DE102021202083A1 - Computer-implemented method for training at least one algorithm for a control unit of a motor vehicle, computer program product, control unit and motor vehicle - Google Patents

Abstract

A method is described for training at least one algorithm for a control unit of an autonomous motor vehicle using the at least one algorithm, the at least one algorithm being trained in a simulation environment by a self-learning neural network by selecting a critical driving situation or an accident in the motor vehicle; Replication of the critical driving situation or the accident in the simulation environment; Calculation of a location area of the motor vehicle model at an earlier point in time based on the driving and environmental parameters using at least one physics model; Simulation of the critical driving situation or the accident; Determining deviations between a state of the critical driving situation or the accident and the simulation thereof, and applying an optimization algorithm to minimize the deviations between the critical driving situation or the accident and its simulation.

Description

A computer-implemented method for training at least one algorithm for a control unit of a motor vehicle, a computer program product, a control unit and a motor vehicle are described here.

Computer-implemented methods for training at least one algorithm for a control unit of a motor vehicle, computer program products, control units and motor vehicles of the type mentioned at the outset are known in the prior art.

Methods, computer program products, motor vehicles and systems of the type mentioned at the outset are known in the prior art. The first semi-automated vehicles (corresponds to SAE Level 2 according to SAE J3016) have reached series maturity in recent years. Automated (corresponds to SAE Level >=3 according to SAE J3016) or autonomously (corresponds to SAE Level 4/5 according to SAE J3016) driving motor vehicles have to react independently to unknown traffic situations with maximum safety based on a variety of specifications, e.g. destination and compliance with current traffic regulations. Since the reality of traffic is highly complex due to the unpredictability of the behavior of other road users, it is considered almost impossible to program corresponding control units in motor vehicles using conventional methods and on the basis of man-made rules.

Instead, it is known to develop algorithms using methods of machine learning or artificial intelligence. On the one hand, such algorithms can react more moderately to complex traffic situations than traditional algorithms. On the other hand, with the help of artificial intelligence, it is in principle possible to further develop the algorithms during the development process and in everyday life through constant learning. Alternatively, a status is frozen by the manufacturer after the end of the training phase in the development process.

However, the validation of correspondingly generated automated driving functions is very complex and often difficult to implement. The state of the art is therefore, among other things, scenario-based simulations with the help of a so-called digital twin. The digital twin, ie a simulation of the vehicle model to be trained, has the algorithm to be trained, which is trained in such a way that the digital twin in the simulation essentially imitates the driving behavior of the real motor vehicle. Relevant scenarios are used for training in order to examine the behavior of the automated vehicle model to be tested (“system under test” or “SuT”). However, it is difficult to select and describe the scenarios in such a way that the vehicle model really gets into a critical situation or to develop a possible high proportion of a wide variety of critical scenarios.

From the DE 10 2006 044 086 A1 a system for the virtual simulation of traffic situations is known, with a display unit for graphically displaying the virtual traffic situations and an input unit, and with at least one modeled reference vehicle and at least one further modeled road user, the reference vehicle having a modeled detection device for capturing the virtual surroundings of the reference vehicle , and the behavior of the reference vehicle during a simulated traffic situation is at least temporarily taking into account the information recorded by the detection device.

The task is therefore to develop methods for training at least one algorithm for a control unit of a motor vehicle, computer program products and motor vehicles of the type mentioned at the outset such that critical situations can be better simulated when training the algorithm.

The object is achieved by a method for training at least one algorithm for a control unit of a motor vehicle according to claim 1, a computer program product according to the independent claim 9, a control unit according to the independent claim 12, and a motor vehicle according to the independent claim 13. Further refinements and developments are the subject of the dependent claims.

