a vehicle as described above, according to one embodiment. The vehicle can, in particular, be a (motorized) road vehicle.

[0108] A vehicle can be a motor vehicle, in particular a driver-only vehicle, a semi-autonomous vehicle, an autonomous vehicle (e.g., a Level 3, 4, or 5 autonomy vehicle (SAE J3016 standard)), or a self-driving vehicle. Level 5 autonomy refers to fully automated vehicles. Preferably, the vehicle is a vehicle from the transportation sector. The vehicle can also be a driverless transport system. The vehicle can be controlled by a driver or drive autonomously. Furthermore, in addition to a road vehicle, the vehicle can also be an air taxi, an aircraft, or another means of transport or another type of vehicle, such as an aircraft, watercraft, or rail vehicle.

[0109] The present invention is further directed to a computer program or computer program product comprising program means, in particular a program code, which represents or encodes at least some and preferably all of the method steps of the method according to the invention and preferably one of the described preferred embodiments and is designed to be executed by a processor device.

[0110] The present invention is further directed to a data memory on which at least one embodiment of the computer program according to the invention or a preferred embodiment of the computer program is stored.

[0111] The present invention has been described with reference to a vehicle and a driver assistance system for a vehicle. However, the method can also be applied outside the automotive industry for safety functions in methods for the automatic or semi-automatic execution of functions, particularly when using reinforcement learning agents, and can be transferred thereto with respect to determining the exploration rate and / or an exploration amount and / or further exploration and / or directed retraining (each depending on an uncertainty variable determined analogously to the above embodiments on the basis of the training data and / or historical data). The applicant reserves the right to also claim a method and an assistance system directed thereto.

[0112] Further advantages and embodiments can be seen from the attached drawings: Fig. 1 shows a schematic overview to illustrate the application of geometric clustering to a route; Fig. 2 shows an exemplary illustration for deriving a statement regarding an exploration amount using the proposed method according to an embodiment; Fig. 3 shows a schematic overview for safe exploration and lifelong learning, which can be achieved using the clustering and uncertainty assessment proposed by the inventive method according to an embodiment; Fig. 4 shows a schematic overview to illustrate the application of the geometrically determined clusters to a new route; Fig. 5 shows a schematic overview to illustrate an assessment of the familiarity of a route based on geometric clustering; Fig. 6 shows an application of a trained reinforcement learning agent; Fig.Fig. 7 shows a representation of the course of uncertainty over the distance meter; Fig. 8 shows a representation of a control signal of the agent over the distance meter compared with a consideration of the uncertainty; Fig. 9 shows a representation of a cross-deviation over the distance meter compared with a consideration of the uncertainty; and Fig. 10 shows a representation of a control activity over the distance meter compared with a consideration of the uncertainty.

[0113] Identical or functionally equivalent elements are provided with the same reference numerals in the figures.

[0114] Machine learning techniques and data-driven methods enable efficient mapping based on previously seen scenarios or situations, with the option of interpolating between known scenarios. However, for many methods, extrapolation to new scenarios, also referred to as the generalisability of the method, poses a problem. In the area of ​​driving maneuvers, however, it is not always trivial to distinguish when an interpolation of already known driving maneuvers or route sections is involved and when an extrapolation is involved. A metric that assesses the familiarity, i.e. the novelty of a new route or individual route sections from the perspective of, for example, a driver assistance system for guiding the vehicle along the new route, can use the data from already known routes to better classify the results of the driver assistance system, i.e. its performance along the new route.

[0115] The driver assistance system can be trained, for example, using reinforcement learning, in which a predefined agent explores its environment and, based on the observed scenarios—i.e., known routes or driving maneuvers—learns an optimized action, for example, for driving a vehicle along a route. Such actions can be or include, for example, adjusting steering angles and speeds during automatic driving, i.e., actions for lateral and longitudinal guidance of the vehicle. Since this is typically data-inefficient, i.e., a large amount of appropriately prepared data is required to optimize the agent's actions, an initial evaluation or learning of the routes or scenarios, or of suitable actions for them, often takes place at the simulation level.The transfer of such a simulation to reality, as well as from known routes or driving maneuvers to unknown routes or driving maneuvers, must then be evaluated. A confidence assessment regarding familiarity is useful for evaluating actions, and ultimately the agent's performance, as it allows for a more accurate assessment of the agent's quality.

