Method for training a central artificial intelligence module

By using multi-vehicle environment data matching processing parameters to train the central artificial intelligence module on the server, the problems of low training efficiency and data security in the existing technology are solved, and efficient and safe automated driving learning and data processing are achieved.

CN109726795BActive Publication Date: 2024-12-03ROBERT BOSCH GMBH
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN201811275902.0
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2017-10-30
Filing Date
2018-10-30
Publication Date
2024-12-03
Estimated Expiration
2038-10-30

AI Technical Summary

Technical Problem

In the prior art, methods for training an automated driving artificial intelligence module require a large number of vehicle test kilometers, and transmission of environmental data may involve privacy information, and computing resources are concentrated on the vehicle, resulting in inefficiency and data security risks.

Method used

Using hardware-in-loop (HiL) simulation to match processing parameters of multi-vehicle environment, train the central artificial intelligence module through the server, use the processing parameters of other KI modules to learn, transmit compressed processing parameters to reduce the amount of data, and realize anonymous learning. And perform efficient computing in the cloud.

Benefits of technology

It reduces the computing load and data transmission on the vehicle, improves learning efficiency and safety, realizes rapid training and wide application of the central KI module, and ensures the diversity of driving manipulation and data privacy protection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN109726795B_ABST
    Figure CN109726795B_ABST
Patent Text Reader

Abstract

A method for training (106) a central artificial intelligence module (KI module) for highly automated or fully automated operation of a vehicle, wherein the central KI module is configured to convert input signals into output signals, wherein the conversion is carried out according to a processing chain that can be adapted by changing the values ​​of internal processing parameters, characterized in that the training (106) of the central KI module is carried out in such a way that internal processing parameters are changed based on further internal processing parameters of further KI modules, wherein the further KI modules are located in a plurality of vehicles and are respectively configured to convert input signals into output signals, wherein the conversion is respectively carried out according to a processing chain that can be adapted by changing the values ​​of further internal processing parameters, wherein the further KI modules are trained based on input signals based on environmental data detected by means of environmental sensor systems installed in the vehicle.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for training a central artificial intelligence module. Background Art

[0002] Learning methods for artificial intelligence are known from the prior art, in which a neural network is trained based on environmental data detected by the vehicle and stored centrally. Raw data from vehicle sensors are used, which may contain sensitive information that may limit the privacy of the vehicle user.

[0003] Furthermore, learning methods are known in which an artificial intelligence intended for automated driving is learned on a vehicle from environmental data detected by the vehicle. In order to collect a sufficiently large amount of data for learning the artificial intelligence, it is necessary for the vehicle to travel a very large number of test kilometers. Summary of the invention

[0004] The present invention describes a method for training a central artificial intelligence module (KI module) located on a server, the KI module being used for highly automated or fully automated operation of a vehicle. The method comprises the following steps:

[0005] performing at least one learning step for training a first KI module in a first vehicle or in a hardware-in-the-loop (HiL) simulation of the first vehicle based on an input signal, wherein the input signal is based on environmental data of the vehicle detected by means of an environmental sensor system installed in the first vehicle or recorded environmental data, wherein first internal processing parameters of the first KI module are matched during the training, and the input signal entering the first KI module is converted into an output signal based on the first internal processing parameters;

[0006] performing at least one learning step for training a second KI module in a second vehicle or a hardware-in-the-loop (HiL) simulation of the second vehicle based on an input signal, wherein the input signal is based on environmental data of the vehicle detected by means of an environmental sensor system installed in the second vehicle or recorded environmental data, wherein second internal processing parameters of the second KI module are adapted during the training, and the input signal entering the second KI module is converted into an output signal based on the second internal processing parameters;

[0007] transmitting the matched first internal processing parameter to the server;

[0008] transmitting the matched second internal processing parameter to the server;

[0009] The central KI module is trained based on the transmitted matched first and second internal processing parameters.

[0010] The server can be, for example, a central computer which is connected to the Internet and can receive data from the vehicle and send data to the vehicle. The training of the central KI module is carried out in particular on the server.

[0011] The artificial intelligence module (hereinafter referred to as KI module in this application) located on the central server can be a module that can process problems independently. As a KI module, for example, an artificial neural network, a recursive neural network, a convolutional neural network or a network based on back propagation can be applied. It is also possible to consider the application of an autoencoder or a similar structure or algorithm. A combination of multiple algorithms or network types is also possible. In general, multiple networks can also be interconnected, and the output values ​​of each network are used as input values ​​to another network.

