Method and apparatus for generating control commands for an autonomous road vehicle

Abstract

Described herein is a method and apparatus (1) for generating validated control commands (2) for an autonomous road vehicle (3). An end-to-end trained neural network system (4) is arranged to receive input of raw sensor data (5) from onboard sensors (6) of the autonomous road vehicle (3) as well as object level data (7) and policy information data (8). The end-to-end trained neural network system (4) is further arranged to map the input data (5, 7, 8) to control commands (10) for the autonomous road vehicle (3) within a preset time range. A safety module (9) is arranged to receive control commands (10) for the autonomous road vehicle (3) within a preset time range and to perform a risk assessment of a planned trajectory resulting from the control commands (10) for the autonomous road vehicle (3) within the preset time range. The safety module (9) is further arranged to verify as safe and output validated control commands (2) for the autonomous road vehicle (3).

Description

Technical Field

[0001] The present disclosure relates generally to autonomous vehicle control technology, and in particular to a method for generating validated control commands for an autonomous road vehicle. The present disclosure also relates to an apparatus for generating validated control commands for an autonomous road vehicle and an autonomous vehicle including such an apparatus. Background Art

[0002] In order to perform travel to some intended destination, autonomous road vehicles typically need to process and interpret a large amount of information. Such information may include visual information, e.g., information captured from a camera, information from radar or lidar, and may also include information obtained from other sources, such as from GPS devices, speed sensors, accelerometers, suspension sensors, etc.

[0003] During the travel of an autonomous road vehicle, decisions need to be made in real time based on available information. Such decisions may include decisions to perform braking, acceleration, lane changes, turns, U-turns, reversing, etc., and the autonomous road vehicle is controlled based on the results of the decisions.

[0004] WO2017125209 A1 discloses a method for operating a motor vehicle system, the motor vehicle system being designed to guide the motor vehicle in different driving situation categories in a fully automated manner, wherein - a computing structure comprising a plurality of analysis units is used to determine control data to be used as driving situation data for guiding the motor vehicle in a fully automated manner from surrounding data describing the surroundings of the motor vehicle and self data describing the state of the motor vehicle and to use the control data in order to guide the motor vehicle, - each analysis unit determines output data and / or output data of driving situation data from at least one other analysis unit, and - at least some of the analysis units are designed as neural networks at least partially on the basis of software. At least some of the analysis units designed as neural networks are dynamically generated from a configuration object, which can be configured during runtime using a configuration parameter set, wherein - a current driving situation category is determined from a plurality of specified driving situation categories, each driving situation category being assigned at least one analysis function using at least some driving situation data, - a configuration parameter set assigned to the analysis function of the current driving situation category is retrieved from a database, and - an analysis unit that performs an analysis function and has not yet been provided is generated by configuring the configuration object using the retrieved configuration parameter set.

[0005] WO2017120336 A2 discloses a system and method for navigating an autonomous vehicle using reinforcement learning techniques. In one embodiment, a navigation system for a main vehicle may include at least one processing device, the processing device being programmed to: receive multiple images representing the environment of the main vehicle from a camera; analyze the multiple images to identify a navigation state associated with the main vehicle; provide the navigation state to a trained navigation system; receive a desired navigation action from the trained navigation system for the main vehicle to perform in response to the identified navigation state; analyze the desired navigation action relative to one or more predefined navigation constraints; determine an actual navigation action for the main vehicle, wherein the actual navigation action includes at least one modification of the desired navigation action determined based on one or more predefined navigation constraints; and cause at least one adjustment of the navigation actuator of the main vehicle in response to the determined actual navigation action for the main vehicle.

[0006] US2017357257 A1 discloses a vehicle control method and device and a method and device for obtaining a decision model. The vehicle control method includes the steps of obtaining current external environment information and map information in real time during the operation of an unmanned vehicle; and determining vehicle state information corresponding to the external environment information and map information obtained each time according to the decision model, wherein the decision model is obtained by pre-training and reflects the corresponding relationship between the external environment information and the map information and the vehicle state information, and the driving state of the unmanned vehicle is controlled according to the determined vehicle state information.

