Method for operating at least an assisted parking function for a vehicle

Abstract

The invention relates to a method for operating an at least assisted parking function (3) for a vehicle (1), comprising: providing sensor information (10) describing a parking space (6); and determining (S2) a parking trajectory (8) from a current location of the vehicle (1) to a parking location (9) in the parking space (6) by applying a trajectory planning algorithm (11) to the provided sensor information (10). The method may include at least assisted operation (S4) of the vehicle (1) to move it along the determined parking trajectory (8). The trajectory planning algorithm (11) comprises an artificial neural network trained by reinforcement learning using a curriculum (23).

Description

[0001] The invention relates to a method for operating at least an assisted parking function for a vehicle. The invention also relates to a vehicle, a control unit for a vehicle, and a computer program product for implementing such a method.

[0002] The vehicle may be equipped with a parking function for at least assisted parking into a parking space. The vehicle may detect a parking space in the vehicle's surroundings and determine a parking trajectory from a current location of the vehicle to a parking location in the parking space. An algorithm used to determine the parking trajectory is typically based on a set of rules. These rules define, for example, a permitted steering angle range and / or speed range for a parking maneuver and / or a minimum distance toward an obstacle in the surroundings. Based on the set of rules, which may include hundreds or thousands of individual rules, a control unit of the vehicle can calculate the parking trajectory. However, the set of rules may include overlapping or intertwined rules, making the algorithm difficult to maintain.In addition, an aggregation of specific rules may lack generality. Furthermore, the rules for some parking functions may be over-adapted. Furthermore, the individual rules were formulated by humans, so they may be cumbersome and / or biased. Therefore, the vehicle's parking function algorithm should be optimized to make it less dependent on human input.

[0003] DE 10 2021 206 413 A1 discloses a method for determining a currently drivable trajectory for an at least partially assisted motor vehicle. A reference trajectory is defined, and several other potentially drivable trajectories are also determined.

[0004] DE 10 2020 127 216 A1 discloses a method for determining a trajectory for a vehicle. A neural network is used to determine at least one anchor point for a trajectory for the vehicle, connecting a starting position of the vehicle with a target position for the vehicle. A training phase of the neural network involves the automatic generation of a plurality of different training scenarios as a function of a number of predefined parameters.

[0005] The object of the invention is to provide a reliable parking function.

[0006] The independent claims solve the problem.

[0007] A first aspect of the invention relates to a method for operating an at least assisted parking function for a vehicle. The at least assisted parking function is configured to provide assisted parking and thus assisted driving of the vehicle into a parking space. In addition, the parking function can also be configured to provide semi-automatic or, in particular, fully automatic parking. Several levels of automation are covered by the method, starting with assisted parking. To carry out the method, the vehicle can include a control unit configured to operate a drive system, a braking system, and / or a steering system of the vehicle.

[0008] The method includes providing or receiving sensor information. The sensor information describes a parking space. The sensor information describes the parking space by at least one line, at least one line sequence, and / or at least one point. The sensor information can describe an environment of the vehicle. The parking space is, for example, a free area in the environment of the vehicle. The area is available for maneuvering. The sensor information also describes at least part of the area available for maneuvering. Therefore, it can describe the environment of the vehicle including the parking space. The sensor information can describe at least one obstacle and / or parking space markings in the environment of the vehicle. The vehicle typically approaches the parking space, for example, with a forward driving maneuver or with a reverse driving maneuver.While approaching the parking space, a sensor device of the vehicle, such as a camera, a lidar device, and / or a radar device, can capture the sensor information. Preferably, the sensor information is camera data provided by at least one camera of the vehicle. The camera can be a front camera, a rear camera, a side camera, and / or a surround-view camera of the vehicle. The vehicle can receive the sensor information from an external device, such as another vehicle, a backend, a server, and / or a cloud server. The external device can transmit the sensor information to the vehicle via vehicle-to-vehicle and / or vehicle-to-infrastructure communication.

