CN114667545A - Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle - Google Patents
Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle Download PDFInfo
- Publication number
- CN114667545A CN114667545A CN202080076990.0A CN202080076990A CN114667545A CN 114667545 A CN114667545 A CN 114667545A CN 202080076990 A CN202080076990 A CN 202080076990A CN 114667545 A CN114667545 A CN 114667545A
- Authority
- CN
- China
- Prior art keywords
- motor vehicle
- algorithm
- simulation
- computer program
- program product
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V20/00—Scenes; Scene-specific elements
- G06V20/50—Context or environment of the image
- G06V20/56—Context or environment of the image exterior to a vehicle by using sensors mounted on the vehicle
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/214—Generating training patterns; Bootstrap methods, e.g. bagging or boosting
- G06F18/2155—Generating training patterns; Bootstrap methods, e.g. bagging or boosting characterised by the incorporation of unlabelled data, e.g. multiple instance learning [MIL], semi-supervised techniques using expectation-maximisation [EM] or naïve labelling
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/24—Classification techniques
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- General Physics & Mathematics (AREA)
- Evolutionary Computation (AREA)
- Multimedia (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Databases & Information Systems (AREA)
- Computing Systems (AREA)
- General Health & Medical Sciences (AREA)
- Medical Informatics (AREA)
- Software Systems (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Evolutionary Biology (AREA)
- General Engineering & Computer Science (AREA)
- Traffic Control Systems (AREA)
Abstract
A method for training at least one algorithm for a control unit of a motor vehicle is described, wherein the algorithm is trained by a self-learning neural network, comprising the following steps: a) providing a computer program product module for automated or autonomous driving functions, b) providing a simulation environment with simulation parameters, wherein the simulation environment contains map data of a real-existing region of use, the motor vehicle and at least one simulated further traffic participant, wherein the behavior of the motor vehicle and of the at least one further traffic participant thereof is determined by a rule set with behavior parameters, wherein the rule set contains behavior parameters which determine permissible limits, c) providing a task for the motor vehicle, d) modifying at least one behavior parameter of the motor vehicle such that the at least one behavior parameter is on the side of the permissible limits, e) performing a simulation of the task.
Description
Technical Field
A method for training at least one algorithm for a controller of a motor vehicle, a computer program product and a motor vehicle are described herein.
Background
Methods, computer program products and motor vehicles of the type mentioned at the outset are known from the prior art. In the last years, the first part of automatically driven vehicles (corresponding to SAE grade 2 according to SAE J3016) has reached the level of mass production (series reife). Motor vehicles that are driven automatically (corresponding to SAE level > 3 according to SAE J3016) or autonomously (corresponding to SAE level 4/5 according to SAE J3016) must react independently to unknown traffic conditions with maximum safety on the basis of various pre-specifications (such as compliance with destinations and common traffic regulations). Since traffic reality is highly complex due to the unpredictability of the behavior of other traffic participants, in particular human traffic participants, it is virtually impossible to program the respective controllers of motor vehicles in a conventional manner and on the basis of rules established by humans.
To solve complex problems with computers, it is also known to develop algorithms with the aid of machine learning methods or artificial intelligence methods or with the aid of self-learning neural networks. On the one hand, such algorithms are able to react more gracefully to complex traffic conditions than traditional algorithms. On the other hand, with the aid of artificial intelligence, it is possible in principle to further develop and continuously improve algorithms during the development process and in daily life by continuous learning. Alternatively, the state of the algorithm may be frozen after the training phase in the development process is terminated and verified by the manufacturer.
DE 102017007136 a1 discloses a method for training a self-learning algorithm for an automatically drivable vehicle by generating learning conditions using a predefined automation module, wherein the learning conditions are generated in the following manner: -performing a traffic simulation in which a virtual host vehicle is placed in a virtual scene using an automation module of the real vehicle, the scene comprises a driving road structure with a preset driving route, and also comprises other automatically generated virtual moving objects with object characteristics and behavior models which can be preset individually, wherein the objects interact with each other independently and adaptively during the progress of the simulation on the basis of corresponding object properties and behavior models, a vehicle dynamics simulation is carried out on the basis of the automation model and on the basis of virtual sensor signals of the moving object which are assigned to a virtual sensor device of the vehicle, the sensor device corresponds to a sensor device of a vehicle that is present in reality, in which a reaction of the vehicle occurs, and the relevant learning situation is detected according to a selection criterion that is determined on the basis of a specifiable measurement.