[0007] Further from the publication "End to End Learning for Self-Driving Cars", arXiv:1604.07316v1 [cs.CV], April 25, 2016, by Mariusz Bojarski et al., a convolutional neural network (CNN) is trained to map raw pixels from a single front-facing camera of an autonomous road vehicle directly to steering commands. With minimal training data from humans, the correspondingly trained system can learn to drive on local roads with or without lane markings and on highways in traffic. It can also operate in areas where visual guidance is unclear, such as in parking lots and on unpaved roads. Such a system can automatically learn internal representations of necessary processing steps such as detecting useful road features using only human-provided steering angles as training signals. Summary of the invention

[0008] It is an object of the present invention to provide a method and an arrangement which provide improved safety when generating control commands for an autonomous road vehicle.

[0009] The invention is defined by the attached independent claim. Embodiments are set forth in the attached dependent claims and in the drawings.

[0010] According to a first aspect, a method for generating verified control commands for an autonomous road vehicle is provided, the method comprising: providing raw sensor data from on-board sensors of the autonomous road vehicle as well as object level data and policy information data as input data to an end-to-end trained neural network system; mapping the input data to control commands for the autonomous road vehicle within a preset time frame by the end-to-end trained neural network system; subjecting the control commands for the autonomous road vehicle within the preset time frame to a safety module, the safety module being arranged to perform a risk assessment of a planned trajectory resulting from the control commands for the autonomous road vehicle within the preset time frame; verifying as safe and outputting the verified control commands for the autonomous road vehicle from the safety module.

[0011] In further embodiments, the method further comprises adding a machine learning component to the end-to-end trained neural network system.

[0012] In still further embodiments, the method further includes providing validated control commands for the autonomous road vehicle verified as safe by the safety module as feedback to the end-to-end trained neural network system.

[0013] In still further embodiments, the method includes providing as raw sensor data at least one of: image data; velocity data and acceleration data from one or more onboard sensors of the autonomous road vehicle.

[0014] In additional embodiments, the method further comprises providing as object level data at least one of: positions of surrounding objects; lane markings and road conditions.

[0015] In yet additional embodiments, the method further comprises providing as strategic information data at least one of the following: electronic horizon (map) information including current traffic regulations and road geometry, and advanced navigation information.

[0016] According to a second aspect, there is provided an apparatus for generating verified control commands for an autonomous road vehicle, comprising: an end-to-end trained neural network system arranged to receive inputs of raw sensor data from onboard sensors of the autonomous road vehicle as well as object-level data and policy information data; the end-to-end trained neural network system is further arranged to map the input data to control commands for the autonomous road vehicle within a preset time range; a safety module arranged to receive control commands for the autonomous road vehicle within a preset time range and to perform a risk assessment of a planned trajectory generated by the control commands for the autonomous road vehicle within the preset time range; the safety module is further arranged to verify as safe and output verified control commands for the autonomous road vehicle.

[0017] In additional embodiments, the apparatus further comprises an end-to-end trained neural network system further comprising a machine learning component.

[0018] In yet further embodiments, the apparatus is further arranged to feed back validated control commands for the autonomous road vehicle verified as safe by the safety module to the end-to-end trained neural network system.

[0019] In still further embodiments, the apparatus further comprises an end-to-end trained neural network system arranged to receive as raw sensor data at least one of: image data; speed data and acceleration data from one or more onboard sensors of the autonomous road vehicle.

[0020] In an additional embodiment, the apparatus further comprises an end-to-end trained neural network system arranged to receive at least one of the following as object-level data: positions of surrounding objects; lane markings and road conditions.

[0021] In yet an additional embodiment, the apparatus further comprises an end-to-end trained neural network system arranged to receive as strategic information data at least one of: electronic horizon (map) information including current traffic regulations and road geometry, and high-level navigation information.

[0022] Furthermore, envisaged herein is an autonomous road vehicle comprising means for generating validated control commands as set out herein.

[0023] The above embodiments have the beneficial effect of providing an end-to-end solution for making improved safe overall decisions for autonomous road vehicles in complex traffic environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] In the following, embodiments herein will be described in more detail by way of example only with reference to the accompanying drawings, in which:

[0025] Figure 1 A method for generating validated control commands for an autonomous road vehicle according to embodiments herein is schematically illustrated.

[0026] Figure 2 An example embodiment of an apparatus comprising an end-to-end trained neural network system for generating validated control commands for an autonomous road vehicle is schematically illustrated.