[0116] Previous approaches often focus on either evaluating the performance of the agent or a strategy (policy) learned by it for unknown scenarios, in this case unknown routes or route sections, or on examining the variance of a given environment and subsequently the performance of a pre-trained strategy. Thus, either scenarios or an underlying Markov decision problem are varied. The disadvantage of this approach is that only the result, i.e., the agent's performance, is considered, but no statement is made about how this performance relates to the respective scenario, how this performance should be evaluated, or what significance this performance has.If new, supposedly unknown scenarios exhibit only a relatively small deviation from the scenarios presented during agent training, this relativizes the agent's performance, thus reducing its generalizability. Furthermore, known solutions are often used in simulation frameworks and, in reality, require experimentally obtained data first. Furthermore, with previous approaches, it is often difficult to make a statement about generalizability, either because a distribution shift in the state space can lead to poor performance, as this changes the underlying Markov decision problem, or because of the coverage of the state space in general.

[0117] In the following, however, a method is described that can already carry out an a priori evaluation of a state space, for example given geometric data or properties of a new route, which can possibly be applied directly in an experimental determination of corresponding data.

[0118] This shows Fig. 1 a schematic overview representation in which, among other things, a training route 10 is schematically depicted. This training route 10 can be driven by a motor vehicle in a specific training direction 12. Along the training route 10, route data 14 of the training route 10 are recorded at a plurality of measuring points, which specify geometric properties as values ​​of predefined geometric parameters. The training route 10 can represent a plurality of other routes that, like the training route 10, can be used to train a driver assistance system or an adaptive model of a driver assistance system, for example, for automated vehicle guidance.

[0119] The route data 14 is schematically represented here by geometric values ​​16 recorded at the individual measurement points along the training route 10, of which only a few are explicitly marked here for the sake of clarity. The geometric values ​​16 are entered into a property space, also only schematically indicated here, spanned by the specified geometric parameters. A clustering method is then applied to the route data 14 or the geometric values ​​16, whereby clusters 18 are formed based on the geometry of the enclosed training route 10.

[0120] To further illustrate the geometric properties of the training route 10, a curve curvature K is plotted centrally against a route coordinate or route position X, which indicates a current position along the training route 10 in the training direction 12. It can be seen that there are some straight, i.e., curvature-free sections along the training route 10, which are connected by curves with different curve curvatures K. The corresponding curves are represented here by deflections from the X-axis. Deflections in the positive and negative K direction represent different curvature directions, i.e., left and right turns, respectively. These deflections, or the straight sections in between, correspond to clusters 18.

[0121] The clusters 18 can be determined by means of a cluster analysis, possibly iterative. For example, multiple clustering can be performed to minimize the mean distance of the individual geometric values ​​16 of each of the clusters 18 to the centroids, i.e., a geometric center of gravity, of the respective cluster 18.

[0122] The clusters 18 ultimately determined in this way can then be applied to the training route 10. This results in a corresponding segmented training route, which is divided into several training route sections 22 according to the clusters 18 or the geometric properties represented by them. Thus, areas of the training route 10 that exhibit certain uniform or similar geometric properties are each compared with the clusters 18 and assigned to one of the clusters 18 that represents the corresponding geometric properties. After training has been completed, the clusters 18 correspond to known route sections. These can, for example, be mastered by the respective driver assistance system with known performance.

[0123] Figur 1 illustrates a first step, clustering based on, for example, curve curvature. The centroids formed by the individual clusters are then used to further assign the clusters.

[0124] Fig. 4 shows a schematic overview of the application of the geometrically determined clusters to a new, potentially completely or partially unknown route. Examples of this are a first (here closed) evaluation route E1, which is identified by reference numeral 80, and a second (here closed) evaluation route E2, which is identified by reference numeral 84. Reference numeral 82 denotes a first direction of travel (here counterclockwise) in which the evaluation route E1 is traveled. Here, the evaluation route E2 results from the evaluation route E1 because the evaluation route E2 is traveled in the opposite direction 86 (here clockwise) compared to the evaluation route E1.

[0125] The two diagrams 88 and 104 each show the curve curvature k of the evaluation section E1 (graph 88) and E2 (graph 104) as a function of the distance (i.e., in particular, as a function of the distance traveled). Since the two evaluation sections E1 and E2 are traveled in opposite directions, the curve curvature in diagram 104 is mirrored to that in diagram 88 on the x-axis (approximately in meters). Both evaluation sections each have two straight, i.e., non-curved sections, which is why the curve curvature runs along the x-axis in two sections each (and the curvature is zero).