[0009] The method includes determining a parking trajectory from a current location of the vehicle to a parking location in the parking space by applying a trajectory planning algorithm to the provided sensor information. The trajectory planning algorithm is software or a part of software configured to calculate the parking trajectory based on information about the parking space. Information about the parking space is provided here by the provided sensor information. The at least one line and / or the at least one line string of the sensor information can describe the parking space, the current location, and / or at least one possible parking location using coordinates and an orientation angle. The inputs are therefore the at least one line, the at least one line string, and / or the at least one point described by the sensor information.In addition, the sensor information may include a position of the at least one point, a contour of the at least one line and / or a motion vector in the case of a dynamic object in the environment.

[0010] The vehicle's control unit can operate the parking planning algorithm to determine the parking trajectory for the vehicle. The current location can be described by a position and orientation of the vehicle when the method starts. In particular, it describes the position of a center of a rear axle of the vehicle. The current location is therefore defined, for example, by coordinates and an orientation angle of the vehicle at these coordinates. The current location can be referred to as the starting position, in particular as the current position, of the vehicle. The current location is preferably near the parking space, for example on a street adjacent to the parking space. From the current location, the parking space is preferably within a coverage area of ​​the sensor device that provides the sensor information. The parking location is a final position and orientation of the vehicle in the parking space.The vehicle should be at the parking location when the parking is complete. At the parking location, the vehicle's engine is typically shut down and a parking brake is activated, allowing, for example, a driver or other occupant of the vehicle to exit the vehicle at the parking location. The parking location can be referred to as the destination location in the parking lot. The determined parking trajectory can be a set of points that define a path the vehicle can follow to travel from the current location to the parking location.

[0011] Possible outputs of the parking planning algorithm can be a final and / or intermediate position on the parking trajectory. The final position can be the parking location. Furthermore, the parking trajectory can be described by at least one polynomial that describes a path to be traveled by the vehicle. The parking trajectory and therefore the output can alternatively or additionally describe at least one geometric curve, in particular a concatenation of several straight, circular, and / or clothoid segments of the trajectory. The method is preferably carried out by the control unit of the vehicle, i.e., for example, an electronic control unit of the vehicle. Therefore, the method is a computer-implemented method.

[0012] The trajectory planning algorithm includes an artificial neural network. The artificial neural network can be a feed-forward neural network. The structure of the artificial neural network can include four layers with 256 neurons. It can have an input size of 43 and an output size of 15 neurons (seven neurons for backward trajectories, seven neurons for forward trajectories, and one neuron to decide which of the two trajectories to choose). The inputs are not given as a map, but rather as a vector of object coordinates, such as the position of a wall in the environment, the orientation of a parking space, and so on. There can be an upstream machine that processes the sensor information to obtain the perception and fuses this perception to obtain the vectorized representation described.

[0013] The trajectory planning algorithm can determine multiple parking trajectories at a time. Preferably, it determines two parking trajectories: one for forward parking and the other for reverse parking. Taking into account, for example, a number of direction changes, a length of the respective parking trajectory, and / or another predetermined condition, one of the two determined parking trajectories is selected. Preferably, the selected trajectory is the one with the fewest direction changes and / or the shortest length.