The disadvantage is that the simulation is performed using traffic participants who comply with the rules. In contrast, it often happens in reality that the traffic participants do not comply with rules, such as driving too fast, seemingly passing lane markings without fail, inattention, right-side overtaking, etc. Thus, human driving behavior is poorly prepared using algorithms trained by other traffic participants who act to comply with the rules alone. This leads to unnatural driving behavior of the motor vehicle equipped with a correspondingly trained algorithm, since the motor vehicle reacts less flexibly.
Furthermore, there are situations in which the host vehicle itself does not comply with the regulations to a hundred percent, for example, the host vehicle passes through a solid line in order to avoid an obstacle, as long as this can be achieved without danger, for example, when there is no oncoming traffic, the traffic flow can be improved and, if necessary, even the risk of accidents can be reduced. Applying braking may result in a rear-end collision by an unprepared human driver traveling behind due to a sudden interruption in traffic flow.
Disclosure of Invention
The object is therefore to develop a method, a computer program product and a motor vehicle of the type mentioned at the outset in such a way that the trained algorithm can be better adapted to the real traffic situation.
This object is achieved by a method for training at least one algorithm for a controller of a motor vehicle according to claim 1, a computer program product according to the parallel claim 12 and a motor vehicle according to the parallel claim 13. Extended configurations and embodiments are the subject matter of the dependent claims.
The following describes a method for training at least one algorithm of a control unit for a motor vehicle, wherein the control unit is provided for carrying out an automated or autonomous driving function by intervening in a group of the motor vehicle based on input data using the at least one algorithm, wherein the algorithm is trained by means of a self-learning neural network, comprising the following steps:
a) providing computer program product modules for the automated or autonomous driving function, wherein the computer program product modules contain the algorithm to be trained and the self-learning neural network,
b) providing a simulation environment with simulation parameters, wherein the simulation environment contains map data of a real-existing use area, the motor vehicle and at least one simulated further traffic participant, wherein the behavior of the motor vehicle and of the at least one further traffic participant is determined by a rule set, wherein the rule set contains behavior parameters which determine permissible limits,
c) a task for the motor vehicle is provided,
d) modifying at least one behavior parameter of the motor vehicle such that the at least one behavior parameter is on the side of the allowed limit,
e) a simulation of the task is performed.
The corresponding behavior parameters can be, for example, the permitted speed, the distance to be observed, a threshold value for which an exceedance is prohibited (for example a time period during which a red-changing traffic light is permitted to pass; a risk parameter in which a solid line is permitted to pass; and/or when a continued travel is still permitted without preemption), the permitted variance of the position of the motor vehicle in the lane, the permitted overtaking side (only the left side or both sides in the case of right-hand traffic), etc., more.
The corresponding task may be, for example, to arrive at a specific destination from a starting point in the shortest possible time or in an energy-consumption-optimized manner.
It has been established that a correspondingly trained algorithm has a driving behavior which differs from conventionally trained algorithms even within more strict parameter limits, as will be subsequently used for use in real motor vehicles. The driving behavior of such an algorithm is more natural, i.e. more in line with human driving behavior, which is more natural for the passengers on the one hand and for other traffic participants on the other hand. One example of this is the need to pass over a solid line past a transport vehicle parked in the second row. An algorithm that absolutely complies with the rules will bring the vehicle to a stop and wait until the delivery vehicle continues to travel. An algorithm trained according to the presently described method, which permits rules applicable in narrow limits to be overridden, for example when this can be carried out without danger because of the absence of oncoming traffic, driving is continued in the situation.
It is thus possible to achieve that a motor vehicle equipped with a corresponding algorithm can react more flexibly to traffic conditions than an algorithm trained with conventional methods.
In a first embodiment, it can be provided that the neural network is learned by means of a reinforcement learning method (also referred to as RFL algorithm, RFL standing for "reinforcement learning"), wherein at least one of the time for completing a task and/or the number of accidents in which the motor vehicle is involved during the task is used as a reward measure (belohn @ measurement), wherein the simulation is repeated until a minimum measure is reached.