[0126] The geometric route data can in turn be displayed in a diagram 90 and by applying the clusters 92, 94, 96, 98 to the evaluation routes E1 and E2, the respective route can be segmented and the individual segments can thus each be assigned to one of the previously determined clusters.

[0127] Reference numeral 100 denotes a region in which data points are arranged at a smaller distance from the cluster center 102 of the cluster 98 previously formed (using the test track). The track sections represented by these data points lying within region 100 exhibit a comparatively higher similarity to sections of the test track with respect to those data points that lie only in the outer ring outside region 100 and within region 98.

[0128] Fig. 5 shows a schematic overview of an application to a new, potentially completely or partially unknown route.

[0129] This can now be done, for example, after the first step of clustering, which is Fig. 1 As illustrated, in a second step, the geometrically determined clusters can be applied to evaluation routes. This preferably involves segmenting the route and assigning the individual segments to predefined clusters.

[0130] Preferably, in a further step, preferably following the second step, the confidence measure is determined based on the agreement of the clusters (the higher the agreement, the more similar the elements of the evaluation route are compared to the training route).

[0131] By way of example, a first evaluation route 26 and a second evaluation route 32 are specified here (see also Fig. 4 ), which are to be automatically evaluated with regard to their familiarity relative to the at least one known training route 10. The first evaluation route 26 can be divided into first route sections 30 by using the clusters 18. Likewise, the second evaluation route 32 can be divided into second route sections 36 by using the clusters 18. In the present case, the first evaluation route 26 and the second evaluation route 32 differ, for example, in that a first direction of travel 28 provided for the first evaluation route 26 runs opposite to a second direction of travel 34 provided for the second evaluation route 32. In other words, the first evaluation route 26 and the second evaluation route 32 are or were traveled in opposite directions.

[0132] As described for the training route 10, geometric data of the evaluation routes 26, 32 can be recorded, which can then be entered into the property space spanned by the specified geometric parameters. These geometric properties of the evaluation routes 26, 32 are then each evaluated based on the determined clusters 18 in order to detect whether the respective evaluation route 26, 32 includes already known route sections, i.e., areas of already known geometry represented by the clusters 18. The geometric properties determined pointwise or sectionally along the evaluation route 26, 32 can thus each be assigned to one of the clusters 18.For this purpose, for example, a distance to the centroids, i.e., cluster centroids 24, of clusters 18 can be determined, whereby the assignment can then be made to the cluster 18 whose cluster centroid 24 has the smallest distance in the property space. Thus, if a new route is traveled, a statement about the familiarity or novelty of a particular route element can be made based on the distance of the geometric properties of its route elements to the cluster centroids 24.

[0133] The certainty or reliability or the probability of error in the assignment of certain geometric properties or measurement points or route sections to one of the clusters 18 can increase with increasing distance from the respective cluster center 24. This results in a certain confidence measure 38 that indicates the familiarity of the respective new route, in this case the evaluation routes 26, 32, in the context of or relative to the at least one known training route 10 or its geometric properties.

[0134] In the present example, the first route sections 30 of the first evaluation route 26 comprise two straight, i.e., curvature-free, route sections for which a match with one of the clusters 18 can be found, since the training route 10 also comprises corresponding straight training route sections 22. However, the first evaluation route 26 also comprises first route sections 30 that correspond to a left-hand bend with a relatively low curve curvature K, for which there is no exact equivalent in the training route 10. Accordingly, the first evaluation route 26 is to be classified as partially known, so that based on the correspondingly relatively low match with the clusters 18 or a relatively large distance to their cluster centroids 24, the confidence measure 38 can, for example, result in a degree of familiarity of 48% for the first evaluation route 26.

[0135] In contrast, the second evaluation route 32 comprises, in addition to the straight, curvature-free sections, second route sections 36 which, due to the opposite second direction of travel 34, represent a right-hand bend with a relatively small curve curvature K. For this, there is at least a partial correspondence in the training route 10, since the training route 10 in the training direction of travel 12 also includes such a right-hand bend, albeit with a slightly different curve curvature K. This means that a correspondingly smaller distance to each of the cluster centroids 24 can be determined for all of the second route sections 36 of the second evaluation route 32. However, since the second evaluation route 32 does not exactly correspond to the segmented training route 20 orwhose training route sections correspond to 22, no 100% agreement is determined here, but for example a level of awareness of 90% as the confidence measure 38 for the second evaluation route 32.