1. a) Bereitstellen wenigstens eines Computerprogrammproduktmoduls für die autonome Fahrfunktion, wobei das wenigstens eine Computerprogrammproduktmodul das Kraftfahrzeugmodell, den wenigstens einen zu trainierenden Algorithmus, das selbstlernende neuronale Netz sowie die Simulationsumgebung enthält;
2. b) Auswahl einer kritischen Fahrsituation oder eines Unfalls des Kraftfahrzeugs, wobei zu der kritischen Fahrsituation oder dem Unfall eine Mehrzahl von Fahrparametern des Kraftfahrzeugs und Umgebungsparametern bestimmt sind;
3. c) Nachbildung der kritischen Fahrsituation oder des Unfalls aus Schritt b) in der Simulationsumgebung und Festlegung eines Zielzeitpunktes (t_n), wobei der Zielzeitpunkt (t_n) von den Fahrparametern des Kraftfahrzeugs und den Umgebungsparametern aus Schritt b) umfasst sind;
4. d) Berechnung eines Aufenthaltsbereichs des Kraftfahrzeugmodells zu einem Zeitpunkt (t_n-1), wobei der Zeitpunkt (t_n-1) vor dem Zielzeitpunkt (t_n) liegt, anhand der Fahrparameter und Umgebungsparameter unter Verwendung wenigstens eines Physikmodells;
5. e) Simulation der kritischen Fahrsituation oder des Unfalls vom Zeitpunkt (t_n-1) zum Zielzeitpunkt (t_n),
6. f) Bestimmen wenigstens einer Abweichung zwischen einem Zustand der kritischen Fahrsituation oder des Unfalls aus Schritt b) und der simulierten kritischen Fahrsituation oder des simulierten Unfalls aus Schritt e);
7. g) Anwenden eines Optimierungsalgorithmus zur Minimierung der wenigstens einen Abweichung zwischen der kritischen Fahrsituation oder des Unfalls aus Schritt b) und der simulierten kritischen Fahrsituation oder des simulierten Unfalls aus Schritt e).

A computer-implemented method for training at least one algorithm for a control unit of an autonomously or semi-autonomously driving motor vehicle to implement an autonomous driving function by intervening in assemblies of the motor vehicle on the basis of input data using the at least one algorithm is described below, with the at least one algorithm in a Simulation environment is trained by a self-learning neural network using a motor vehicle model of the motor vehicle, comprising the following steps:

1. a) Providing at least one computer program product module for the autonomous driving function, wherein the at least one computer program product module comprises the motor vehicle model, the at least one algorithm to be trained, the self-learning contains the neural network and the simulation environment;
2. b) Selection of a critical driving situation or an accident of the motor vehicle, a plurality of driving parameters of the motor vehicle and environmental parameters being determined for the critical driving situation or the accident;
3. c) simulating the critical driving situation or the accident from step b) in the simulation environment and defining a target time (t_n), the target time (t_n) being comprised of the driving parameters of the motor vehicle and the environmental parameters from step b);
4. d) calculation of a location area of the motor vehicle model at a point in time (t_n-1), the point in time (t_n-1) being before the target point in time (t_n), on the basis of the driving parameters and environmental parameters using at least one physical model;
5. e) simulation of the critical driving situation or the accident from time (t_n-1) to target time (t_n),
6. f) determining at least one deviation between a state of the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e);
7. g) Application of an optimization algorithm to minimize the at least one deviation between the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e).

An autonomous driving function takes over the control of a motor vehicle in whole or in part by detecting the surroundings of the motor vehicle, deriving a sensible behavior of the motor vehicle from this and controlling the motor vehicle by intervention in the steering, accelerator, brakes and/or other units, e.g. lights, turn signals, etc .Controls according to the planned behavior. There is a general desire, particularly in mixed autonomous/non-autonomous traffic, to imitate the behavior of natural drivers so that unpredictable behavior does not cause any safety risks.

Depending on the configuration, the driving function can access navigation data. A corresponding autonomous driving function can be a traffic jam assistant, for example, which automatically controls the motor vehicle in heavy traffic at slow speeds.

In general, current data relating to the environment can be considered as input data, for example data obtained by means of various sensors. Such sensors can include, for example, cameras, radar, lidar and/or ultrasonic sensors, but also other sensors such as position sensors, e.g. GPS, magnetic field detecting sensors and the like.

Further possible input data are color planning data, which are obtained from a navigation destination, for example, and possibly traffic data, which determine the traffic flow on the route. Further data can be communication data, which is obtained from car-to-car or car-to-infrastructure systems, e.g. on traffic light phases or similar.

The self-learning neural network can be based on various learning principles, in particular it can use methods of reinforcement learning. Reinforcement learning is a set of machine learning methods in which the neural network independently learns a strategy to maximize rewards received. The neural network is not shown which action is best in which situation, but receives a reward at certain times, which can also be negative. Using these rewards, it approximates a utility function that describes the value of a particular state or action.

One way of achieving driving behavior that is as natural as possible is to define a virtual twin of a real motor vehicle with a real driver in a simulation environment that is modeled on a real environment. The real driver and the virtual twin are given identical driving tasks, with the real journey serving as the reference. The task of the virtual twin is to emulate the driving behavior of the real driver. A deviation between the trajectory of the real driver and the trajectory of the virtual twin are suitable as reward criteria. The self-learning neural network is given the task of minimizing the deviation. In this way, the self-learning neural network imitates the behavior of real drivers.