In particular, it can be provided that no accident is required for successful completion of the task. The extended metric may be an indirect trigger of an accident that is not a participant in other traffic, for example due to sudden, unexpected hard braking.
By using reinforcement learning methods, neural networks learn better and better strategies for carrying out a predetermined task in the case of simulations which are carried out one after the other.
In a further embodiment, it can be provided that the at least one anomalous traffic participant is a motor vehicle, a motorcycle driver or a pedestrian.
Partially holding humans out of compliance with regulations causes vehicle motion. Therefore, simulations using such abnormally behaving traffic participants are particularly realistic.
In a further development, it can be provided that the computer program product module has an algorithm which has been pre-trained by means of the traffic participants following the rules.
In this way the already learned behavior pattern is refined and training becomes more efficient and faster.
In a further extended configuration, provision can be made for a predefined percentage to be higher or lower than at least one of the behavior parameters.
This can be significant in particular in the case of behavior parameters which can be expressed numerically, such as speed, distance, deviation from a predefined driving route, etc.
In a further embodiment, it can be provided that the simulation is repeated a plurality of times, wherein at least one simulation parameter is changed in each case.
Such simulation parameters may be, for example, behavior parameters. Over-specification of the algorithm to a specific situation can be avoided by varying the respective parameters: ()。
In another extended configuration, provision may be made for the simulated environment to be altered.
This also prevents the algorithm from being over-trained to existing simulation environments. The change may occur, for example, by making a modification within the same traffic zone (e.g., by changing road widths, preemption rules, traffic light switches, road blocks, etc.) or by changing the traffic zone as a whole.
In a further extended configuration, provision can be made for at least one of the behavior parameters to be varied.
Such behavior parameters cover common driving behaviors of different driver types, e.g. drivers who tend to drive too fast, drivers with less driving accuracy, etc.
In a further embodiment, it can be provided that the number, positioning and/or assignment of the other road users is varied.
This presents new conditions, with the aid of which the algorithm can be further trained.
In a further extended configuration it may be provided that the algorithm is further trained by a self-learning neural network, the method comprising the steps of:
a) providing a computer program product module for the automated or autonomous driving function, wherein the computer program product module contains an algorithm to be trained and a self-learning neural network,
b) providing a simulated environment with simulation parameters, wherein the simulated environment contains map data of a real-existing region of use, the motor vehicle and at least one simulated further traffic participant, wherein the further simulated traffic participant is simulated by means of an algorithm, wherein the algorithm is trained according to the method described above,
c) a task is provided for the motor vehicle,
d) a simulation of the task is performed.
In this way, it is possible to configure not only the motor vehicle but also other motor vehicles in a non-behavioural manner, thereby increasing the robustness of the algorithm.
The algorithm can be applied to other traffic participants, such as pedestrians or cyclists, wherein in this case, instead of using computer program product modules for automated or autonomous driving functions, computer program product modules for motor behavior simulation are used. These agents can thus be configured more realistically and can subsequently be used in future training tasks of the type described above, thereby improving the simulation quality.
In a further embodiment, it can be provided that the computer program product module is integrated into a control unit of the motor vehicle, and wherein the algorithm is tested and/or trained in a real-life region of use.
In this way, the influence of a real motor vehicle, which may not be completely simulated, can be taken into account. Thus, a vehicle that is actually traveling may react differently than the simulated situation.
A first independent subject matter relates to a device for training at least one algorithm of a controller for a motor vehicle, wherein the controller is provided for carrying out an automated or autonomous driving function by intervening in a unit of the motor vehicle based on input data using the at least one algorithm, wherein the algorithm is trained by means of a self-learning neural network, wherein:
a) means for providing a computer program product module for the automated or autonomous driving function, wherein the computer program product module contains an algorithm to be trained and the self-learning neural network,
b) means for providing a simulated environment with simulation parameters, wherein the simulated environment contains map data of a real-existing region of use, the motor vehicle and at least one simulated further traffic participant, wherein the behavior of the motor vehicle and of the at least one further traffic participant is determined by a rule set, wherein the rule set contains behavior parameters which determine permissible limits,
c) means for providing a mission for the motor vehicle,
d) means for modifying at least one behavior parameter of the motor vehicle such that the at least one behavior parameter is on the side of the allowed limit,
e) means for performing a simulation of the task.