[0136] The statement on familiarity provided by the confidence measure 38 enables, for example, the use of corresponding information for safety functions, which can monitor a data-driven approach for route sections of lesser familiarity, as well as a particularly meaningful assessment of how, for example, a machine learning approach used to train the respective driver assistance system, in particular a reinforcement learning agent, will act or perform on unknown routes.

[0137] The higher the confidence measure 38 is, i.e. the more precisely, unambiguously or reliably the geometric properties or the corresponding data points in the property space of the respective route being evaluated or to be evaluated can be assigned to one of the clusters 18, the more likely it is that the route or a respective route section is known.

[0138] The confidence measure 38, i.e., the familiarity of the corresponding route sections or routes as a whole, can then be incorporated into an assessment of the generalizability or generalization capability of the respective driver assistance system or of a training method or training system used to train it, for example, the reinforcement learning agent. For this purpose, the performance of the driver assistance system is examined after completion of its training on various test routes, in this case, for example, using the evaluation routes 26, 32. In order to be able to make a statement about the adaptation to new environments, i.e., routes or route sections, their route geometry is objectively described, for example, in an unsupervised learning approach using the geometric clustering described above.This highlights differences between the route geometry in a training data set—here, for example, training route 10—and in the respective test data set—here, for example, evaluation routes 26–32. If the route geometries differ significantly from one another, this can be used in the form of a correspondingly low confidence measure 38 as a weighting for assessing the driver assistance system in unknown driving maneuvers or along unknown routes. However, if the differences in route geometries are relatively small, i.e., relatively familiar or a relatively large confidence measure 38, the assessment of generalizability can be weakened.To assess generalizability, the performance of the driver assistance system determined on a new route and the confidence measure 38 determined for this route can be taken into account, for example, through a predefined combination or weighting. This can be done in an intermediate step during training to evaluate the training progress of the driver assistance system, as well as after training has been completed, i.e., at inference time.

[0139] The methods described here enable a more objective assessment of the performance of a data-driven learning approach in new environments—i.e., when processing new data that was not used during training, for example—than is possible with conventional approaches. Such objectification can be automated by using geometric clustering as an evaluation metric or as a basis for determining familiarity as an evaluation metric. This enables particularly rapid testing or verification of the robustness of a corresponding algorithm through the particularly simple detection of previously unconsidered, i.e., unknown, driving maneuvers or route sections.Previous approaches, in which, for example, a subjective test route is compared with at least one training route 10 purely visually, can in practice only be of limited significance with regard to the generalization capability of the respective algorithms, since it is not determined how strongly or to what extent the routes objectively differ from one another. Thus, such approaches require a relatively high degree of expert knowledge and manual effort, which also entails corresponding costs. This can be eliminated by the method described here, which is also not subject to any restrictions regarding the possible geometric complexity or intricacy of the routes to be evaluated.

[0140] Overall, the described examples show how geometric clustering can be used as a measure of confidence for an evaluation of the generalizability of a learning model, for example a reinforcement learning agent for adaptive control strategies of automatic driving functions, in order to ultimately enable a particularly safe use of a corresponding driver assistance device, i.e. a particularly safe automatic driving operation of a motor vehicle.

[0141] Based on the Figuren 1 , 4 and 5 Geometric clustering was already introduced above as an objective quality functional for assessing the generalizability of a reinforcement learning agent for adaptive control strategies of automatic driving functions.

[0142] Building on this, the clustering method is used to obtain a confidence measure regarding the novelty of the current situation based on geometric parameters and parameters describing a trajectory, in particular the speed and acceleration profile. Based on this, the agent's interaction scope is then modified. The classification of already known data points into so-called clusters will then be used to detect white spots. These are subsets of new data clouds that either cannot be clearly assigned to a cluster or have only a very low correlation with a cluster. This information is then used in a reinforcement learning framework for the directed retraining of an agent.Thus, the exploration rate can either be adjusted based on the information, thereby achieving increased driving safety despite possible exploration, or the agent's positioning activity can be restricted. Furthermore, an assessment of the need for retraining can be carried out (so-called "live-long learning").