[0014] The artificial neural network was trained using reinforcement learning with a curriculum. In other words, the trajectory planning algorithm is based on machine learning. The training of the trajectory planning algorithm is assumed to be complete when the procedure takes place. The training of the artificial neural network uses reinforcement learning. Reinforcement learning is a machine learning technique that focuses on training an algorithm using an empirical method and, consequently, through trial and error. The algorithm, typically referred to as an agent, evaluates a current situation, which can be referred to as a state. The algorithm also takes an action and receives feedback from an environment after each action. The feedback is typically referred to as a reward if it is positive feedback.If the feedback is negative, it can be considered a punishment for making a mistake. The environment can be considered a computer game and / or multiple simulations that mimic various artificially generated parking scenarios. Based on trial and error, the algorithm learns how to solve the computer game or simulations. Through this, the algorithm learns how to solve a parking problem and, therefore, how to determine a parking trajectory for a random parking situation. More precisely, the reward is the ultimate goal that the agent learns while interacting with the environment through numerous trials and errors. The algorithm typically receives short-term rewards that, when combined, lead to the cumulative long-term reward.This means that the artificial neural network learns, using reinforcement learning, how to define the best sequence of decisions that will allow the agent to solve the problem while maximizing long-term rewards. For example, determining a parking trajectory in the environment is considered successful if the vehicle is parked without colliding with any obstacle. A reward would then be given. In a similar scenario in the future, a parking trajectory similar to the one that received the reward is more likely to be chosen compared to other possible trajectories that might include, for example, a potential collision with the obstacle.

[0015] Reinforcement learning is one of three learning styles of artificial neural networks, along with supervised learning and unsupervised learning. Consequently, the artificial neural network used can be based on known techniques.

[0016] Furthermore, reinforcement learning with a curriculum is applied. This means that curriculum learning is used to train the artificial neural network. Curriculum learning means that training takes place in a meaningful order. Training can therefore start with an easy simulation, then increase the difficulty of the simulations, and end with a difficult simulation. This means, for example, that a parking scenario involving only one obstacle is first trained, and later, after the agent has mastered this relatively easy parking scenario, the number of obstacles is increased to simulate more complex parking scenarios. Thanks to curriculum learning, the artificial neural network can learn how to determine the parking trajectory in a relatively short time.The total time required to train the artificial neural network of the trajectory planning algorithm is therefore minimized compared to training without curriculum learning. Therefore, it is particularly reasonable to use reinforcement learning with the curriculum to generate the trajectory planning algorithm.

[0017] Furthermore, the described trajectory planning algorithm is free of human input because it is not based on a set of rules but was formulated by humans. In contrast, the trajectory planning algorithm is completely artificially generated. In other words, the artificial neural network is trained only through simulations without human input. Therefore, no training data and / or supervised training are required to provide a reliable trajectory planning algorithm. The provided parking function is therefore reliable and independent of human input.

[0018] According to one embodiment, the method includes at least assisted operation of the vehicle to move it along the determined parking trajectory. In the case of assisted parking, the control unit transmits, for example, commands for manually driving, steering, and / or braking the vehicle, which are to be displayed by a display device in the vehicle to instruct the driver to follow the parking trajectory. However, the parking function preferably operates the vehicle along the parking trajectory semi-automatically or, in particular, fully automatically. In particular, the parking function controls lateral and / or longitudinal guidance of the vehicle at least semi-automatically. The lateral guidance includes at least semi-automatic steering of the vehicle. The longitudinal guidance includes at least semi-automatic acceleration and / or braking of the vehicle. The lateral and / or longitudinal guidance is preferably fully automatic.

[0019] According to another embodiment, reinforcement learning with the curriculum includes the continuous generation of multiple parking scenarios. These multiple parking scenarios sequentially increase the difficulty of a parking environment. Alternatively or additionally, the multiple parking scenarios sequentially decrease the difficulty within a given tolerance range of the parking location in the parking scenario. The complexity of the parking scenarios, and therefore the levels of the computer game or simulation, continuously increases from one parking scenario to the next. A first parking scenario may include a wide parking lot, so that the parking location can be located in a relatively large area. In this example, the tolerance range of the parking location is therefore high. Furthermore, the first parking scenario only includes one or a few objects in the surroundings of the parking lot.The number of possible obstacles is therefore small, so the difficulty of the parking environment is relatively low, for example, because multiple parking trajectories are possible. However, if the artificial neural network can solve the first parking scenario, training continues with a second parking scenario that differs from the first. The second parking scenario may involve a narrower parking space and / or more obstacles than the first. Analogously, several parking scenarios are presented to the artificial neural network one after the other until a final parking scenario with the highest complexity is presented and solved. The described reinforcement learning with the curriculum has a positive effect on training time, so that the trajectory planning algorithm is trained in a shorter amount of time compared to, for example, reinforcement learning without the curriculum.