With regard to critical situations or even accidents, however, this is not so easy, since it is ethically unacceptable to deliberately put real drivers in dangerous situations or accident situations. With the method proposed here, it is possible to replace the known reference journeys with simulated journeys generated using the method, the behavior of which then depends on the self-learning neural network are to be imitated. For this purpose, according to the method, existing data can be accessed that contain driving parameters and environmental parameters.

A point in time immediately before the occurrence of the critical situation or the accident can be selected as the starting time t_n, since there is often no longer any reliable data available after a critical situation or accident.

By using a physics model, which is part of the simulation environment, for example, a probable trajectory before the critical situation arises or before the accident can be reconstructed according to the method described above.

By using an optimization algorithm, it is possible to minimize deviations between the critical driving situation or the accident and the simulated critical driving situation or the simulated accident in order to obtain the most realistic possible image of the critical driving situation or the accident in the simulation environment.

Further process steps for training the at least one algorithm can then be carried out, as explained above.

In a first further development it can be provided that h) steps d) to g) are carried out for a time period between (t_n-1) and an earlier time period (t_n-2).

In this way, an iterative back-calculation approach is used, with which the simulated critical driving situation or the simulated accident can be reconstructed.

These process steps can be dependent on an increment size | t_n - t_n-1 | be repeated several times or multiple times until a sensible starting time before the critical driving situation or the accident has been reached, for example a few seconds before the critical driving situation or the accident.

In a further refinement, provision can be made for the at least one deviation to be evaluated, with step h) only being initiated if the at least one deviation is below a threshold value.

In this way it can be achieved that the simulated trajectory of the critical situation or the accident corresponds to the real model with sufficient precision.

Both deviations from the previous step and deviations over several iteration stages are possible as deviations, e.g. between the start time and the target time. Several deviations can be measured, e.g. distances, speeds or times.

In a further refinement, it can be provided that i) the driving parameters and/or the environmental parameters have variable driving parameters and/or variable environmental parameters, with at least one of the variable driving parameters and/or environmental parameters being varied, with the simulation being carried out again with the varied parameters becomes.

1. 1. Festgelegte Parameter - solche Parameter haben einen Wert, der für die kritische Fahrsituation oder den Unfall festgelegt ist. Das können z. B. alle Werte der Verkehrsinfrastruktur wie Straßenverlauf und -breite und dergleichen sein oder auch ein Reibwert der Straße, falls dieser bekannt ist.
2. 2. Variable Umgebungsparameter - Hier können Wertebereiche (kontinuierlich oder diskret) festgelegt werden, deren konkrete Werte jedoch vor dem eigentlichen Start der Simulation festgelegt werden. Dies kann z.B. durch eine Versuchsplanung geschehen. Beispiele wären ebenfalls der Reibwert der Straße, weitere Verkehrsteilnehmer wie Fußgänger, Autofahrer und dergleichen, aber auch Sonnenstand, Sensorzustand etc.
3. 3. Designparameter - Hier können Wertebereiche vorgegeben werden. Diese Wertebereiche für die werden von dem (numerischen) Optimierungsalgorithmus in der Simulationsphase variiert, um den Zielpunkt zu erreichen. Hierzu gehören beispielsweise Trajektorien verschiedener Verkehrsteilnehmer.

Three types of parameters can be distinguished:

1. 1. Specified parameters - such parameters have a value that is specified for the critical driving situation or accident. This can e.g. B. all values of the traffic infrastructure such as the course and width of the road and the like or a coefficient of friction of the road, if this is known.
2. 2. Variable environmental parameters - Value ranges (continuous or discrete) can be specified here, but their concrete values are specified before the actual start of the simulation. This can be done, for example, by designing an experiment. Examples would also be the friction coefficient of the road, other road users such as pedestrians, drivers and the like, but also the position of the sun, sensor status, etc.
3. 3. Design parameters - Value ranges can be specified here. These value ranges for are varied by the (numerical) optimization algorithm in the simulation phase in order to reach the target point. These include, for example, trajectories of different road users.

In a further refinement, it can be provided that the optimization algorithm according to step g) varies the variable driving parameters and/or the variable environmental parameters between two iterations of the optimization.

This can be used as a reinforcement learning method and increases the automation of the process.

In a further refinement, provision can be made for the at least one deviation to be evaluated, with step h) being carried out if the at least one deviation is less than the threshold value.