[0027] Figure 3 An autonomous road vehicle according to an example embodiment is schematically illustrated, comprising means for generating validated control commands for the autonomous road vehicle. DETAILED DESCRIPTION

[0028] In the following, some example embodiments of a method and apparatus 1 for generating validated control commands for an autonomous road vehicle 3 will be described. The autonomous road vehicle 3 may be a car, a truck, a bus, etc. The autonomous road vehicle 3 may also be a fully autonomous (AD) vehicle or a partially autonomous vehicle with an advanced driver assistance system (ADAS).

[0029] According to the proposed method, Figure 1 As illustrated in , raw sensor data 5 from onboard sensors 6 of an autonomous road vehicle 3 as well as object level data 7 and policy information data 8 are provided 16 as input data to an end-to-end trained neural network system 4 .

[0030] The raw sensor data 5 may be provided as at least one of: image data; speed data and acceleration data from one or more onboard sensors 6 of the autonomous road vehicle 3. Thus, the raw sensor data 5 may be, for example, images from a vision system of the autonomous road vehicle and information about speed and acceleration, such as, for example, available via a controller area network (CAN) or similar network tapped into the vehicle. The image data may include images from a surround view camera and also consist of several recent images.

[0031] The object level data 7 may, for example, be provided as at least one of: positions of surrounding objects; lane markings and road conditions.

[0032] The strategy information data 8 may be provided, for example, as at least one of the following: electronic horizon (map) information including current traffic regulations and road geometry, and advanced navigation information. The electronic horizon (map) information about current traffic regulations and road geometry enables the neural network system 4 to take into account things such as permissible speeds, motorway exits, roundabouts and distances to intersections when making decisions.

[0033] In order to guide the neural network system 4 in situations where a lane change, highway exit, or similar action is required, high-level navigation information should be included as input to the neural network system 4. In this way, the neural network system 4 will be able to take into account user preferences, such as driving to a user-specified location.

[0034] The raw sensor data 5 (such as image data) and the object level data 7 do not have to be synchronized. Previously sampled object level data 7 may be used, for example, together with current image data.

[0035] The neural network system 4 can be based on a convolutional neural network (CNN) or on a recurrent neural network (RNN). End-to-end training of the neural network 4 is performed before using the neural network 4 according to the proposed method and should be performed end-to-end in a supervised manner, where the neural network 4 starts observing expert driving behavior for a very large number of possible traffic scenarios and / or is performed end-to-end with the help of reinforcement learning in a simulation environment.

[0036] In the supervised case, using a very large and diverse data set will enable the model to become as general as possible across the range of operations. In the reinforcement setting, the neural network system 4 will be trained in a simulation environment where the neural network system 4 will attempt to improve good behavior and suppress bad behavior. Doing so in simulation will allow the training of the neural network system 4 to be more exploratory without imposing risks on people or expensive hardware.

[0037] Convolutional neural networks are particularly suitable for use with the proposed method because features can be automatically learned from training examples, e.g., large labeled datasets can be used for training and validation. The training data may, for example, have been previously collected by vehicles driving on a wide variety of roads and in a diverse set of lighting and weather conditions. The CNN learning algorithm can be implemented on massively parallel graphics processing units (GPUs) to accelerate learning and inference.

[0038] The method further comprises performing, by an end-to-end trained neural network system 4, a mapping 17 of input data, such as raw sensor data 5 from onboard sensors 6 of the autonomous road vehicle 3, as well as object level data 7 and policy information data 8, to control commands 10 required for controlling the travel of the autonomous road vehicle 3 on a planned trajectory generated by the control commands 10 for the autonomous road vehicle 3 within a pre-set time frame. Therefore, it is suggested to use a trained neural network 4, such as a convolutional neural network (CNN), to directly map the input data 5, 7, 8 to the control commands 10.

[0039] The output of the end-to-end trained neural network system 4 consists of an estimate of the control command 10 within a time horizon. The time horizon should be long enough to enable prediction of the system behavior, for example 1 second. In this way, the neural network system 4 learns to plan for the long term. The output can also be fed back into the model of the neural network system 4 to ensure that the model takes into account smooth driving commands of previous decisions. The method can, for example, include automatically setting the control command for speed to the electronic horizon recommendation if no obstacles obscure the expected trajectory and if the road conditions are suitable. Otherwise, it can include estimating a suitable speed.