In a first embodiment, it can be provided that the neural network has a device for learning by means of a reinforcement learning method, wherein at least one of the time for completing a task and/or the number of accidents in which the motor vehicle is involved during the task is used as a reward measure, wherein the simulation is repeated until a minimum measure is reached.
In a further embodiment, it can be provided that the at least one anomalous traffic participant is a motor vehicle, a motorcycle or a pedestrian.
In a further embodiment, it can be provided that the computer program product module has an algorithm that has been pre-trained by the traffic participants in compliance with the rules.
In a further embodiment, provision can be made for means to be provided for the at least one behavior parameter to be higher or lower by a predetermined percentage.
In a further embodiment, it can be provided that means are provided for repeating the simulation a plurality of times, wherein at least one simulation parameter is changed in each case in the simulation.
In a further embodiment, provision can be made for means for changing the simulated environment to be provided.
In a further embodiment, provision can be made for means to be provided for varying at least one of the behavior parameters.
In a further embodiment, it can be provided that means are provided for varying the number, positioning and/or tasks of the other road users.
In a further embodiment, means for training the algorithm by means of a self-learning neural network can be provided, wherein:
a) means for providing a computer program product module for the automated or autonomous driving function, wherein the computer program product module contains an algorithm to be trained and the self-learning neural network,
b) means for providing a simulated environment with simulated parameters, wherein the simulated environment contains map data of a real-existing region of use, the motor vehicle and at least one simulated further traffic participant, wherein the further simulated traffic participant is simulated by means of an algorithm, wherein the algorithm is trained according to the method described above,
c) means for providing a mission for the motor vehicle,
d) means for performing a simulation of the task.
In a further embodiment, it can be provided that the computer program product module is integrated into a control unit of the motor vehicle, and that means are provided for testing and/or training the algorithm in the actually present region of use.
A further independent subject matter relates to a computer program product having a computer-readable storage medium on which instructions are embedded, which instructions, when executed by at least one computing unit, cause the at least one computing unit to be arranged for carrying out a method of the above-mentioned type.
The method can be implemented on one computing unit or distributed over a plurality of computing units, so that certain method steps are carried out on one computing unit and other implementation steps are carried out on at least one further computing unit, wherein the computed data (if required) can be transferred between the computing units.
Another independent subject matter relates to a motor vehicle having a computer program product of the above-mentioned type.
Drawings
Further features and details emerge from the following description, in which (if appropriate with reference to the drawings) at least one embodiment is described in detail. The features described and/or illustrated graphically form the subject matter individually or in any meaningful combination, if appropriate also independently of the claims, and can in particular additionally also be the subject matter of one or more separate applications. Identical, similar and/or functionally identical components are provided with the same reference numerals. Here, schematically shown:
FIG. 1: a motor vehicle provided for automated or autonomous driving;
FIG. 2: a computer program product for the motor vehicle of fig. 1;
FIG. 3: the simulated environment of the motor vehicle in fig. 1, and
FIG. 4: a flow chart of the method.
Detailed Description
Fig. 1 shows a motor vehicle 2, which is provided for automated or autonomous driving.
The motor vehicle 2 has a controller 4 with a computing unit 6 and a memory 8. In the memory 8 a computer program product is stored, which is described in more depth below in connection with fig. 2 to 4.
The control unit 4 is connected on the one hand to a series of environmental sensors which allow the current position of the motor vehicle 2 and the corresponding traffic situation to be detected. The environmental sensor includes: environmental sensors 10, 11 at the front of the motor vehicle 2, environmental sensors 12, 13 at the rear of the motor vehicle 2, a camera 14 and a GPS module 16. The environmental sensors 10 to 13 can comprise, for example, radar sensors, lidar sensors and/or ultrasonic sensors.
Furthermore, sensors for detecting the state of the motor vehicle 2 are provided, in particular a wheel speed sensor 16, an acceleration sensor 18 and a pedal sensor 20, which are connected to the controller 4. The present state of the motor vehicle 2 can be reliably detected by means of these motor vehicle sensor systems.