[0143] In current approaches, exploration rates for deterministic policies (English: policy, German: "strategy") are usually regulated over time, so that extensive exploration takes place at the beginning and decreases over time. In stochastic policies, the exploration rate is adjusted based on past experience and is thus directly part of the policy.

[0144] There are heuristic methods for exploring the search space, such as Tabu Search (see, for example, Abramson et al., Tabu Search Exploration for On-Policy Reinforcement Learning, In: 4. 10.1109 / IJCNN.2003.1224033, 2013), which evaluate possible solutions and restrict their selection over time. The Tabu List can serve as a selection criterion based on how often a certain action has been used, but in a high-dimensional state and action space, it provides only a reduced information gain.

[0145] In addition, Bayesian approaches can provide additional information regarding the certainty of an action in a state. Another possibility in the area of ​​exploration for reinforcement learning is meta-learning, in which an exploration function is learned in addition to the policy (Garcia, Francisco M., and Philip S. Thomas. "A meta-MDP approach to exploration for lifelong reinforcement learning." arXiv preprint arXiv:1902.00843 (2019)).

[0146] In the case of deterministic policies and the associated exploration, this is time-dependent and also takes place under fixed specifications and does not take into account the current situational information content.

[0147] However, even in this case, exploration quality cannot be predicted, since the agent's observation state is not known a priori, and thus, it cannot be determined in which situation the agent will explore to what extent. Since in some situations, regardless of the information content, exploration is rather undesirable or only desired to a very limited extent (driving at the limit without a safety margin), predicting the uncertainty of future driving situations in combination with the exploration rate offers significant added value.

[0148] The taboo list is not suitable for high-dimensional state or action spaces, as ambiguity between individual data points is usually less present in these cases. The aforementioned publication in the field of Bayesian methods (Garcia, Francisco M., and Philip S. Thomas. "A meta-MDP approach to exploration for lifelong reinforcement learning." arXiv preprint arXiv:1902.00843 (2019)) focuses on the appropriate selection of actions during exploration rather than a general decision as to whether to explore. It also focuses on exploration for different tasks (different MDPs - MDP is an acronym for "Markov Decision Process"), which can be usefully applied, for example, to various vehicle derivatives. However, this publication does not consider improving or adapting the exploration of a policy for a specific MDP (e.g., a vehicle).

[0149] So far, safety features have been added that monitor vehicle stability and trained safety drivers in combination with further manual monitoring of the algorithm during driving tests.

[0150] Clusters are formed based on parameters from trajectory planning (e.g., reference curvature, speed, and acceleration) as well as previously experienced driving scenarios. The assignment of new driving situations to new clusters can then take place online in the vehicle, as described above in the context of Figur 1 explained.

[0151] A statement regarding the amount of exploration can be made using a metric to be defined, for example, distance to the center of the cluster, possibly the temporal course of driving dynamics parameters or the determination of the agent in the respective driving situation (combination with Bayesian approaches is possible here). This is done using the Figuren 2a - 2f illustrated.

[0152] The Figuren 2a - 2b illustrate a first step in which clustering is carried out based on the route data of a training route 40. A cluster analysis divides the training route into a plurality of route sections, with the route points of each route section being assigned (exactly) to one cluster.

[0153] At the Fig. 2a For the training track 40 used, the cluster analysis performed here resulted in 10 different clusters, each characterized by different sampling types. In other words, each of the ten sampling types F1 - F10 indicates membership in (exactly) one specific cluster.

[0154] In Fig. 2a The assignment of the individual route sections is illustrated by the respective sampling type of the cluster to which this route section is assigned. It can be seen that several (non-adjacent) route sections (which have a similar curvature) are assigned to the same cluster.

[0155] For each route point of the training route, it can then be determined how well the respective route points are approximated by the formed clusters, for example by determining the distance of the respective data point (which is assigned to the route point) to the previously determined centroid (of the cluster in which the data point lies). Fig. 2b In the representation 42, a normalized distance (in the range between 0 and 1) to the (respective) cluster centroid is shown for each route point.