[0012] For example, further KI modules used in vehicle or hardware-in-the-loop (HiL) simulations can be constructed accordingly. In one embodiment of the invention, the further KI modules can even be designed identically. The further KI modules also have the task of solving problems independently. To this end, input data are converted into output data based on variable internal processing parameters. When the further KI modules are learning, these processing parameters are matched based on the input data corresponding to the detected environmental data. For this purpose, methods such as reinforcement learning can be used, for example. In addition or as an alternative, the desired output data can also be provided based on the input data and thus a learning set (Lernset), according to which the KI modules are trained. The matched processing parameters of the further KI modules can be transmitted to the central KI module independently of the learning state of the KI modules.

[0013] The central KI module is characterized in that it is trained based on the transmitted processing parameters of the other KI modules, in particular the first KI module and the second KI module. The processing parameters of the other KI modules can be used to determine the central processing parameters of the central KI module, for example, by means of an algorithm based on the stochastic gradient method. The central processing parameters can also be determined, for example, by averaging the corresponding other processing parameters.

[0014] If the KI module is designed as a neural network, the processing parameters can include in particular the values ​​of the weights, which determine the information forwarding between different neurons. In the learning of neural networks, it is common practice to randomly select the starting values ​​of the weights. For distributed learning with multiple KI modules, it is suitable to select all initial values ​​of the weights in the individual KI modules identically. This can increase the probability that the weights converge in the central KI.

[0015] The transmission of the processing parameters of the first KI module and / or the second KI module is realized, for example, by a conventional Car2X communication unit and a Car2X communication method. For example, transmission is possible by radio, for example by WLAN and / or UMTS and / or LTE.

[0016] If the first and / or second KI module is located in a HiL simulation, environmental data recorded from any vehicle can be used for the training of the KI module. Thus, environmental data from different vehicles or from the same vehicle can be used for the training of the first and second KI module. It is expedient here to select environmental data from different driving situations if environmental data from the same vehicle are used. On the one hand, the environmental data can be fed into the KI module in real time, and on the other hand, the environmental data can first be recorded and fed / input into the KI module with a time offset. For the HiL simulation, recorded data are particularly suitable.

[0017] A highly automated or highly automated vehicle is understood to be a vehicle that can independently carry out at least predefined driving functions without driver intervention. Exemplary driving functions are systems with automated lateral and / or longitudinal guidance, such as parking assistance, traffic jam assistance, highway pilot (Autobahnpilot), etc.

[0018] Fully automated vehicles or vehicles that operate fully automatically are able to perform many of these driving functions, so that even complex routes can be driven independently without driver intervention if necessary. Fully automated vehicles can master almost all typical driving situations if necessary, so that they can participate in road traffic independently without driver intervention. If necessary, there is no longer any possibility for driver intervention in these vehicles, since a driver is no longer required.

[0019] Vehicles are understood here to mean land vehicles, air vehicles, space vehicles and water vehicles, in particular passenger cars, trucks, buses, shuttles, two-wheeled vehicles, boats, airplanes, helicopters and drones. For example, it is also possible to train a central KI module with distributed KI modules in an automated robot and / or forklift for operating an automated robot and / or forklift.

[0020] The method claimed in the present application offers the following advantages: the learning of the central KI module does not necessarily require the use of input data, but learning can also be achieved based on the processing parameters of other KI modules. In addition, it can be sufficient for the other KI modules located in the vehicle to perform only a small number of learning steps, which are not sufficient for the training of the KI module by themselves, but provide a sufficiently good learning result in combination with the other KI modules. The individual vehicles therefore also require less computing power if necessary and can thus be manufactured more cost-effectively. In addition, the learning results can be used for training the central KI module in a time-staggered manner with the help of the method, because the processing parameters can be temporarily stored. Another important advantage of the method compared to the following learning methods - in which environmental data detected by the vehicle are used for training the central KI module - is that these environmental data no longer need to be transmitted. The input data in the form of matched processing parameters are compressed according to the learning process of the KI module located in the vehicle. Since only the compressed processing parameters are transmitted to the central KI module, the amount of data to be transmitted can be greatly reduced. As a result, the transmission capacity of the Car-to-X communication connection from the individual vehicles is almost unloaded, because only the calculation results of the learning steps are transmitted to the cloud. In addition, the input signal and / or the environmental data cannot be re-established / reconstructed from the transmitted processing parameters. As a result, the input signal and / or the environmental data containing personal data are protected if necessary, and anonymous learning without sensitive data can be achieved. By installing the KI module in as many vehicles as possible, especially in different areas, it can be ensured that the most extensive learning results of driving maneuvers are available on the server for the calculation of the central KI module. The learning process of the central KI module can also be parallelized based on the parallel learning process in the vehicle and thus extremely accelerated. Learning on the central server also provides the following advantage, namely that a lot of computing power can be provided to the server. Therefore, the training of the central KI module can be achieved in an efficient manner by HPC (high performance computing).