During operation of the motor vehicle 2, the computing unit 6 downloads the computer program product stored in the memory 8 and executes it. Based on the algorithm and the input signals, the computing unit 6 decides on the control of the motor vehicle 2, which control is to be effected by the computing unit 6 by intervening the steering 22, the motor control 24 and the brake 26, which are each connected to the controller 4.
The data of the sensors 10 to 20 are continuously buffered in the memory 8 and removed after a predetermined time duration, and these environmental data can thus be provided for further evaluation.
The algorithm has been trained according to the method described below.
Fig. 2 shows a computer program product 28 with computer program product modules 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 according to a reinforcement learning method, i.e. the neural network 32 tries to obtain a reward by the algorithm 34 for improving the behavior corresponding to one or more metrics or scales (Ma β stab), i.e. for improving the algorithm 34. Alternatively, known learning methods of supervised learning and unsupervised learning, and combinations of these learning methods may also be used.
The algorithm 34 can essentially consist of a complex filter with a matrix of values, which are generally referred to as weights by those skilled in the art, which define a filter function which determines the behavior of the algorithm 34 as a function of input variables which are recorded in the present case by the ambient sensors 10 to 20 and which generates control signals for controlling the motor vehicle 2.
The computer program product module 30 can be used not only in the motor vehicle 2 but also outside the motor vehicle 2. It is thus possible to train the computer program product module 30 not only in a real environment but also in a simulated environment. According to the teachings described herein, the training is started especially in a simulated environment, as this is safer than training in a real environment.
The computer program product module 30 is arranged for formulating a metric that should be improved. Such a measure may be, for example, the time until a predetermined task is performed (e.g., the destination is reached). If the metric has risen above a certain threshold, e.g. the time is less than a limit time, it can be considered that the metric has been met and the algorithm associated therewith is frozen. The algorithm can then either be optimized and further trained in terms of additional metrics or can be tested in a real environment.
Fig. 3 shows a simulated environment 36 of the motor vehicle 2 in fig. 1.
A road intersection 38 is provided in the simulated environment 36, at which a road 40 intersects a road 42. The road intersection 38 is based on map data that is actually present, so that the behavior of the algorithm 34 at the road intersection 38 is specifically simulated.
The motor vehicle 44 is parked at the road edge of the road 40 in such a way that it is not possible to drive past without passing the solid line 46. At the same time, according to the simulation, the motorcycle driver 48 wishes to turn from the road 42 into the road 40. Furthermore, the pedestrian 50 moves at high speed without noticing traffic in the direction of movement 52 towards the road 40, which it is obviously intended to traverse.
In the situation involved, there are multiple complex decisions to be made for the algorithm 34. The first decision to be made is whether or not it is permitted to pass the solid line 46. Since it is not possible to pass by a parked vehicle 44 without exceeding the solid line 46, a yes decision must be made to the algorithm 34, however, as to which driving parameters the question is based on. The algorithm 34 must make predictions for this purpose regarding: how a motorcycle driver 48, who may approach the motor vehicle 2 relatively closely in his normal trajectory, will act. However, in everyday life it is often the case that the respective motorcycle driver can avoid or drive to the right on his/her lane without problems due to the small motorcycle width and the low speed in the intersection region.
In addition, the speed of the motor vehicle 2 needs to be taken into account. When the vehicle 2 is slightly accelerated in order to pass the parked vehicle 44, the probability of the vehicle 2 obstructing the planned trajectory of the motorcycle driver 48 is reduced. However, this may lead to the motor vehicle 2 intersecting the trajectory of the pedestrian 50, which is to traverse the road 40 in the process, and which may be inattentive, as a result of which an accident may occur.
In successive iterations, the algorithm 34 may first attempt to pass the vehicle 44 without stopping. For this purpose, the motor vehicle 2 may first increase its speed beyond the maximum speed permitted in order to pass the motor vehicle 44. However, this may result in being below the minimum distance between the pedestrian 50 and the vehicle 2.
In subsequent interactions, the algorithm 34 may move the vehicle 2 more slowly, however this may cause a hazard to the motorcycle driver 48.