[0156] The Figuren 2c and 2eillustrate a second step in which an assignment of new data points of the new route to the previously determined clusters (and thus centroids) from Fig. 2a In this step, the parameters defined in the first step (illustrated by Fig. 2a ) formed clusters applied to the route data of the two previously unknown routes 44 (see Fig. 2c ) and 48 (see Fig. 2e ). For this purpose, the two routes 44 and 46 are divided into sections according to the clusters formed based on the training route 40. Each of the resulting sections is assigned to one of the previously formed clusters. This creates Fig. 2c and in Fig. 2e a clustered course.

[0157] The assignment to a specific cluster is here, as in Fig. 2a , illustrated by the different sampling of the respective route sections according to one of the ten different sampling types F1 - F10.

[0158] For the two unknown distances 44 and 48, a characteristic value for uncertainty can now be determined for each data point or distance point (or distance section), preferably by determining a distance, here normalized, to the (respective) cluster centroid. Such a normalized distance to the cluster centroid is shown along the respective distances 44 and 48 in the two illustrations 46 (see Fig. 2d ) and 50 ( Fig. 2f ) is illustrated.

[0159] From this quantity, which is characteristic of the respective uncertainty (of a route point), a statement can then be made regarding an exploration amount.

[0160] In addition, a decision about further exploration ("live-long learning") can be made using the metric, for example, if the uncertainty is high in certain areas and further exploration does not raise any safety concerns. This is Fig. 3 illustrated.

[0161] The upper left image 56 illustrates the clustering procedure once again. For data selection, two (here closed) paths or segments 52, 54 are clustered based on characteristics of a planned trajectory, such as curvature, speed, or acceleration. This results in the paths or segments 52, 54, which are divided into different clusters L1 - L6.

[0162] An uncertainty assessment is illustrated in the lower left Figure 58. A performance evaluation can be performed by calculating an uncertainty based on, for example, a distance metric (and the previous clustering).

[0163] This enables safe exploration, as illustrated in Figure 64, top right. For example, during exploration, a prediction could be made regarding a trajectory 68 that deviates from a known trajectory 66 with respect to performance (of the control system or agent), and based on this, an exploration strategy could be adapted depending on an uncertainty. For safe exploration, performance in an unknown state space is preferably determined or predicted. Preferably, an exploration strategy can be adapted with respect to the uncertainty.

[0164] By predicting the uncertainty of various future driving situations, the exploration rate can be influenced even before this situation occurs.

[0165] This is particularly advantageous from a control engineering perspective, as it allows for improved signal flow. This allows exploration to be restricted early on during safety-critical maneuvers, preventing dangerous situations from arising due to additional exploration.

[0166] In addition, lifelong learning is made possible.

[0167] Another added value is an approach for determining when exploration is necessary and when not. Many scientific papers on exploration strategies address the question of which action should be explored (Garcia, Francisco M., and Philip S. Thomas. "A meta-MDP approach to exploration for lifelong reinforcement learning." arXiv preprint arXiv:1902.00843 (2019)) but not when exploration should be carried out and when not. However, the latter is directly related to the area of ​​"live-long learning." A convergence of the exploration rate is also necessary for the convergence of the algorithm, so continuous exploration is not possible indefinitely. However, in areas where the state space has not yet been sufficiently explored or in new driving situations, further exploration is certainly desirable.A metric for determining when to continue exploration and when not to do so can address the problem of convergence while simultaneously dealing with new driving situations, changing vehicle characteristics, or customer requests.

[0168] In deterministic policies (procedures) known from the state of the art, exploration is purely a function of time [y(t)].

[0169] The method proposed here enables a decoupling of exploration in the deterministic case from the temporal component, so that exploration is no longer a function exclusively of time, but rather a function of time and uncertainty, e.g., y(t, Sigma). Typically, the exploration rate should decrease over time and can reach zero over a predefined period. However, this is undesirable in "live long learning," since neither a specific period should be specified nor should learning be stopped. However, continuous exploration should not take place across the board.

[0170] In the case of stochastic policies, in which the policy does not specify an explicit value but rather the parameters of a probability distribution, exploration decreases over time due to the reward and a corresponding decision as to whether exploration or exploitation is preferable. The reason for this is that the familiarity of the respective state significantly influences the probability distribution. However, if an exploration rate is adjusted with respect to uncertainty, as is possible, for example, with the method proposed here, exploration can be carried out in unknown state spaces while still meeting specified requirements in terms of safety. Such safe exploration advantageously avoids unstable driving situations, so that the proposed method can achieve a high level of driving safety while simultaneously maintaining exploration.