[0040] In combination with the lateral control command, the method should preferably provide for executing a planned lane change, e.g. changing to the appropriate lane for turning in an intersection at an early stage. Therefore, in such cases, one of the outputs of the system should be a lane change signal 13. This lane change signal 13 should be fed back into the neural network system 4 in a recursive manner to allow smooth overtaking etc.

[0041] Furthermore, according to the proposed method, the control commands 10 for the autonomous road vehicle 3 are subjected 18 to the safety module 9 within a predefined time frame. A risk assessment 19 of the planned trajectory resulting from the control commands 10 for the autonomous road vehicle is performed by the safety module 9 within a predefined time frame. Figure 1 As illustrated by the Y option in FIG. 1 , the control commands 2 to be used to control the travel of the autonomous road vehicle 3 are verified as safe and then output 20 from the safety module 9 to control the travel of the autonomous road vehicle 3. These verified control commands 2 may, for example, be output to the vehicle control module 15 of the autonomous road vehicle 3. Option X is used to schematically illustrate the rejection of control commands 10 which the safety module 9 could not verify as safe.

[0042] To enhance planning capabilities from predicted system behavior, some embodiments of the method include adding a machine learning component 11 to the end-to-end trained neural network system 4. Such a machine learning component 11 adds to the method the ability to "learn", i.e., gradually improve performance from new data without explicit programming. Another alternative for this is through more advanced methods, such as value iteration networks.

[0043] In some further embodiments, the method further comprises providing the verified control commands 2 for the autonomous road vehicle 3 verified as safe by the safety module 9 as feedback 12 to the end-to-end trained neural network system 4. In this way, it becomes possible to further train the neural network system 4 to avoid an infeasible or unsafe series of control commands. In such embodiments, the safety module 9 must be available for system training.

[0044] Thus, a method is provided which can be used as an end-to-end solution for overall decision making for an autonomous road vehicle 3 in a complex traffic environment. The method operates on rich sensor information 5, 7, 8 to allow appropriate decisions to be made and planning for what control commands 2 to take in a future time horizon.

[0045] The proposed apparatus 1 for generating validated control commands 2 for an autonomous road vehicle 3 comprises an end-to-end trained neural network system 4 arranged to receive input of raw sensor data 5 from onboard sensors 6 of the autonomous road vehicle 3 as well as object level data 7 and strategy information data 8.

[0046] The end-to-end trained neural network system 4 of the proposed apparatus 1 for generating validated control commands for an autonomous road vehicle 3 may also be arranged to receive as raw sensor data 5 at least one of the following: image data; speed data and acceleration data from one or more on-board sensors 6 of the autonomous road vehicle 3. The raw sensor data 5 may be, for example, images from a vision system of the autonomous road vehicle 3 and information about its speed and acceleration, for example obtained by tapping into a controller area network (CAN) or similar network of the autonomous road vehicle. The image data may include images from a surround view camera and also consist of several recent images.

[0047] The end-to-end trained neural network system 4 may further be arranged to receive at least one of the following as object level data 7: positions of surrounding objects; lane markings and road conditions.

[0048] Furthermore, the end-to-end trained neural network system 4 may be arranged to receive at least one of the following as strategy information data 8: electronic horizon (map) information including current traffic rules and road geometry, and high-level navigation information. The electronic horizon (map) information about current traffic rules and road geometry enables the end-to-end trained neural network system 4 to take into account things such as permissible speeds, motorway exits, roundabouts, and distances to intersections when making decisions.

[0049] In order to guide the end-to-end trained neural network system 4 in situations where a lane change, freeway exit or similar action is required, the system may be arranged to include high-level navigation information as an input to the end-to-end trained neural network system 4. In this way, the end-to-end trained neural network system 4 will be able to take into account user preferences, such as driving to a user-specified location.

[0050] The end-to-end trained neural network system 4 should be based on a convolutional neural network (CNN) or on a recurrent neural network (RNN). The end-to-end training of the neural network 4 should be carried out before using the neural network 4 in the proposed device 1 and should have been performed end-to-end in a supervised manner, wherein the neural network 4 has been allowed to observe expert driving behavior for a very large number of possible traffic scenarios and / or has been performed end-to-end with the help of reinforcement learning in a simulation environment.

[0051] In the supervised setting, using very large and diverse data sets will allow the model to become as general as possible across the range of operations. In the reinforcement setting, the system will be trained in a simulation environment where the system will try to improve good behaviors and suppress bad behaviors. Doing so in simulation will allow the training of the system to be more exploratory without putting risks on people or expensive hardware.