Subsequently, the algorithm 34 may first accelerate the vehicle 2 to pass the parked vehicle 44 and then re-brake. This solution is preferred because it enables, on the one hand, passing by the parked motor vehicle 44 and smoothly completing the current task, and, on the other hand, optimizing the measure of the hazard of the further traffic participants 48, 50.
Optimization may then be performed on additional criteria and metrics to further improve the maturity of the algorithm 34.
Fig. 4 shows a flow chart of the method.
First, computer program product modules are provided after the start. The computer program product module contains an algorithm to be trained and a self-learning neural network.
Subsequently, a simulated environment is provided on the basis of the real map data. The simulated environment may contain other traffic participants and their tasks in addition to roads and specific rules.
Starting from the basic algorithm, the rule set of the vehicle can be changed, which contains behavior rules, such as speed compliance, passing through a solid line, position on the driving lane, etc.
Simulations may then be performed in which various metrics are attempted to be achieved according to a reinforcement learning method. If this is not the case, the strategy or algorithm is changed and the simulation is repeated until a specific individual metric is reached. The method is repeated for all metrics.
Once all metrics are reached, the own vehicle's rule set is changed and the method is repeated until the algorithm is sufficiently mature. The algorithm may then be frozen.
The algorithm can be used, for example, in traffic simulation for other simulated vehicles than the motor vehicle to be trained. The method can also be applied to other traffic participants.
The training can be continued in a real environment that is either fully real or mixed real (gemischt-real).
Although the subject matter has been illustrated and described in detail by way of examples, the invention is not limited to the examples disclosed and further variants can be derived therefrom by the person skilled in the art. Therefore, a plurality of variant possibilities are evident. It is also clear that the exemplary embodiments mentioned merely show examples, which should not be construed in any way as limiting the scope of protection, application possibilities or configurations of the invention. Rather, the foregoing description and drawings describe and enable others skilled in the art to practice the exemplary embodiments, and it is, for example, contemplated that various changes may be made in the function and arrangement of elements described in the exemplary embodiments without departing from the scope of protection defined by the claims and their legal equivalents (e.g., as further set forth in the specification) in view of the inventive concepts disclosed.
List of reference numerals
2 Motor vehicle
4 controller
6 calculating unit
8 memory
10 environmental sensor
11 environmental sensor
12 Environment sensor
13 environmental sensor
14 vidicon
15 GPS module
16 wheel revolution sensor
18 acceleration sensor
20 pedal sensor
22 steering device
24 motor control device
26 brake
28 computer program product
30 computer program product module
32 neural network
34 Algorithm
36 simulation environment
38 road intersection
40. 42 road
44 parked motor vehicle
46 solid line
48 motorcycle driver
50 pedestrian
52 direction of movement of the pedestrian 50
54 planned trajectory
Claims (13)
1. Method for training at least one algorithm (34) for a controller (4) of a motor vehicle (2), wherein the controller (4) is provided for implementing an automated or autonomous driving function by intervening on a unit (22, 24, 26) of the motor vehicle (2) based on input data using the at least one algorithm (34), wherein the algorithm (34) is trained by means of a self-learning neural network (32), comprising the following steps:
a) providing a computer program product module (30) for the automated or autonomous driving function, wherein the computer program product module (30) contains the algorithm (34) to be trained and the self-learning neural network (32),
b) providing a simulation environment (36) having simulation parameters, wherein the simulation environment (36) contains map data (38) of a real-existing region of use, the motor vehicle (2) and at least one simulated further traffic participant (48, 50), wherein the behavior of the motor vehicle (2) and of the at least one further traffic participant (48, 50) is determined by a rule set, wherein the rule set contains behavior parameters which determine permissible limits,
c) providing a task for the motor vehicle (2),
d) -modifying at least one behavior parameter of the motor vehicle (2) such that the at least one behavior parameter is on the side of the allowed limit,
e) a simulation of the task is performed.
2. The method according to claim 1, wherein the neural network (32) is learned by a reinforcement learning method, wherein at least one of the time for completing the task and/or the number of accidents in which the motor vehicle (2) is involved during the task is used as a reward measure, wherein the simulation is repeated until a minimum measure is reached.