[0171] The Figuren 6 - 10 explain an exemplary implementation using measurement data from a driving test at the Ehra-Lessin test site.

[0172] The Fig. 6 The route 106 shown here shows a route to be evaluated, or the evaluation route, with the recorded distance meters. The two axes represent a spatial extension of the evaluation route in two mutually perpendicular directions of a plane, such as x- and y-position coordinates (which can be converted from a world coordinate system, such as GPS data).

[0173] The track meter 0 marks the start point of the lap. A Reinforcement Learning Agent was tested on track E2 from Figur 4 trained and applied on this route. Thus, many system states (curve combinations, speed profiles, etc.) are unknown to the agent. Uncertainty is preferably determined using a relative distance metric, which is derived from the clusters determined based on the training route.

[0174] Fig. 7 shows the uncertainty 108 ("Uncertainty") over the distance meter. The threshold value here was assumed to be 1 or higher. If the uncertainty is greater than or equal to the threshold, the additive control signal of the reinforcement learning agent is preferentially switched off.

[0175] To achieve a smooth transition, a look ahead and temporal reduction of the signal is used.

[0176] Fig. 8 shows the agent's control signal ML plotted against the distance meter. Reference numeral 114 denotes the zero line. Reference numeral 112 represents the control signal curve without the influence of uncertainty, and reference numeral 110 represents the control signal curve taking uncertainty into account. It can be seen that in the areas of high uncertainty (cf. Fig. 7 and reference numeral 108) the agent is deactivated.

[0177] Fig. 9 shows the lateral deviation dy over the distance meter. Signal 124 represents the course of the lateral deviation without an active agent, signal 122 with an agent permanently active, and the signal marked with reference number 123 represents the lateral deviation with an active agent based on the uncertainty. When calculating the averaged lateral deviation, the agent, taking the uncertainty into account when adjusting its control signal, leads to a minimal averaged lateral deviation when comparing the three variants.

[0178] Fig. 10 shows the control activity FB over the distance meter. Reference numeral 134 denotes the control activity without an active agent, reference numeral 132 denotes the control activity with a permanently active agent, and reference numeral 130 denotes the control activity with an active agent based on uncertainty.

[0179] When calculating the averaged control activity, the agent also takes the uncertainty into account to adapt its control signal and reduces the averaged control activity to a maximum in comparison of the three variants. Bezugszeichenliste

[0180] 1Vehicle 10Training route, known route 12Training direction 14Route data 16Geometry values ​​18Cluster 20Segmented training route 22Training route sections 24Cluster centroids 26First evaluation route 28First direction 30First route sections 32Second evaluation route 34Second direction 36Second route sections 38Confidence measure 40Training route 42Representation of the normalized distance to cluster centroid 44Unknown route 48Unknown route 46, 50Representation of uncertainty 52, 54Path 56Image illustrating clustering 58Image illustrating an uncertainty assessment 64Image illustrating a safe exploration 66,68Trajectory 70Image illustrating lifelong learning 80Evaluation route 82First direction of travel 84Evaluation route 86Second direction of travel 106Route to be evaluated 108Uncertainty 110Control signal taking uncertainty into account 112Control signal without uncertainty influence 114Zero line 122Cross-departure with permanently active agent 123Cross-departure with active agent based on uncertainty 124Cross-departure without active agent E1Evaluation route 1 E2Evaluation route 2 F1 - F10Segments KKurve curvature XRoute position,

Claims

1. A method for performing an at least partially automatic vehicle function, in particular a driving function, of a vehicle depending on a travel route (106) to be assessed using a driving assistance device, comprising the computer-implemented method steps of: - providing a plurality of clusters (18) from route data with respect to at least one known travel route (10, 84), wherein the clusters group the route data section-wise according to predefined geometric parameters; - providing course data, in particular recorded by means of a sensor apparatus, that indicate a course of the travel route (106) to be assessed, and applying the clusters (18) to the course data in order to divide the travel route (106) to be assessed into route sections (30, 36) corresponding to the clusters (18) and, as a result, to assign one cluster (18) in each case to the individual route sections (30, 36); - determining an uncertainty variable which is characteristic of an uncertainty with respect to the assignment made between at least one of the route sections (30, 36) and the cluster (18) assigned to said route section (30, 36), wherein the uncertainty variable is characteristic of the uncertainty of the chosen cluster assignment of exactly one route section; - determining a control variable as a function of the uncertainty variable and providing the control variable for performing the vehicle function, wherein for each route section (30, 36) of the travel route to be assessed, to which route section a cluster (18) is assigned, a control variable is determined depending on the respectively assigned cluster and provided for performing the vehicle function.