The precision of the method can thus be ensured.

In a further refinement, it can be provided that if the at least one deviation is below the threshold value, the at least one algorithm is stored.

The storage can be an intermediate storage. After saving, a further iteration step can be carried out or, after validation of the driving function, another situation can be selected for this case, on which the at least one pre-trained algorithm is trained further.

In a further refinement, it can be provided that the at least one algorithm is used and trained in further simulations of the same critical driving situation or the accident or other critical driving situations or accidents.

In this way, the at least one algorithm can be trained in many different critical situations or accident situations.

In a further refinement, it can be provided that if the at least one deviation from a start time (t_s) to the target time (t_n) is not below the threshold value, the driving function for the selected critical driving situation or the accident is validated.

According to a further embodiment, after a certain time it can happen that the at least one deviation becomes too large. One reason for this can be that the motor vehicle model can avoid the critical situation or the accident by driving differently than the human driver. This result can be used as a validation of the corresponding situation, since the neural network knows how to avoid the critical situation or the accident.

1. a) Bereitstellen wenigstens eines Computerprogrammproduktmoduls für die autonome Fahrfunktion, wobei das wenigstens eine Computerprogrammproduktmodul das Kraftfahrzeugmodell, den wenigstens einen zu trainierenden Algorithmus, das selbstlernende neuronale Netz sowie die Simulationsumgebung enthält;
2. b) Auswahl einer kritischen Fahrsituation oder eines Unfalls des Kraftfahrzeugs, wobei zu der kritischen Fahrsituation oder dem Unfall eine Mehrzahl von Fahrparametern des Kraftfahrzeugs und Umgebungsparametern bestimmt sind;
3. c) Nachbildung der kritischen Fahrsituation oder des Unfalls aus Schritt b) in der Simulationsumgebung und Festlegung eines Zielzeitpunktes (t_n), wobei der Zielzeitpunkt (t_n) von den Fahrparametern des Kraftfahrzeugs und den Umgebungsparametern aus Schritt b) umfasst sind;
4. d) Berechnung eines Aufenthaltsbereichs des Kraftfahrzeugmodells zu einem Zeitpunkt (t_n-1), wobei der Zeitpunkt (t_n-1) vor dem Zielzeitpunkt (t_n) liegt, anhand der Fahrparameter und Umgebungsparameter unter Verwendung wenigstens eines Physikmodells;
5. e) Simulation der kritischen Fahrsituation oder des Unfalls vom Zeitpunkt (t_n-1) zum Zielzeitpunkt (t_n),
6. f) Bestimmen wenigstens einer Abweichung zwischen einem Zustand der kritischen Fahrsituation oder des Unfalls aus Schritt b) und der simulierten kritischen Fahrsituation oder des simulierten Unfalls aus Schritt e);
7. g) Anwenden eines Optimierungsalgorithmus zur Minimierung der wenigstens einen Abweichungen zwischen der kritischen Fahrsituation oder des Unfalls aus Schritt b) und der simulierten kritischen Fahrsituation oder des simulierten Unfalls aus Schritt e).

A first independent subject relates to a device for training at least one algorithm for a control unit of an autonomously or semi-autonomously driving motor vehicle for implementing an autonomous driving function by intervening in aggregates of the motor vehicle on the basis of input data using the at least one algorithm described, with a simulation environment for training the at least one algorithm is provided by a self-learning neural network based on a motor vehicle model of the motor vehicle, the simulation environment being designed to carry out the following steps:

1. a) providing at least one computer program product module for the autonomous driving function, the at least one computer program product module containing the motor vehicle model, the at least one algorithm to be trained, the self-learning neural network and the simulation environment;
2. b) Selection of a critical driving situation or an accident of the motor vehicle, a plurality of driving parameters of the motor vehicle and environmental parameters being determined for the critical driving situation or the accident;
3. c) simulating the critical driving situation or the accident from step b) in the simulation environment and defining a target time (t_n), the target time (t_n) being comprised of the driving parameters of the motor vehicle and the environmental parameters from step b);
4. d) calculation of a location area of the motor vehicle model at a point in time (t_n-1), the point in time (t_n-1) being before the target point in time (t_n), on the basis of the driving parameters and environmental parameters using at least one physical model;
5. e) simulation of the critical driving situation or the accident from time (t_n-1) to target time (t_n),
6. f) determining at least one deviation between a state of the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e);
7. g) Application of an optimization algorithm to minimize the at least one deviation between the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e).