[0020] To automatically generate the parking scenarios for training, at least one specification may be given. The specification may be a definition of an obstacle, a parking location, a maximum steering angle of a vehicle, a current location, and / or at least one rule or regulation for driving and / or parking. Such a rule or regulation may include that a vehicle cannot penetrate an obstacle and / or that there are tolerance distances towards obstacles and / or parking space boundaries that the vehicle cannot cross.

[0021] According to a further embodiment, the method includes checking whether the determined parking trajectory is free of obstacles. This is achieved by applying a collision detection algorithm to the determined parking trajectory and the provided sensor information. The collision detection algorithm includes applying at least one rule to determine whether the vehicle can follow the determined parking trajectory without colliding with an obstacle or whether a collision with the obstacle exists when the vehicle follows the parking trajectory. The parking trajectory can, for example, be compared with an environment map generated based on the sensor information. The environment map includes objects in the environment of the vehicle that could be obstacles when the vehicle travels along the parking trajectory. Taking the environment map into account, possible collisions can be predicted.

[0022] Only when the parking trajectory is clear of obstacles will the vehicle be operated to move along the parking trajectory. This means that the algorithm or method does not simply trust the trajectory planning algorithm, but first checks whether the specific parking trajectory is trustworthy. This is a precautionary measure to ensure that the trajectory planning algorithm has been sufficiently and accurately trained. For example, if operation according to the specific parking trajectory is stopped due to a potential collision with an obstacle, retraining of the trajectory planning algorithm can be automatically activated. Retraining can be based on additional parking scenarios, allowing, for example, additional levels to be added to the curriculum.

[0023] Another embodiment involves determining a repositioning trajectory along which the vehicle is movable from the current location to a starting location. However, the repositioning trajectory is only determined if the determined parking trajectory is not free of obstacles. The method then involves reapplying the trajectory planning algorithm to the starting location as the current location of the vehicle in order to determine a new parking trajectory. This means that the method as described above, in particular the application of the trajectory planning algorithm, is repeated, but taking into account a different current location. This different current location is the starting location. Based on the new location of the vehicle, which is the target of the repositioning trajectory, a different parking trajectory is determined that differs from the previously determined parking trajectory.The repositioning trajectory is a trajectory and therefore a path from the current location to another location near the current location, referred to as the starting location. Typically, the repositioning trajectory is a relatively short trajectory, for example, a few centimeters or meters. The positioning trajectory can lead to a change in the vehicle's orientation with respect to the parking space. Therefore, the process is repeated to provide a parking trajectory, even if this has failed so far. This means that a solution for parking into the parking space is provided if the previously determined parking trajectory is not suitable for the environment and the parking space.

[0024] According to a further embodiment, the method includes applying a rule-based algorithm to the provided sensor information to determine the repositioning trajectory. This means that the repositioning trajectory is not determined, for example, based on reinforcement learning, but can be based on one or more rules. The rule-based algorithm can, for example, include a maximum distance and / or a maximum change in orientation of the vehicle that is permitted for repositioning the vehicle. In addition, the rule-based algorithm can take obstacles and / or the parking space into account. The rule-based algorithm can be trained based on supervised and / or unsupervised learning.The rule-based algorithm does not include an artificial neural network trained through reinforcement learning, as determining the repositioning trajectory is considered less complex than determining the parking trajectory, so the rule-based algorithm is sufficient to determine the repositioning trajectory. This can save computing power.