[0052] Convolutional neural networks are particularly suitable for use with the proposed apparatus because they enable automatic learning of features from training examples, e.g., by using large labeled datasets for their training and validation. The training data may, for example, have been previously collected by vehicles driving on a wide variety of roads and in a diverse set of lighting and weather conditions. Training may alternatively be based on a predefined set of categories, e.g., overtaking, roundabouts, etc., but also on "generic" categories that capture all situations that are not explicitly defined. The CNN learning algorithm may be implemented on massively parallel graphics processing units (GPUs) to accelerate learning and inference.

[0053] The end-to-end trained neural network system 4 should also be arranged to map input data, such as raw sensor data 5 from onboard sensors 6 of the autonomous road vehicle 3, as well as object level data 7 and policy information data 8, to control commands 10 required for controlling the travel of the autonomous road vehicle 3 on the planned trajectory generated by the control commands 10 for the autonomous road vehicle 3 within a pre-set time frame. Therefore, it is recommended to use a trained neural network 4, such as a convolutional neural network (CNN), to map the input data 5, 7, 8 directly to the control commands 10.

[0054] The output of the end-to-end trained neural network system 4 consists of an estimate of the control command 10 within a time horizon. The time horizon should be long enough to be able to predict the system behavior, for example 1 second. In this way, the neural network system 4 learns to plan for the long term. The output 10 can also be fed back into the model of the neural network system 4 to ensure that the model takes into account smooth driving commands of previous decisions. The control command for speed can be automatically set by the neural network system 4 to the electronic horizon recommendation, for example, if there are no obstacles obscuring the expected trajectory and if the road conditions are suitable. Otherwise, a suitable speed can be estimated by the neural network system 4.

[0055] In combination with the lateral control commands, the neural network system 4 should preferably be able to execute planned lane changes, e.g. changing to the appropriate lane for turning at an intersection at an early stage. Therefore, one of the outputs of the system should be a lane change signal 13. This lane change signal 13 should be fed back into the system in a recursive manner to allow smooth overtaking etc. Figure 2 , a lane change signal 13 is depicted in two settings. The first includes a feedback controller 14 and the second does not. An optional feedback controller 14 is provided as a link between the driver or safety monitor and the system. The link can also be used as an ADAS feature, where the driver can trigger a lane change by, for example, engaging a turn signal. The optional feedback controller 14 can also be arranged to provide feedback based on other strategy information 8 in addition to lane change information.

[0056] The safety module 9 is arranged to receive control commands 10 for the autonomous road vehicle 3 within a pre-set time frame and to perform a risk assessment of the planned trajectory resulting from the control commands 10 for the autonomous road vehicle 3 within the pre-set time frame. The safety module 9 is also arranged to verify control commands 2 to be used to control the travel of the autonomous road vehicle 3 as safe and then output such verified control commands 2 to control the travel of the autonomous road vehicle 3. These verified control commands 2 may, for example, be output to the vehicle control module 15. The safety module 9 may also be arranged to receive raw sensor data 5 and object level data 7.

[0057] To enhance planning capabilities from predicted system behavior, the end-to-end trained neural network system 4 of the device 1 may further include a machine learning component 11. Such a machine learning component 11 adds to the device 1 the ability to "learn", i.e., gradually improve performance from new data without explicit programming. Another alternative for this is through more advanced methods, such as value iteration networks.

[0058] The device 1 may also be arranged to feed back 12 verified control commands 2 for the autonomous road vehicle 3 verified as safe by the safety module 9 to the end-to-end trained neural network system 4. This makes it possible to further train the neural network system 4 to avoid an infeasible or unsafe series of control commands. In such embodiments, the safety module 9 must be available for system training.

[0059] Thus, a device 1 is thereby provided which can be used as an end-to-end solution for overall decision making for an autonomous road vehicle 3 in complex traffic environments and in any driving situation. Its neural network system 4 is arranged to operate on the rich sensor information 5, 7, 8 to allow for appropriate decisions and planning of what control commands 2 to take in a future time horizon. The neural network system 4 of the device 1 may consist of only one learned neural network, which is used to generate verified control commands 2 for performing driving of an autonomous road vehicle 3.

[0060] Furthermore, it is envisaged here that an autonomous road vehicle 3 comprises a device 1 as set out herein.