In a first further development, it can be provided that the simulation environment is set up to
h) performing steps d) to g) for a time period between (t_n-1) and an earlier time period (t_n-2).

In a further refinement, it can be provided that the simulation environment is set up to evaluate the at least one deviation, with step h) only being initiated if the at least one deviation is below a threshold value.

In a further refinement, it can be provided that i) the driving parameters and/or the environmental parameters have variable driving parameters and/or variable environmental parameters, with the simulation environment being set up to vary at least one of the variable driving parameters and/or environmental parameters and the simulation with to be carried out again with the varied parameters.

In a further refinement, it can be provided that the simulation environment is set up so that the optimization algorithm according to step g) varies the variable driving parameters and/or the variable environmental parameters between two iterations of the optimization.

In a further refinement, it can be provided that the simulation environment is set up to evaluate the at least one deviation, step h) being carried out if the at least one deviation is smaller than the threshold value.

In a further refinement, it can be provided that the simulation environment is set up to store the at least one algorithm if the at least one deviation is below the threshold value.

In a further refinement, it can be provided that the simulation environment is set up to use and train the at least one algorithm in further simulations of the same critical driving situation or the accident or other critical driving situations or accidents.

In a further refinement, it can be provided that the simulation environment is set up to validate the driving function for the selected critical driving situation or the accident if the at least one deviation from a starting time (t_s) to the target time (t_n) is not below the threshold value lies.

Another independent subject relates to a computer program product with a permanent, computer-readable storage medium on which instructions are embedded which, when executed by at least one processing unit, cause the at least one processing unit to be set up to carry out the method of the aforementioned type.

The method can be executed on one or more computing units so that certain method steps are executed on one computing unit and other method steps are executed on at least one other computing unit, with calculated data being able to be transmitted between the computing units if necessary.

The processing unit can be part of the control unit.

In a first further development, it can be provided that the commands have the computer program product module of the type described above.

Another independent subject relates to a control unit with a permanent, computer-readable storage medium, with a computer program product of the type described above being stored on the storage medium.

Another independent subject relates to a motor vehicle with a control unit of the type described above.

In a first further development it can be provided that the computing unit is part of the control unit.

In a further refinement, provision can be made for the computing unit to be networked with environmental sensors.

• 1 ein Kraftfahrzeug, das zum automatisierten bzw. autonomen Fahren eingerichtet ist;
• 2 ein Computerprogrammprodukt für das Kraftfahrzeug aus 1;
• 3 eine Simulationsumgebung mit dem Kraftfahrzeug aus 1, sowie
• 4 ein Ablaufdiagramm des Verfahrens.

Further features and details emerge from the following description, in which at least one exemplary embodiment is described in detail—if appropriate with reference to the drawing. Described and/or illustrated features form the subject matter alone or in any meaningful combination, possibly also independently of the claims, and in particular can also be the subject matter of one or more separate applications. Identical, similar and/or functionally identical parts are provided with the same reference symbols. Schematically show:

• 1 a motor vehicle that is set up for automated or autonomous driving;
• 2 a computer program product for the motor vehicle 1 ;
• 3 a simulation environment with the motor vehicle 1 , such as
• 4 a flowchart of the process.

1 shows a motor vehicle 2 that is set up for automated or autonomous driving.

The motor vehicle 2 has a control unit 4 with a computing unit 6 and a memory 8 on. The memory 8 is a persistent memory whose data is not lost when the memory 8 is powered off. In memory 8, a computer program product is stored, which in the following in connection with the 2 until 4 will be described in more detail.

On the one hand, the control unit 4 is connected to a series of environmental sensors that allow the current position of the motor vehicle 2 and the respective traffic situation to be detected. These include environmental sensors 10, 11 on the front of motor vehicle 2, environmental sensors 12, 13 on the rear of motor vehicle 2, a camera 14 and a GPS module 16. Environmental sensors 10 to 13 can include radar, lidar and/or ultrasonic sensors, for example .

Furthermore, sensors for detecting the state of the motor vehicle 2 are provided, including wheel speed sensors 16, acceleration sensors 18 and pedal sensors 20, which are connected to the control unit 4. With the help of this motor vehicle sensor system 16, 18, 20, the current state of the motor vehicle 2 can be reliably detected.

During the operation of the motor vehicle 2, the computing unit 6 has loaded the computer program product stored in the memory 8 and executes it. On the basis of an algorithm and the input signals, the arithmetic unit 6 decides on the control of the motor vehicle 2, which the arithmetic unit 6 would achieve by intervening in the steering 22, engine control 24 and brakes 26, each of which is connected to the control unit 4.