[0021] If the first and / or second KI module performs a learning step in a HiL simulation, input data can also be fed in more frequently than in the vehicle, since the processing speed in a correspondingly designed HiL is higher than in the vehicle. As a result, the individual learning steps can be carried out more quickly and efficiently.

[0022] In a further embodiment of the method, the central KI module is configured to convert the input signal into the output signal based on adaptable central internal processing parameters. In the step of training the central KI module, the central internal processing parameters are adapted based on the adapted first and second internal processing parameters. In particular, no input signal of the first vehicle or of the second vehicle or of the first HiL simulation or of the second HiL simulation is used for adapting the central internal processing parameters.

[0023] In another embodiment of the method, individual driving functions and / or driving maneuvers are learned with the aid of the first and / or second KI module. This makes it possible to assign different learning tasks to different vehicles, wherein the vehicles are driven, for example, only in certain areas, such as on highways. By distributing the individual learning tasks, the performance of the individual KI modules and the central KI module can be improved.

[0024] In another embodiment of the method, the same internal output processing parameters are selected for the first KI module and the second KI module, wherein the internal output processing parameters correspond to the internal processing parameters before the first learning process of the KI module. In particular, the first internal output processing parameter and the second internal output processing parameter have the same output value.

[0025] This embodiment of the invention offers the advantage that, due to the start of the learning process in different vehicles with the same KI module, a better integration of the processing parameters can be achieved. This means that the processing parameters converge better when determining the processing parameters for the central KI module. This makes it possible to achieve a more reliable performance of the central KI module and, if necessary, a faster training process for the central KI module.

[0026] In another embodiment of the method, the server receives adapted internal processing parameters of further vehicles having a KI module, and the method comprises an additional step of selecting at least one of the further vehicles. Here, only the adapted internal processing parameters of the selected at least one vehicle are used for training the central KI module.

[0027] This embodiment of the invention offers the advantage that certain vehicles are taken into account and other vehicles are not taken into account when training the central KI module. By intelligent preselection, the learning process can be accelerated and the final result / performance of the central KI module can be optimized. This selection can be achieved, for example, based on the number of learning steps that have been performed for the other KI modules. In addition, convergence criteria for the selection can also be considered. For example, changes in processing parameters can be analyzed based on the last learning step. If the change is too large, this requires a processing chain of the KI module that is not yet strongly converged. If such processing parameters are not used to train the KI module, the corresponding vehicle or its matched processing parameters can be obtained in this way and therefore remain unconsidered. According to the embodiment of the method, the following processing parameters can also be used: the processing parameters have been matched based on only one learning step or a very small number of learning steps.

[0028] In a further embodiment of the method, the method comprises the additional step of transmitting the central KI module to at least one vehicle.

[0029] This embodiment of the invention offers the advantage that, by transferring the central KI module, a KI module with the same processing parameters is available for the vehicle. Based on this KI module, further learning processes can be carried out in the vehicle. Since all vehicles have the same intermediate state, the convergence probability of the central processing parameters can be increased.

[0030] In a further embodiment of the method, the central KI module transmitted to the vehicle is used for operating the vehicle, so that the method has an additional step of operating the vehicle in a highly automated or fully automated manner with the aid of the transmitted central KI module.

[0031] This embodiment of the invention offers the advantage that all vehicles benefit from the training of the central KI module and the optimized KI module can be used to control the vehicle. This increases the safety of vehicle passengers and living beings in the vehicle environment. In addition, infrastructure and other traffic participants are protected.