3. Method according to claim 1 or 2, wherein the at least one anomalous traffic participant is a motor vehicle (2), a motorcycle driver (48) or a pedestrian (50).
4. The method as claimed in one of the preceding claims, wherein the computer program product module (30) has an algorithm (34) which has been pre-trained by means of rule-compliant traffic participants (48, 50).
5. The method according to any of the preceding claims, wherein the at least one behavioural parameter is higher or lower by a predefined percentage.
6. The method according to any of the preceding claims, wherein the simulation is repeated a plurality of times, wherein at least one simulation parameter is changed each time separately.
7. The method of claim 6, wherein the simulated environment (36) is caused to change.
8. The method of claim 6 or 7, wherein at least one behavior parameter is varied.
9. Method according to one of claims 6 to 8, wherein the number, positioning and/or tasks of the further traffic participants (48, 50) are varied.
10. The method according to any one of the preceding claims, wherein the algorithm (34) is trained by a self-learning neural network (32), the method comprising the steps of:
a) providing a computer program product module (30) for the automated or autonomous driving function, wherein the computer program product module (30) contains the algorithm (34) to be trained and the self-learning neural network (32),
b) providing a simulated environment (36) with simulation parameters, wherein the simulated environment (36) contains map data (38) of a real-existing region of use, the motor vehicle (2) and at least one simulated further traffic participant (48, 50), wherein the simulated further traffic participant (48, 50) is simulated by an algorithm which has been trained according to one of claims 1 to 9,
c) providing a task for the motor vehicle (2),
e) a simulation of the task is performed.
11. The method according to any one of the preceding claims, wherein the computer program product module (30) is integrated in a controller (4) of a motor vehicle (2), and wherein the algorithm (34) is tested and/or trained in the real-existing region of use (38).
12. A computer program product having a computer-readable storage medium (8) on which are embedded instructions that, when executed by at least one computing unit (6), cause the computing unit (6) to be arranged for carrying out a method according to any one of the preceding claims.
13. A motor vehicle having the computer product of claim 12.
Applications Claiming Priority (3)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
DE102019216836.1A DE102019216836A1 (en) | 2019-10-31 | 2019-10-31 | Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle |
DE102019216836.1 | 2019-10-31 | ||
PCT/EP2020/079764 WO2021083785A1 (en) | 2019-10-31 | 2020-10-22 | Method for training at least one algorithm for a control device of a motor vehicle, computer program product, and motor vehicle |
Publications (1)
Publication Number | Publication Date |
---|---|
CN114667545A true CN114667545A (en) | 2022-06-24 |
Family
ID=73013437
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN202080076990.0A Pending CN114667545A (en) | 2019-10-31 | 2020-10-22 | Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle |
Country Status (4)
Country | Link |
---|---|
EP (1) | EP4052178A1 (en) |
CN (1) | CN114667545A (en) |
DE (1) | DE102019216836A1 (en) |
WO (1) | WO2021083785A1 (en) |
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117246345A (en) * | 2023-11-06 | 2023-12-19 | 镁佳(武汉)科技有限公司 | Method, device, equipment and medium for controlling generating type vehicle |
Families Citing this family (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN113345227B (en) * | 2021-05-31 | 2022-09-20 | 上海涵润汽车电子有限公司 | Random traffic flow generation method and device |
CN113392467A (en) * | 2021-06-15 | 2021-09-14 | 广东洪裕智能制造研究院有限公司 | Automobile part complex product collaborative simulation platform based on HLATH _ RTI |
DE102021206737A1 (en) * | 2021-06-29 | 2022-12-29 | Psa Automobiles Sa | Designing an automation algorithm for a vehicle |
CN114333493B (en) * | 2021-12-31 | 2022-10-14 | 江苏普旭科技股份有限公司 | Human body simulation system and method of driving simulator |
DE102022208519A1 (en) | 2022-08-17 | 2024-02-22 | STTech GmbH | Computer-implemented method and computer program for the movement planning of an ego driving system in a traffic situation, computer-implemented method for the movement planning of an ego driving system in a real traffic situation, control device for an ego vehicle |