Data from the sensors 10 to 20 are continuously buffered in the memory 8 and discarded after a predetermined period of time so that these environmental data can be made available for further evaluation.

The algorithm was trained according to the procedure described below.

2 shows a computer program product 28 with a computer program product module 30.

The computer program product module 30 has a self-learning neural network 32 that trains an algorithm 34 . The self-learning neural network 32 learns using methods of reinforcement learning, i. H. the algorithm 34 tries by varying the neural network 32 to obtain rewards for improved behavior according to one or more metrics or benchmarks, ie for improvements of the algorithm 34. In other embodiments, known learning methods of supervised and unsupervised learning, as well as combinations of these learning methods, can also be used.

The neural network 32 may essentially be a matrix of values, typically called weights, that define a complex filter function that determines the behavior of the algorithm 34 depending on inputs, which are presently received via the environmental sensors 10-20, and control signals generated to control the motor vehicle 2.

The algorithm present here is what is known as a deep neural network, which has at least one hidden layer in addition to an input layer and an output layer.

The computer program product module 30 can be used both in the motor vehicle 2 and outside of the motor vehicle 2 . It is thus possible to train the computer program product module 30 both in a real environment and in a simulation environment in which a virtual twin of the motor vehicle 2 is trained.

In particular, training according to the teachings herein begins in a simulation environment, as this is safer than training in a real environment.

The computer program product module 30 is configured to establish a metric to be improved. Such a metric can be the achievement of a specific target state, that of the real motor vehicle 2 . If the metric has exceeded a certain threshold, e.g. a deviation smaller than a certain threshold value, the metric can be considered fulfilled and the algorithm can be frozen or saved in this regard. Then it can either be optimized with regard to another metric and further trained using another mission, or the algorithm can be tested in a real environment.

The computer program product module 30 has a driving function 31, e.g. an autonomous driving program, which is based on input data, for example environmental data from the environmental sensors 10 to 13, the camera 14 and the GPS module 15 and possibly environmental databases and driving data from the sensors 16 to 18 on the basis of a mission, for example to reach a certain destination, plans interventions in steering 22, engine control 24 and brakes 26.

Furthermore, a neural network 32 is provided, which is designed here as a deep neural network with at least one hidden level and is part of an algorithm 34 that implements parts of the driving function 31 .

The motor vehicle model 2 ′, which represents a virtual representation of the motor vehicle 2 , has a simulation environment 36 . The motor vehicle model 2' simulates the motor vehicle 2 with regard to its driving characteristics.

Furthermore, a physics model 38 is provided that simulates the dynamic behavior of the components of the simulation environment 36 on the basis of driving parameters pF, for example speed, cornering speed, acceleration, brake actuation, mass, etc., and environmental parameters pU, for example weather and road friction values. The parameters pF and pU are partly fixed and partly incremental or variable in stages

In addition, an optimization algorithm 39 is provided, which varies the driving parameters pF and/or environmental parameters pU in order to influence the behavior of the motor vehicle model 2'.

3 shows a representation of the simulation environment 36 with a motor vehicle model 2 ', which is a virtual twin of the motor vehicle 2 from 1 is.

The model for the simulation environment 36 is an accident description in an accident database, e.g. GIDAS, in which an accident between two motor vehicles took place.

• Technische Bestandsaufnahme zu Unfall:
  • • Kraftfahrzeug: Baujahr 2021, Bremse: ABS, Motorisierung: 103 KW (1997 ccm), Höchstgeschwindigkeit: 195 km/h
  • • Ausstattung: Antischlupfregelung, Elektronisches Stabilitätsprogramm, Bremsassistent, Spurverlasswarnung, Reifendruckkontrolle, Geschwindigkeitsregelanlage
  • • Zustand nach Unfall:
    • ◯ Frontairbags Fahrer- und Beifahrerseite (aktiviert), Knieairbag Fahrerseite (aktiviert), Seitenairbag sitzintegriert 1. Reihe, Seiten-Kopfairbag 1. und 2. Reihe
    • ◯ Gurtstraffer, Gurtkraftbegrenzer 1. Sitzreihe (Auslösung an der Unfallstelle nicht feststellbar)
    • ◯ keine offensichtlichen technischen Mängel am Kraftfahrzeug feststellbar
    • ◯ Beschädigungen im gesamten Frontbereich durch Auffahren auf das stehende Kraftfahrzeug
    • ◯ Beschädigungen im Innenraum im Armaturenbereich