[0032] In another embodiment of the method, environmental data detected by means of an environmental sensor system of the first vehicle and / or the second vehicle are temporarily stored in the first vehicle and / or the second vehicle, and the first KI module and / or the second KI module are trained based on the temporarily stored environmental data. In particular, the data are temporarily stored for longer than 10 seconds, further in particular longer than 10 minutes, further in particular longer than 24 hours.

[0033] This embodiment of the invention offers the advantage that the learning step does not have to be carried out in real time in the individual KI modules located in the vehicle, but can be carried out at any time. Depending on the required computing power and the required energy requirements for carrying out the calculation or learning step, the calculation can be carried out, for example, only when the respective vehicle is stationary and / or during a battery charging process. If the KI module is trained during driving, a buffer time of 10 seconds or 10 minutes may be sufficient. If the KI module is trained only during a charging process, a significantly longer buffer time of more than 24 hours may be necessary.

[0034] Based on a small number of learning steps of the first and / or second KI module, these learning steps are not normally used for vehicle control for safety reasons. The KI module is preferably used as an intermediate KI module, whose learning results are primarily used to train the central KI module. The weights adapted in the individual learning steps are transmitted to a central server or cloud.

[0035] Furthermore, a method for training a central artificial intelligence module KI module located on a server, which is used for highly automated or fully automated operation of a vehicle, is claimed, and the method can only be executed on the server. The method comprises the following steps:

[0036] receiving a value of a matched first internal processing parameter from a first KI module of a first vehicle and / or a first HiL simulation, wherein an input signal entering the first KI module is transformed into an output signal according to the first internal processing parameter, wherein at least one learning step has been performed by the first KI module, in which the first internal processing parameter has been matched, wherein the matching of the first internal processing parameter is performed based on environmental data of the vehicle detected by means of an environmental sensor installed in the first vehicle or recorded environmental data;

[0037] receiving values ​​of matched second internal processing parameters from a second KI module of a second vehicle and / or a second HiL simulation, wherein an input signal entering the second KI module is transformed into an output signal according to the second internal processing parameters, wherein at least one learning step has been performed by the second KI module, in which the second internal processing parameters have been matched, wherein the matching of the second internal processing parameters is carried out based on environmental data of the vehicle detected by means of an environmental sensor installed in the second vehicle or recorded environmental data;

[0038] The central KI module is trained based on the received matched first and second internal processing parameters.

[0039] Furthermore, a central KI module for highly automated or fully automated operation of a vehicle is claimed, wherein the central KI module is configured to convert an input signal into an output signal, wherein the conversion takes place based on internal processing parameters. The KI module is trained in accordance with the method for training a KI module based on the present application.

[0040] Furthermore, a control system for a highly automated or fully automated vehicle is claimed, which control system comprises the central KI module claimed in the present application.

[0041] This embodiment of the invention offers the advantage that an intelligent control system for a vehicle is provided, by means of which safe operation of the vehicle can be achieved.

[0042] Furthermore, protection is claimed for a computer program which contains parameters for a KI module, which, if loaded into the KI module, upgrade the KI module into a central KI module according to the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 : shows a schematic flow chart;

[0044] Figure 2 : Another schematic flow chart is shown. DETAILED DESCRIPTION

[0045] In one embodiment, the central KI module is located in a cloud computer, which includes a communication module for exchanging data with the vehicle. The central KI module has the following task, namely, converting input signals into output signals based on internal processing parameters. In particular, the KI module should be able to realize vehicle control based on incoming environmental data. The KI module is constructed as a neural network with multiple hidden layers, wherein the information transmission between neurons located in the layers is defined by weights. These weights are variable and are decisively responsible for the generation of output signals. In order to train the central KI module, in this embodiment, the input data is not used, but the processing parameters, especially weights, of another KI module are used, wherein the other KI module is located in the vehicle. In this embodiment, the other KI module is located in a plurality of cars of the same structure, wherein the plurality of cars detect the same environmental data, especially by means of the same environmental sensor, and in the plurality of cars, the same actuator is controlled.

[0046] The method used to train the central KI module is implemented in Figure 1 Schematically depicted in and starting in step 101.