[0025] Another embodiment includes checking whether the determined new parking trajectory is free of obstacles by applying the collision detection algorithm to the determined new parking trajectory and the provided sensor information. Only if this is the case does the method include at least assisted operation of the vehicle to move it along the new parking trajectory. This means that the new parking trajectory determined after determining the repositioning trajectory is not immediately trusted, but must first be checked for possible collisions with obstacles. This further increases the reliability and robustness of the method.

[0026] A preferred embodiment involves operating the vehicle to move it along the determined repositioning trajectory to the starting location before reapplying the trajectory planning algorithm. This means that the repositioning trajectory is determined and immediately thereafter the vehicle is moved or is moving from the current location to the starting location. When the vehicle reaches the starting location, the new parking trajectory is determined and checked, for example, for possible collisions with obstacles. Alternatively, after checking that the determined new parking trajectory is free of obstacles, the vehicle is moved to the starting location. In this case, the method can first simulate or assume the repositioning of the vehicle, and only when the new parking trajectory is selected for operating the vehicle does the vehicle leave the current location, travel to the starting location, and from there follow the new parking trajectory.This means that the new parking trajectory can be determined before the vehicle is even moved. Consequently, unnecessary repositioning of the vehicle can be avoided if the new parking trajectory does not at least meet the collision avoidance requirement.

[0027] Another embodiment involves converting an output of the trajectory planning algorithm into a steering angle curve along the parking trajectory. It also involves providing the steering angle curve to the parking function. This means determining a graph showing the steering angle as a function of a distance from the current location. The distance can be considered a position on the parking trajectory. The steering angle curve allows for a more precise definition of the parking trajectory compared to, for example, a parking trajectory based on Bezier functions. It solves the disadvantage of Bezier functions, which is that the set of possible curves is limited and / or some curves are unacceptable due to a given maximum curvature. However, Bezier functions are typically easy for a reader or user to understand.The method described here for describing the parking trajectory involves predicting curvature profiles for the vehicle that include multiple segments, particularly three segments. Therefore, compared to using Bezier functions, a larger set of curves is possible, and all types of curvatures are acceptable. The steering angle profile is easily predicted using artificial intelligence, such as the artificial neural network used here. Therefore, the steering angle profile is advantageous.

[0028] Another embodiment involves the determined parking trajectory including multiple trajectory sections. Each time the trajectory planning algorithm is applied, one of the trajectory sections is determined, so that the trajectory planning algorithm is applied multiple times to determine the entire parking trajectory. The parking trajectory can, for example, be divided into multiple sections so that the trajectory planning algorithm can determine an optimal path for each step of the parking maneuver. If a parking maneuver that includes at least one change of direction can be divided into two trajectory sections, each of which is individually determined by applying the trajectory planning algorithm, this allows each section of the parking trajectory to be better predicted and planned. In summary, an optimized parking trajectory is provided that includes individually optimized trajectory sections.

[0029] However, it is alternatively or additionally possible to determine the entire parking trajectory by applying the trajectory planning algorithm.

[0030] A preferred embodiment includes the trajectory planning algorithm determining the parking trajectory with the fewest changes in direction compared to other possible parking trajectories. Alternatively or additionally, the trajectory planning algorithm determines the parking trajectory with the shortest trajectory length compared to other possible parking trajectories. This means that preferably only those parking trajectories are output by the artificial neural network that have few changes in direction and therefore few steps and / or that have a minimal total trajectory length. For example, if the trajectory planning algorithm determines a second parking trajectory that has a longer length and / or an additional change in direction compared to a first parking trajectory, the first parking trajectory is selected as the output of the trajectory planning algorithm.This makes it possible to always provide the best or optimal possible parking trajectory and discard less advantageous parking trajectories. The method is therefore particularly effective.