The following information, for example, is stored in such a database:

• Technical inventory of the accident:
  • • Motor vehicle: year of construction 2021, brakes: ABS, engine: 103 KW (1997 ccm), top speed: 195 km/h
  • • Equipment: traction control, electronic stability program, brake assistant, lane departure warning, tire pressure monitor, cruise control system
  • • Condition after accident:
    • ◯ Driver's and front passenger's front airbags (activated), driver's knee airbag (activated), seat-integrated side airbags in 1st row, side head airbags in 1st and 2nd rows
    • ◯ Belt tensioners, belt force limiters in the 1st row of seats (triggering cannot be determined at the scene of the accident)
    • ◯ no obvious technical defects can be found on the vehicle
    • ◯ Damage in the entire front area due to driving into the stationary motor vehicle
    • ◯ Damage to the interior in the dashboard area

Furthermore, sketches of the accident environment, e.g. road course, road condition, weather, position of the sun, time, and driving dynamics data, e.g. speeds, trajectories, etc. are usually stored.

In the simulation environment 36, a road 40 is shown to simulate the previously described accident, on which a further motor vehicle 42 is located next to the motor vehicle 2 at the edge of the road 40. The time at which motor vehicle 2 collides with motor vehicle 40 is shown.

In order to train the neural network 32 of the algorithm 34, the method of reinforcement learning is used, as already described above, with the algorithm 34 being trained to imitate the behavior of the real motor vehicle 2 as precisely as possible. However, since there is no real driving data from the accident to be simulated that can be imitated directly, comparative driving data that the algorithm 34 is intended to imitate are instead generated using the data from the accident database.

For this purpose, the known data on a trajectory 44 of motor vehicle 2 are iteratively imitated in simulation environment 36 and optimized as far as possible using optimization algorithm 39, so that motor vehicle model 2' emulates the known behavior of motor vehicle 2 in the best possible way.

In order to accomplish this, the trajectory 44 is divided into a large number of short sections, each of which is a time increment [t_n−1; t_n] with a duration of Δt = | t_n - t_n-1 | are spaced. The journey of the motor vehicle model 2' is then simulated using the physics model 38, initially from a time period t_n-1 lying a time increment Δt before a target time t_n (time of the collision) to the target time t_n and varied by varying the changeable driving parameters pF until the simulated motor vehicle model 2' collides with motor vehicle 42 at the same point, except for a deviation 46 that is less than a predetermined threshold value S_46 with the same values.

Once this has been achieved, the intermediate status is saved and the process in the previous one the time increment [t_n-2; t_n-1] repeated. This takes place until a start time t_s has been reached.

Algorithm 34 is then used and trained to validate driving function 31 by being trained on the driving behavior of motor vehicle model 2′, which was developed using the method described above and imitates the driving behavior of motor vehicle 2 involved in the accident.

After a certain distance, the algorithm 34 may be able to avoid the accident. This can be understood as a validation of the algorithm 34.

4 shows a flow chart of the method.

After the start, an accident situation is selected from a database and entered into the simulation environment. The driving situation is simulated up to the target time t_n.

The simulation of driving the motor vehicle model then begins for a time increment duration Δt from a time t_n−1 to the target time t_n. The location of the motor vehicle model 2 ′ is compared with the data on the real motor vehicle 2 . If this is not within an acceptable deviation 46, the optimization algorithm 39 is applied and variable driving parameters pF are varied.

If an acceptable deviation 46 is reached, the method described above is repeated starting from the previous point in time t_n−2 until the method has arrived at the start time t_s.

The neural network 32 is then trained until the accident can be avoided.

Although the subject matter has been illustrated and explained in more detail by exemplary embodiments, the invention is not restricted by the disclosed examples and other variations can be derived therefrom by a person skilled in the art. It is therefore clear that a large number of possible variations exist. It is also understood that exemplary embodiments are only examples and should not be construed as limiting in any way the scope, applications, or configuration of the invention. Rather, the preceding description and the description of the figures enable the person skilled in the art to concretely implement the exemplary embodiments, whereby the person skilled in the art, knowing the disclosed inventive idea, can make a variety of changes, for example with regard to the function or the arrangement of individual elements mentioned in an exemplary embodiment, without having to Leaving the scope of protection defined by the claims and their legal equivalents, such as a further explanation in the description.