2. The method according to claim 1, characterized in that the predefined geometric parameters comprise a local curve curvature (K), a local curvature direction (K) in the travel direction (12, 28, 34), a local road width, a local distance of a vehicle trajectory along the travel route (10, 26, 32) to a road edge, a local distance of a vehicle trajectory along the travel route (10, 26, 32) to a lane edge, local spatial route coordinates (X) along the travel route (10, 26, 32), a local speed of a vehicle along the travel route (10, 26, 32) and / or a local acceleration of a vehicle along the travel route (10, 26, 32).

3. The method according to any of the preceding claims, characterized in that an extent of interaction of the driving assistance device for performing the vehicle function is changed and / or checked depending on the uncertainty variable in order to determine the control variable.

4. The method according to any of the preceding claims, characterized in that the uncertainty variable is assessed with respect to a future potential travel route as the travel route to be assessed.

5. The method according to any of the preceding claims, characterized in that the driving assistance device uses an agent, in particular a reinforcement learning agent, to determine the control variable, wherein the agent is suitable and intended for exploring an environment of the agent, wherein the exploration, in particular an exploration rate and / or a decision about further exploration, depends on the uncertainty variable.

6. The method according to the preceding claim, characterized in that it is determined based on the uncertainty variable when exploration is required.

7. The method according to any of the preceding claims, characterized in that a correlation of at least a portion of the course data with respect to the route section and the assigned cluster (18) is determined in order to determine the uncertainty variable of a particular route section of the travel route to be assessed.

8. The method according to any of the preceding claims, characterized in that a cluster center point (24) is determined for each cluster and at least the cluster center point of the cluster assigned to the course data to be assessed is taken into account during the determination of the uncertainty variable.

9. The method according to any of the preceding claims, characterized in that the uncertainty variable is determined based on a relative distance metric on the basis of an arrangement of at least a portion of the course data with respect to an arrangement of at least one cluster center point of a cluster in a property space spanned by the predefined geometric parameters.

10. The method according to any of the preceding claims, characterized in that a driving situation of the vehicle is assessed with respect to a vehicle stability and / or driving safety and the control variable is determined depending hereon and as a function of the uncertainty variable.

11. The method according to any of the preceding claims, characterized in that the control variable is characteristic of a manipulated variable in an adaptive control method for performing the vehicle function.

12. The method according to any of the preceding claims, characterized in that the driving assistance device uses a teachable model and a further training process is initiated depending on the uncertainty variable.

13. The method according to the preceding claim, characterized in that route data and / or course data are identified which cannot be uniquely assigned to a cluster (18) or the correlation of which with an assigned cluster (18) is merely low, and an additional training process of the teachable model is carried out based on the identified route data and / or course data.

14. A driving assistance device for a vehicle for performing an at least partially automatic vehicle function, in particular a driving function, of a vehicle depending on a travel route (106) to be assessed, which driving assistance device is suitable and intended for performing the following computer-implemented method steps: - providing a plurality of clusters (18) from route data with respect to at least one known travel route (10, 84), wherein the clusters group the route data section-wise according to predefined geometric parameters; - determining course data, in particular recorded by means of a sensor apparatus, that indicate a course of the travel route (106) to be assessed and applying the clusters (18) to the course data to be assessed in order to divide the travel route (106) to be assessed into route sections (30, 36) corresponding to the clusters (18) and, as a result, to assign one cluster (18) in each case to the individual route sections (30, 36); - determining an uncertainty variable which is characteristic of an uncertainty with respect to the assignment made between at least one of the route sections (30, 36) and the cluster (18) respectively assigned to said route section (30, 36), wherein the uncertainty variable is characteristic of the uncertainty of the chosen cluster assignment of exactly one route section; - determining a control variable as a function of the uncertainty variable and providing the control variable for performing the vehicle function, wherein for each route section of the travel route to be assessed, to which route section a cluster is assigned, a control variable is determined depending on the respectively assigned cluster and provided for performing the vehicle function.

15. A vehicle, in particular a motor vehicle, comprising a driving assistance device according to the preceding claim.