[0031] According to a further embodiment, the method includes determining a further parking trajectory by applying a rule-based trajectory planning algorithm to the sensor information and selecting the parking trajectory with the fewest changes in direction and / or the shortest trajectory length from the determined parking trajectory and the determined further parking trajectory. The method may thus include a general parallel method. This means that the rule-based trajectory planning algorithm and the trajectory planning algorithm are both applied, in particular simultaneously. Operating both algorithms in parallel may only be possible if sufficient computing power is available. The method then includes choosing between the determined parking trajectory and the determined further parking trajectory. It thereby selects the parking trajectory with the fewest changes in direction and / or the shortest length.

[0032] Alternatively or additionally, the trajectory planning algorithm may be applied first. If the determined parking trajectory has a complexity level that lies outside a predetermined complexity level range, the method may include applying the trajectory planning algorithm on a rule-based basis. The method may then automatically select the further parking trajectory or select between the determined parking trajectory and the determined further parking trajectory. This allows it to select the parking trajectory with the fewest changes of direction and / or the shortest length.

[0033] According to another embodiment, the trajectory planning algorithm is applied only if the rule-based trajectory planning algorithm fails to determine the further parking trajectory. Therefore, it is possible to first rely on well-known rule-based trajectory planning algorithms to determine the parking trajectory. The parking trajectory determined thereby is referred to as the further parking trajectory. This rule-based trajectory planning algorithm may, for example, include an artificial neural network based on supervised and / or unsupervised learning. However, if the rule-based trajectory planning algorithm determines a parking trajectory, this parking trajectory is accepted, and the vehicle is operated accordingly.However, if the rule-based trajectory planning algorithm cannot determine a parking trajectory that is, for example, collision-free and / or meets predetermined requirements, the trajectory planning algorithm trained using reinforcement learning with a curriculum is activated and applied to the sensor information. This means that it is possible to use the trajectory planning algorithm based on reinforcement learning only in those cases where established rule-based trajectory planning algorithms cannot provide a parking trajectory. This can be particularly efficient and / or fast in terms of computational power.

[0034] Furthermore, one embodiment includes the determined parking trajectory describing a forward or reverse parking maneuver. In particular, the determined parking trajectory is suitable for parking in a parallel, perpendicular, and / or herringbone parking space. The herringbone parking space may be referred to as an angled parking space. In general, the trajectory planning algorithm can be trained for different types of parking maneuvers. An individual trajectory planning algorithm can be present for each parking direction and / or each type of parking space. Preferably, a trajectory planning algorithm is trained for all possible parking scenarios, such as forward, reverse, parallel, perpendicular, and / or angled parking. The described method is therefore suitable for all types of parking situations.

[0035] The vehicle's control unit can perform all described steps of the procedure, if possible.

[0036] Furthermore, one aspect of the invention relates to a vehicle. The vehicle is configured to carry out the method described above. The vehicle is preferably a motor vehicle. The vehicle is, for example, a passenger car, a truck, a bus, a motorcycle, and / or a moped. The motor vehicle includes a control unit for carrying out the method.

[0037] Another aspect of the invention relates to a control unit for a vehicle, a mobile device, or a server. The control unit is configured to perform the method described above. The control unit performs the method.

[0038] The control unit is a computing unit. The control unit can be understood, in particular, as a data processing device that includes a processing circuit. The control unit can therefore, in particular, process data to perform computing operations. This can also include operations for performing indexed access to a data structure, for example, a lookup table (LUT).

[0039] In particular, the control unit may include one or more computers, one or more microcontrollers, and / or one or more integrated circuits, for example, one or more application-specific integrated circuits (ASICs), one or more field-programmable gate arrays (FPGAs), and / or one or more systems-on-chip (SoCs). The control unit may also include one or more processors, for example, one or more microprocessors, one or more central processing units (CPUs), one or more graphics processing units (GPUs), and / or one or more signal processors, in particular one or more digital signal processors (DSPs). The control unit may also include a physical or virtual cluster of computers or other of the units.

[0040] In various embodiments, the control unit includes one or more hardware and / or software interfaces and / or one or more memory units.

