Method for generating a training data set for training an artificial intelligence module for a control device of a vehicle

1. A method for generating a training data set for training an artificial intelligence (AI) module, comprising the following steps:

providing an image sequence in which surroundings of a robot are recorded;

determining at least one trajectory which is situatable in the recorded surroundings of the robot;

generating at least one future image sequence which extends to a time segment in the future with respect to a sequence ending point in time, and, based on the at least one determined trajectory, encompasses a prediction of images for a future event that the determined trajectory is followed during the time segment in the future;

assessing at least one sub-section of the determined trajectory included in the generated image sequence as positive when a movement predicted by following the determined trajectory corresponds to a valid movement situation, or as negative when the movement predicted by following the determined trajectory corresponds to an invalid movement situation, wherein the valid movement situation encompasses a collision-avoiding and/or road-following continued movement along the determined trajectory, and the invalid driving situation encompasses a veering off a roadway and/or a departure from a lane and/or a collision with another object; and

combining the generated future image sequence with the assessment assigned to the determined trajectory for generating a training data set for the AI module;

training the AI module using the generated training data set.

2. The method as recited in claim 1 , wherein the training of the AI module includes the following step:

feeding the generated training data set into the AI module.

3. The method as recited in claim 1 , wherein only a single image sequence is provided, and multiple trajectories which each differ from one another are generated from which a multitude of future image sequences is generated.

4. The method as recited in claim 1 , wherein the generated image sequence for the determined trajectory encompasses a number of: (i) depth images, and/or (ii) real images and/or (iii) images of a semantic segmentation, along the same trajectory.

5. The method as recited in claim 1 , wherein the trajectory is determined for a dynamic object included in the provided image sequence and, based the determined trajectory, the future image sequence is generated.

6. The method as recited in claim 1 , wherein the trajectory is determined for the robot and, based on the determined trajectory, the future image sequence is generated.

7. The method as recited in claim 1 , wherein, prior to the determination, a preselection of the trajectory situatable in the surroundings is made, based on a predetermined probability distribution.

8. The method as recited in claim 1 , wherein the determination is only made for the one trajectory or multiple trajectories which are implementable based on a driving situation, and is based on an assigned vehicle dynamics model of the robot configured as a vehicle.

9. The method as recited in claim 1 , wherein the time segment is established with a duration between 0.5 s and 1.5 s.

10. The method as recited in claim 1 , wherein the time segment is established with a duration of 1 s.

11. The method as recited in claim 1 , wherein the prediction includes at least one or multiple of the following methods: monocular depth estimation, stereo depth estimation, LIDAR data processing, and estimation from optical flow.

12. The method as recited in claim 1 , wherein the prediction includes generation of a semantic segmentation of at least several individual images of the image sequence.

13. The method as recited in claim 12 , wherein, during the assessment, an object recognition and/or a feature recognition obtained from the semantic segmentation is used to weight the positive or negative assessment.

14. The method as recited in claim 1 further comprising:

generating signals, based the trained AI module, for controlling a robot; and

controlling the robot using the generated signals, including activating an actuator of the robot.

15. The method as recited in claim 14 , wherein the actuator is a drive of the robot or a steering system of the robot.

16. A data processing unit for training an artificial intelligence (AI) module, which is configured to:

provide an image sequence in which surroundings of a robot are recorded;

determine at least one trajectory which is situatable in the recorded surroundings of the robot;

generate at least one future image sequence which extends to a time segment in the future with respect to a sequence ending point in time, and, based on the at least one determined trajectory, encompasses a prediction of images for a future event that the determined trajectory is followed during the time segment in the future;

assess at least one sub-section of the determined trajectory included in the generated image sequence as positive when a movement predicted by following the determined trajectory corresponds to a valid movement situation, or as negative when the movement predicted by following the determined trajectory corresponds to an invalid movement situation, wherein the valid movement situation encompasses a collision-avoiding and/or road-following continued movement along the determined trajectory, and the invalid driving situation encompasses a veering off a roadway and/or a departure from a lane and/or a collision with another object; and

combine the generated future image sequence with the assessment assigned to the determined trajectory for generating a training data set for the AI module;

train the AI module using the generated training data set.

17. A device for controlling an at least semi-autonomous robot, the device being configured to:

provide an image sequence in which surroundings of a robot are recorded;

determine at least one trajectory which is situatable in the recorded surroundings of the robot;

generate at least one future image sequence which extends to a time segment in the future with respect to a sequence ending point in time, and, based on the at least one determined trajectory, encompasses a prediction of images for a future event that the determined trajectory is followed during the time segment in the future;

assess at least one sub-section of the determined trajectory included in the generated image sequence as positive when a movement predicted by following the determined trajectory corresponds to a valid movement situation, or as negative when the movement predicted by following the determined trajectory corresponds to an invalid movement situation, wherein the valid movement situation encompasses a collision-avoiding and/or road-following continued movement along the determined trajectory, and the invalid driving situation encompasses a veering off a roadway and/or a departure from a lane and/or a collision with another object;

combine the generated future image sequence with the assessment assigned to the determined trajectory for generating a training data set for an artificial intelligence (AI) module;

train the AI module using the generated training data set;

select an assessed trajectory using the trained AI module; and

activate the robot according to the selected trajectory.

18. The data processing unit according to claim 16 , where the data processing unit is further configured to:

generate signals, based the trained AI module, for controlling a robot; and

control the robot using the generated signals, including activating an actuator of the robot.

19. The data processing unit according to claim 18 , wherein the actuator is a drive of the robot or a steering system of the robot.