Reference List

2

motor vehicle

2'

motor vehicle model

4

control unit

6

unit of account

8th

Storage

10-13

environmental sensor

14

camera

15

GPS module

16

wheel speed sensor

18

accelerometer

20

pedal sensor

22

steering

24

engine control

26

brakes

28

computer program product

30

computer program product module

31

driving function

32

neural network

34

algorithm

36

simulation environment

38

physics model

39

optimization algorithm

40

Street

42

parked motor vehicle

44

trajectory

46

deviation

t_n

target time

t_n-1, t_n-2 ...

time

t_s

start time

Δt

time increment duration

pF

driving parameters

PU

environmental parameters

S_46

threshold

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• DE 102006044086 A1 [0006]

Claims (14)

Computer-implemented method for training at least one algorithm (34) for a control unit (4) of an autonomously or semi-autonomously driving motor vehicle (2) for implementing an autonomous driving function (31) by engaging in units (22, 24, 26) of the motor vehicle (2). on the basis of input data using the at least one algorithm (34), the at least one algorithm (34) being trained in a simulation environment (36) by a self-learning neural network (32) using a motor vehicle model (2') of the motor vehicle (2). , comprising the following steps: a) Providing at least one computer program product module (30) for the autonomous driving function (31), the at least one computer program product module (30) containing the motor vehicle model (2'), the at least one algorithm (34) to be trained, the self-learning neural network (32) and the simulation environment (36) includes; b) selection of a critical driving situation or an accident of the motor vehicle (2), a plurality of driving parameters (pF) of the motor vehicle (2) and environmental parameters (pU) being determined for the critical driving situation or the accident; c) Replication of the critical driving situation or the accident from step b) in the simulation environment (36) and determination of a target time (t_n), the target time (t_n) of the driving parameters (pF) of the motor vehicle (2) and the environmental parameters (pU) from step b); d) calculation of a location area () of the motor vehicle model (2') at a point in time (t_n-1), the point in time (t_n-1) being before the target point in time (t_n), based on the driving parameters (pF) and environmental parameters (pU). using at least one physics model (38); e) simulation of the critical driving situation or the accident from time (t_n-1) to target time (t_n), f) determining at least one deviation (46) between a state of the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e); g) Application of an optimization algorithm (39) to minimize the at least one deviation (46) between the critical driving situation or the accident from step b) and the simulated critical driving situation or the simulated accident from step e). Computer-implemented method claim 1 , where h) steps d) to g) are performed for a period between (t_n-1) and an earlier period (t_n-2). Computer-implemented method claim 1 or 2 , wherein the at least one deviation (46) is evaluated, step h) only being initiated if the at least one deviation (46) is below a threshold value (S_46). Computer-implemented method according to one of the preceding claims, wherein i) the driving parameters (pF) and/or the environmental parameters (pU) have variable driving parameters (pF) and/or variable environmental parameters (pU), wherein at least one of the variable driving parameters (pF) and/or or environmental parameters (pU) are varied, with the simulation being carried out again with the varied parameters (pF, pU). Computer-implemented method according to one of the preceding claims, wherein the optimization algorithm (39) according to step g) varies the variable driving parameters (pF) and/or the variable environmental parameters (pU) between two iterations of the optimization. Computer-implemented method according to one of the preceding claims, wherein the at least one deviation (46) is evaluated, wherein if the at least one deviation (46) is smaller than the threshold value (S_46), step h) is carried out. Computer-simulated method according to one of the preceding claims, wherein if the at least one deviation (46) is below the threshold value (S_46), the at least one algorithm (34) is stored. Computer-simulated procedure claim 7 , wherein the at least one algorithm (34) is used and trained in further simulations of the same critical driving situation or the accident or other critical driving situations or accidents. Computer-simulated method according to one of the preceding claims, wherein if the at least one deviation (46) from a start time (t_s) to the target time (t_n) is not below the threshold value (S_46), the driving function (31) for the selected critical driving situation or the accident is validated. Computer program product with a permanent, computer-readable storage medium (8) on which instructions are embedded which, when executed by at least one processing unit (4), cause the at least one processing unit (4) to be set up to carry out the method according to one of the preceding claims. computer program product claim 10 , wherein the instructions the computer program product module (30) according to one of Claims 1 until 9 exhibit. Control unit (4) with a permanent, computer-readable storage medium (8), on the storage medium (8) according to a computer program product claim 10 or 11 is saved. Motor vehicle with a control unit (4). claim 12 . motor vehicle after Claim 13 , wherein the control unit (4) is networked with environmental sensors (10, 11, 12, 13, 14, 15).

Images