[0047] In step 102, a learning step for training a first KI module located in a first vehicle is implemented based on the input signal and the implemented reinforcement learning method, which is designed to convert the input signal into an output signal based on internal processing parameters. The output signal is environmental data, which is recorded with the help of an environmental sensor installed in the first vehicle. For this purpose, the vehicle has a plurality of cameras, a 360° lidar, a radar sensor, a global navigation satellite system sensor and an ultrasonic sensor. In the learning step, first internal processing parameters of the first KI module are adapted based on the input signal, in particular weights between artificial neurons.

[0048] In step 103 , a learning step is also carried out in the second vehicle, which is configured identically to the first vehicle and has a second KI module. In this learning step, second internal processing parameters of the second KI module are adapted based on the input signal.

[0049] In step 104 , the adapted first process parameters are transmitted from the first vehicle to the central KI module. For this purpose, the first vehicle has a communication unit, by means of which the process parameters are transmitted by radio.

[0050] The adapted second process parameters are transmitted from the second vehicle to the central KI module in step 105. For this purpose, the second vehicle has a communication unit, by means of which the process parameters are transmitted by radio.

[0051] In step 106 , the central KI module is trained based on the transmitted first and second processing parameters. In this case, the values ​​of the individual processing parameters of the first and second vehicles are averaged, wherein the averaged processing parameters thus generated form the central processing parameters.

[0052] The method ends in step 107 .

[0053] In a further embodiment, in step 106, the processing parameters are not simply averaged during the training of the central KI module, but the processing parameters are fed into the training process in a weighted manner. For this purpose, on the one hand, a convergence criterion for the individual KI modules is included, which indicates how strongly the weights of the individual KI modules have converged. The stronger the convergence criterion, i.e. the better the convergence, in the individual KI modules, the more strongly its weight is weighted during the training. In addition, the number of learning steps that have been carried out is fed into the weighting of the processing parameters during the training of the central KI module. In addition, the frequency of the individual driving maneuvers carried out and the criticality of the driving maneuvers can also be included in the weighting of the processing parameters.

[0054] In a further exemplary embodiment, the central KI module is trained based on individual KI modules which are located both in the vehicle and in the HiL.

[0055] In another embodiment, if the vehicle is at a charging station or is connected to the Internet separately by means of a cable, the processing parameters are transmitted via the cable connection. If the vehicle is stopped and connected to a power supply, the learning step in the vehicle is preferably performed here. This allows the full power of the computing unit on the vehicle to be fully utilized.

[0056] In another embodiment, steps 101 to 106 are repeated for such a long time until the value of the central processing parameter has sufficiently converged. Here, the conditions defining sufficient convergence are first specified. In this embodiment, during the step of training the central KI module, the value of the processing parameter is not allowed to change by a value > 0.2, wherein the processing parameter can only have values ​​between 0 and 1.

[0057] In a further exemplary embodiment, the method has an additional method step 116 .

[0058] In step 116, the trained central KI module is transmitted to the first and second vehicles. The transmission is used to match the processing parameters of the KI modules located in the vehicles to each other. This allows for better convergence.

[0059] In another exemplary embodiment, the trained central KI module is also transmitted to other vehicles in step 116 .

[0060] In a further step 126 , the vehicle to which the central KI module has been transmitted is operated automatically based on the central KI module. The KI module can either take over individual driving tasks or enable fully automated control of the vehicle.

[0061] In another exemplary embodiment, a plurality of central KI modules located on a cloud computer (hereinafter referred to as the cloud) are trained for an autonomous vehicle variant using a distributed system of a plurality of vehicles of the same variant, wherein the KI modules are respectively implemented in the vehicles. The learning process of the central KI module is divided into Figure 2 The steps shown in and start with step 201.

[0062] In step 202 , a learning step for a specific driving maneuver, in this case an overtaking process on a highway, is carried out on a vehicle using the vehicle's surroundings data. In this case, a KI module located in the vehicle is trained.

[0063] In step 203 , a learning step for a specific driving maneuver, in this case an overtaking process on a highway, is carried out on the other vehicle using the environmental data of the other vehicle. In this process, the KI module located in the other vehicle is trained.

[0064] In step 204 , the results of the step of learning the driving maneuver are transmitted from the two vehicles in steps 202 and 203 to the cloud.