Claims

1. A method for generating verified control commands (2) for an autonomous road vehicle (3), characterized in that The method comprises: providing (16) raw sensor data (5) from onboard sensors (6) of the autonomous road vehicle (3) as well as object level data (7) and policy information data (8) as input data to an end-to-end trained neural network system (4); mapping (17) input data (5, 7, 8) to an estimate of a control command (10) for the autonomous road vehicle (3) by the end-to-end trained neural network system (4) within a preset time frame, and outputting the estimate of the control command (10) from the end-to-end trained neural network system (4) within a preset time frame; the estimation of the control command (10) for the autonomous road vehicle (3) within the pre-set time range being subject (18) to a safety module (9), the safety module (9) being arranged to perform a risk assessment of a planned trajectory resulting from the estimation of the control command (10) for the autonomous road vehicle (3) within the pre-set time range, the safety module (9) also receiving as input raw sensor data (5) and object level data (7); verifying (19) that it is safe and outputting from the safety module (9) a verified control command (2) for the autonomous road vehicle (3); The method (1) further comprises providing at least one of the following as strategic information data (8): electronic horizon information including current traffic regulations and road geometry, and advanced navigation information.

2. The method (1) according to claim 1, wherein the method (1) further comprises adding a machine learning component (11) to the end-to-end trained neural network system (4).

3. The method (1) according to claim 1 or 2, wherein the method (1) further comprises providing verified control commands (2) for the autonomous road vehicle (3) verified as safe by the safety module (9) as feedback (12) to the end-to-end trained neural network system (4).

4. A method (1) according to claim 1 or 2, wherein the method (1) also includes providing at least one of the following items as raw sensor data (5): image data; speed data and acceleration data from one or more on-board sensors (6) of the autonomous road vehicle (3).

5. The method (1) according to claim 1 or 2, wherein the method (1) further comprises providing at least one of the following as object level data (7): positions of surrounding objects; lane markings and road conditions.

6. A device (1) for generating a verified control command (2) for an autonomous road vehicle (3), It is characterized in that The device (1) comprises: an end-to-end trained neural network system (4) arranged to receive as input data raw sensor data (5) from onboard sensors (6) of the autonomous road vehicle (3) as well as object level data (7) and policy information data (8); The end-to-end trained neural network system (4) is further arranged to map input data (5, 7, 8) to an estimate of a control command (10) for the autonomous road vehicle (3) within a pre-set time frame and to output the estimate of the control command (10) from the end-to-end trained neural network system (4) within a pre-set time frame; a safety module (9) arranged to receive an estimate of the control command (10) for the autonomous road vehicle (3) within the pre-set time range and to perform a risk assessment of a planned trajectory resulting from the control command (10) for the autonomous road vehicle (3) within the pre-set time range, the safety module (9) being further arranged to receive as input raw sensor data (5) and object level data (7); The safety module (9) is further arranged to verify as safe and output a verified control command (2) for the autonomous road vehicle (3); The device (1) further comprises the end-to-end trained neural network system (4) arranged to receive at least one of the following as strategy information data (8): electronic horizon information including current traffic regulations and road geometry, and high-level navigation information.

7. The device (1) for generating verified control commands (2) for an autonomous road vehicle (3) according to claim 6, wherein the device (1) also includes the end-to-end trained neural network system (4) further including a machine learning component (11).

8. A device (1) for generating verified control commands (2) for an autonomous road vehicle (3) according to claim 6 or 7, wherein the device (1) is also arranged to feed back (12) the verified control commands (2) for the autonomous road vehicle (3) verified as safe by the safety module (9) to the end-to-end trained neural network system (4).

9. A device (1) for generating verified control commands (2) for an autonomous road vehicle (3) according to claim 6 or 7, wherein the device (1) also includes the end-to-end trained neural network system (4) arranged to receive at least one of the following items as raw sensor data (5): image data; speed data and acceleration data from one or more on-board sensors (6) of the autonomous road vehicle (3).

10. A device (1) for generating verified control commands (2) for an autonomous road vehicle (3) according to claim 6 or 7, wherein the device (1) also includes the end-to-end trained neural network system (4) arranged to receive at least one of the following items as object-level data (7): positions of surrounding objects; lane markings and road conditions.

11. An autonomous road vehicle (3), characterized in that The autonomous road vehicle (3) comprises a device (1) for generating a validated control command (2) according to any one of claims 6 to 10.

Images