[0041] A memory unit can be implemented as a volatile data memory, for example dynamic random access memory, DRAM, or static random access memory, SRAM, or as a non-volatile data memory, for example read-only memory, ROM, programmable read-only memory, PROM, erasable programmable read-only memory, EPROM, electrically erasable programmable read-only memory, EEPROM, flash memory or flash EEPROM, ferroelectric random access memory, FRAM, magnetoresistive random access memory, MRAM, or phase-change random access memory, PCRAM.

[0042] In addition, one aspect of the invention relates to a computer program product that includes instructions. When the program is executed by a computer, such as the control unit, the instructions cause the computer to perform the method described above. The computer program product is a computer program.

[0043] Furthermore, the invention may include a computer-readable storage medium having instructions which, when executed by a computer such as the control unit, cause the computer to carry out the method described above.

[0044] Further embodiments of the invention are apparent from the claims, the figures, and the description. The invention includes combinations of the described embodiments. The embodiments described in connection with the method apply individually and in combination with one another, as applicable, to the vehicle, the control unit, and the computer program product.

[0045] Respective information within the meaning of the invention can include respective data. This means that the sensor information, for example, can include or be sensor data.

[0046] The figures show in: Fig. 1 a schematic representation of a vehicle in a parking lot environment; Fig. 2 a schematic representation of a method for operating at least an assisted parking function for a vehicle; and Fig. 3 a schematic representation of a trajectory planning algorithm to implement a method according to Fig. 2 to be carried out.

[0047] In the figures, functionally identical elements are given the same reference numerals.

[0048] Fig. 1 shows a vehicle 1 including a control unit 2. Alternatively, a mobile device and / or a server connected to the vehicle 1 may include the control unit 2. The control unit 2 may provide a parking function 3. The parking function 3 may be configured for forward and / or reverse parking into a parking space 6. The parking space 6 here is a perpendicular parking space. Alternatively or additionally, the parking function 3 may be configured for parking into a parallel and / or angled parking space 6. The angled parking space 6 may be a herringbone parking space 6.

[0049] The vehicle 1 may include a sensor device 4, such as a camera, a radar device, a lidar device, and / or an ultrasonic sensor. The sensor device 4 is preferably a front camera, a side camera, a rear camera, and / or an all-around camera. The sensor device 4 can detect the parking space 6 and the surroundings 5 ​​of the vehicle. The surroundings 5 ​​include an area available for maneuvering.

[0050] Vehicle 1 intends to park in parking space 6. Parking space 6 is bordered laterally by two obstacles 7. Here, the obstacles 7 are parked vehicles. A parking maneuver into parking space 6 can follow a parking trajectory 8 from a current location of a vehicle 1 to a parking location 9 in parking space 6. In the Fig. In the example outlined in Figure 1, a forward parking maneuver is determined by the parking function 3. However, more complex parking trajectories 8 are possible, which may include at least one change of direction at a turning point (not outlined here). Such a change of direction increases the number of steps required for parking.

[0051] Fig. Figure 2 shows steps of a method for operating an at least assisted parking function 3 for the vehicle 1. Assisted means that steering, braking, and / or driving commands are provided to a manual driver of the vehicle 1. Alternatively or additionally, the parking function 3 can provide semi-automatic or even fully automatic parking. In this case, the parking function 3 and therefore the control unit 2 are configured to provide control commands for a drive system, a steering system, and / or a braking system of the vehicle 1.

[0052] In a step S1, the sensor device 4 can provide sensor information 10. The sensor information 10 describes the parking space 6. In general, it describes the surroundings 5 ​​of the vehicle 1, which include the parking space 6. Here, the sensor information 10 also describes the obstacles 7. The acquired sensor information 10 is provided to the control unit 2 of the vehicle 1. The subsequent steps can all be performed by the control unit 2.