[0065] In step 205 , a calculation of a specific driving maneuver is performed in the cloud at the central KI module based on the results of the individual learning steps of the vehicles by combining the learning results of both vehicles.

[0066] In step 206 , these learning steps are repeated, wherein further vehicles learn further specific driving maneuvers and the learning results generated in this process are used to train further central KI modules. In this case, the same vehicle can also learn further driving maneuvers.

[0067] In step 207, the individual results of the specific driving maneuvers learned at the central KI modules are combined in a central master KI module. The master KI module is composed of central KI modules and has an additional situation recognition module, which detects the current driving situation and recognizes the driving maneuver required in the driving situation and, based on the detection and recognition, supplies the corresponding central KI module with environmental data and uses its output signals for controlling the vehicle.

[0068] In step 208 , the central master KI module is transmitted to a plurality of vehicles so as to enable automated control based on the KI modules in the central master KI module.

[0069] The method ends in step 209 .

[0070] In another embodiment, the environmental data are temporarily stored in a vehicle memory of the vehicle in step 202, so that the learning steps for the driving maneuver can also be calculated in the vehicle afterwards or gradually. Therefore, the calculation of the learning step does not have to be carried out in real time. As a result, the computing power of the vehicle computer can be designed to be relatively small. After the learning step has been carried out on a set of environmental data, the environmental data of the learning step are deleted from the vehicle memory or overwritten with environmental data for another learning step. By transmitting the learning results of the same driving maneuver of different vehicles to the cloud step 204, it is ensured that the most extensive learning results of the driving maneuver are available for the calculation of the central KI module in the cloud. The transmission capacity of the Car-to-X communication connection from the individual vehicles is hardly loaded, because only the calculation results of one learning step are transmitted to the cloud.

[0071] Repeating steps 202 to 205 for multiple vehicles and driving maneuvers leads to more efficient and faster learning of the central KI module, since any number of input data (learning steps) from any number of vehicles can be used in parallel. The computationally intensive training of the central KI module is carried out in an efficient manner in the cloud by combining the learning results of the individual learning steps via HPC (High Performance Computing).

[0072] After the central KI module has been transmitted back to the individual vehicles in step 208 , the central KI module can be executed there in order to implement autonomous driving functions in the vehicle.

[0073] In another embodiment, the learning of the KI is performed for autonomous vehicle variants by means of a HiL system (hardware in the loop). In this case, the learning steps that previously took place in the vehicle are implemented on the HiL system. The system accesses environmental data recorded by a plurality of vehicles from different driving situations.

[0074] In another embodiment, the server is located on the HiL system on which the individual KI modules are additionally learned. Here, after each learning step of the individual KI modules, the central KI module is updated in parallel, while further learning steps are performed in parallel on the individual KI modules of the HiL.

Claims

1. A method for training a central artificial intelligence module, a central KI module, located on a server, for highly automated or fully automated operation of a vehicle, the method comprising the following steps: performing (102) at least one learning step for training a first KI module located in a first vehicle or in a hardware-in-the-loop simulation of the first vehicle based on an input signal, wherein: The input signal is based on environmental data detected by the vehicle with the aid of an environmental sensor mechanism installed in the first vehicle or recorded environmental data, wherein a first internal processing parameter of the first KI module is matched during training, and the input signal entering the first KI module is converted into an output signal based on the first internal processing parameter; at least one learning step is performed (103) according to the input signal for training a second KI module located in a second vehicle or a hardware-in-the-loop simulation of the second vehicle, wherein the input signal is based on environmental data detected by the vehicle with the aid of an environmental sensor mechanism installed in the second vehicle or recorded environmental data, wherein a second internal processing parameter of the second KI module is matched during training, and the input signal entering the second KI module is converted into an output signal based on the second internal processing parameter; and the matched first internal processing parameter is transmitted (104) to the second vehicle. said server; transmitting (105) the matched second internal processing parameter to said server; training (106) said central KI module based on the transmitted matched first and second internal processing parameters, wherein said central KI module is configured to transform an input signal into an output signal based on the matched central internal processing parameters, wherein in the step of training (106) said central KI module, said central internal processing parameters are matched based on the matched first and second internal processing parameters, wherein no input signal of the first vehicle or of the second vehicle or of a first hardware-in-the-loop simulation of the first vehicle or of a second hardware-in-the-loop simulation of the second vehicle is used for matching said central internal processing parameters, wherein the values ​​of the matched first and second internal processing parameters are averaged, wherein the average processing parameters thus generated form said central internal processing parameters.