CN116424332B (en) * | 2023-04-10 | 2023-11-21 | 重庆大学 | Energy management strategy enhancement updating method for deep reinforcement learning type hybrid electric vehicle |
Family Cites Families (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
DE102017007136A1 (en) * | 2017-07-27 | 2019-01-31 | Opel Automobile Gmbh | Method and device for training self-learning algorithms for an automated mobile vehicle |
DE102018217004A1 (en) * | 2017-10-12 | 2019-04-18 | Honda Motor Co., Ltd. | Autonomous vehicle strategy generation |
CN111919225B (en) * | 2018-03-27 | 2024-03-26 | 辉达公司 | Training, testing, and validating autonomous machines using a simulated environment |
-
2019
- 2019-10-31 DE DE102019216836.1A patent/DE102019216836A1/en active Pending
-
2020
- 2020-10-22 CN CN202080076990.0A patent/CN114667545A/en active Pending
- 2020-10-22 EP EP20796785.2A patent/EP4052178A1/en active Pending
- 2020-10-22 WO PCT/EP2020/079764 patent/WO2021083785A1/en unknown
Cited By (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN117246345A (en) * | 2023-11-06 | 2023-12-19 | 镁佳(武汉)科技有限公司 | Method, device, equipment and medium for controlling generating type vehicle |
Also Published As
Publication number | Publication date |
---|---|
WO2021083785A1 (en) | 2021-05-06 |
DE102019216836A1 (en) | 2021-05-06 |
EP4052178A1 (en) | 2022-09-07 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN114667545A (en) | Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle | |
EP3579211B1 (en) | Method and vehicle for assisting an operator of an ego-vehicle in controlling the ego-vehicle by determining a future behavior and an associated trajectory for the ego-vehicle | |
CN109598934B (en) | Rule and learning model-based method for enabling unmanned vehicle to drive away from high speed | |
CN110789528B (en) | Vehicle driving track prediction method, device, equipment and storage medium | |
CN112292719B (en) | Adapting the trajectory of an ego-vehicle to a moving foreign object | |
CN102320301B (en) | For the method making the ride characteristic of vehicle adapt to chaufeur conversion | |
US10831208B2 (en) | Vehicle neural network processing | |
CN111301419A (en) | Reinforcement learning based method for SAE4 level automated lane change | |
US12019414B2 (en) | Method for generating a training data set for training an artificial intelligence module for a control device of a vehicle | |
CN113168570A (en) | Method for training at least one algorithm for a control device of a motor vehicle, computer program product and motor vehicle | |
JP6790258B2 (en) | Policy generator and vehicle | |
CN111661045B (en) | Training generator unit and discriminator unit for collision-aware trajectory prediction | |
EP3629105A1 (en) | High-level decision making for safe and reasonable autonomous lane changing using reinforcement learning | |
Li et al. | An explicit decision tree approach for automated driving | |
CN114793460A (en) | Method for creating an algorithm for the computer simulation of a traffic participant, method for training an algorithm of at least one control device for a motor vehicle, computer program product and motor vehicle | |
EP3725609B1 (en) | Calibrating method for vehicle anti-collision parameters, vehicle controller and storage medium | |
CN111413973A (en) | Lane change decision method and device for vehicle, electronic equipment and storage medium | |
KR20230069940A (en) | Methods and systems for testing driver assistance systems | |
JP2023551944A (en) | Selection of driving maneuvers for at least partially automated vehicles | |
CN115176297A (en) | Method for training at least one algorithm for a control unit of a motor vehicle, computer program product and motor vehicle | |
JP2023546367A (en) | Method and system for testing driver assistance systems for vehicles | |
CN117242438A (en) | Method for testing a driver assistance system of a vehicle | |
CN115136081A (en) | Method for training at least one algorithm for a controller of a motor vehicle, method for optimizing a traffic flow in a region, computer program product and motor vehicle | |
JP7347252B2 (en) | Vehicle behavior evaluation device, vehicle behavior evaluation method, and vehicle behavior evaluation program | |
CN114987511A (en) | Method for simulating human driving behavior to train neural network-based motion controller |
Legal Events
Date | Code | Title | Description |
---|---|---|---|
PB01 | Publication | ||
PB01 | Publication | ||
SE01 | Entry into force of request for substantive examination | ||
SE01 | Entry into force of request for substantive examination |