2. The method according to claim 1, characterized in that: Individual driving functions and / or driving maneuvers are learned with the aid of the first KI module and / or the second KI module.

3. The method according to claim 1 or 2, characterized in that: The same internal output processing parameters are selected for the first KI module and the second KI module, wherein the internal output processing parameters correspond to the internal processing parameters before the first learning process of the KI module, and wherein the first internal output processing parameters of the first KI module and the second internal output processing parameters of the second KI module have the same output value.

4. The method according to claim 1 or 2, characterized in that: The server receives matched internal processing parameters of further vehicles having KI modules, and the method comprises the additional step of selecting at least one of the further vehicles, and using only the matched internal processing parameters of the selected at least one vehicle for training the central KI module (106).

5. The method according to claim 1 or 2, comprising the additional step of transmitting (116) the central KI module to at least one vehicle. 6 . The method according to claim 5 , comprising the additional step of operating ( 126 ) the vehicle in a highly automated or fully automated manner using the transmitted central KI module.

7. The method according to claim 1 or 2, characterized in that: Environmental data detected by means of an environmental sensor system of the first vehicle and / or the second vehicle are temporarily stored in the first vehicle and / or the second vehicle, and the first KI module and / or the second KI module are trained based on the temporarily stored environmental data.

8. The method according to claim 7, characterized in that The data is temporarily stored for longer than 10 seconds.

9. The method according to claim 7, characterized in that: The data is temporarily stored for longer than 10 minutes.

10. The method according to claim 7, characterized in that The data is temporarily stored for longer than 24 hours.

11. A method for training a central artificial intelligence module located on a server, a central KI module, for highly automated or fully automated operation of a vehicle, the method comprising the following steps: receiving values ​​of matched first internal processing parameters from a first KI module of a first vehicle and / or a first hardware-in-the-loop simulation, wherein: The input signal entering the first KI module is converted into an output signal according to the first internal processing parameter, wherein at least one learning step has been performed by the first KI module, and the first internal processing parameter has been matched in the at least one learning step, wherein the matching of the first internal processing parameter is achieved based on environmental data detected by the vehicle by means of an environmental sensor mechanism installed in the first vehicle or recorded environmental data; receiving a value of a matched second internal processing parameter from a second KI module of a second vehicle and / or a second hardware-in-the-loop simulation, wherein the input signal entering the second KI module is converted into an output signal according to the second internal processing parameter, wherein at least one learning step has been performed by the second KI module, and the second internal processing parameter has been matched in the at least one learning step, wherein the matching of the second internal processing parameter is based on the ... The invention relates to a method for adjusting the control parameters of a vehicle by using environmental data detected by an environmental sensor installed in the second vehicle or recorded environmental data; training the central KI module based on the received matched first and second internal processing parameters, wherein the central KI module is configured to convert an input signal into an output signal based on the matched central internal processing parameters, wherein in the step of training (106) the central KI module, the central internal processing parameters are matched based on the matched first and second internal processing parameters, wherein the input signal of the first vehicle or the second vehicle or the first hardware-in-the-loop simulation of the first vehicle or the second hardware-in-the-loop simulation of the second vehicle is not used for matching the central internal processing parameters, wherein the values ​​of the matched first and second internal processing parameters are averaged, wherein the average processing parameters thus generated form the central internal processing parameters.

12. A central KI module for highly automated or fully automated operation of a vehicle, wherein: The central KI module is provided for converting an input signal into an output signal, wherein the conversion is carried out on the basis of internal processing parameters, characterized in that the central KI module has been trained in accordance with one of the training methods according to any one of claims 1 to 11.

13. A control system for a highly automated or fully automated vehicle, the control system comprising at least one central KI module according to claim 12.

14. A computer program product comprising parameters for a KI module, which if loaded into a KI module upgrades the KI module to a central KI module according to claim 12.

Citation Information

Patent Citations

  • Distributed training of reinforcement learning systems

    CN107209872A

  • Non-robot car, robot car, road traffic system, vehicle sharing system, robot car training system, and robot car training method

    WO2017057528A1