[0053] A step S2 involves determining the parking trajectory 8 from the current location of the vehicle 1 to the parking location 9 in the parking space 6 by applying a trajectory planning algorithm 11 to the provided sensor information 10. The method may then include checking, in a step S3, whether the determined parking trajectory 8 is free of obstacles 7. Therefore, a collision detection algorithm 12 may be applied to the determined parking trajectory 8 and the provided sensor information 10. If the parking trajectory 8 is free of obstacles 7, the vehicle 1 may be operated in a step S4 to move it at least with assistance along the determined parking trajectory 8.

[0054] However, if the verification step S3 results in the observation that the parking trajectory 8 is not free of obstacles 7, a step S5 may involve determining a repositioning trajectory 13 along which the vehicle 1 is movable from the current location to a starting location. The repositioning trajectory 13 is determined by applying a rule-based algorithm 14 to the provided sensor information 10. In a step S6, the vehicle 1 may be operated to move it along the determined repositioning trajectory 13 to the starting location. The trajectory planning algorithm 11 may then be applied again, taking the starting location as the current location. A new parking trajectory 8 may then be determined. This means that steps S2 and preferably also S4 are repeated based on the starting location as the current location.If a specific new parking trajectory 8 is free of obstacles 7, this is determined by performing step S4 on the new parking trajectory 8. Therefore, it is also possible to repeat steps S3 and S4 for the starting location as the current location. It can be checked whether a specific new parking trajectory 8 is free of obstacles 7 by applying the collision detection algorithm 12 to the specific new parking trajectory 8 and the provided sensor information 10. Only if this is the case can the at least assisted operation of the vehicle 1 to move it along the new parking trajectory 8 preferably take place in step S4.

[0055] As an alternative to the outlined sequence of steps, it is possible to operate the vehicle 1 only after checking that the determined new parking trajectory 8 is free of obstacles 7, in order to move it along the determined repositioning trajectory 13. In this case, after steps S3 and S5, step S2 is repeated directly, followed by steps S3, S6, and then S4.

[0056] The method may include a step S7, which involves determining a further parking trajectory 15 by applying a rule-based trajectory planning algorithm 16 to the sensor information 10 and selecting the parking trajectory 8, 15 from the determined parking trajectory 8 and the determined further parking trajectory 15 with the fewest changes of direction and / or the shortest trajectory length. The rule-based trajectory planning algorithm 11 and the rule-based trajectory planning algorithm 16 can be applied simultaneously or sequentially.

[0057] In particular, only if the rule-based trajectory planning algorithm 16 fails to determine the further parking trajectory 15 can the trajectory planning algorithm 11 be applied in step S2. However, if the rule-based trajectory planning algorithm 16 succeeds in determining the further parking trajectory 15, a step S9 can take place in which the vehicle 1 is operated along the further parking trajectory 15. In this case, it is possible to avoid performing steps S2 to S6. Therefore, it is possible to perform step S7 directly after step S1, before performing step S2.

[0058] Fig.Figure 3 shows the trajectory planning algorithm 11 used in more detail. The trajectory planning algorithm 11 includes an artificial neural network 20. The artificial neural network 20 is trained by reinforcement learning with a curriculum 23. Known reinforcement learning techniques can be used here. Consequently, the trajectory planning algorithm 11 includes an agent 21 and an environment 22. The environment 22 can be viewed as a computer game or simulation that generates various parking scenarios 24, 25, 26 for training the agent 21. The agent 21 evaluates a current situation, which is referred to as state s t referred to, undertakes an action a t and receives feedback from the environment 22 after each action. The feedback is used as a reward r t However, negative feedback is typically referred to as a punishment for making a mistake.

[0059] The trajectory planning algorithm 11 determines the parking trajectory 8 with the fewest changes of direction and / or the shortest trajectory length compared to other possible parking trajectories 8. Only if such a parking trajectory 8 is determined, the reward is given.

[0060] An output of the trajectory planning algorithm 11 can be converted into a steering angle curve 27 along the parking trajectory 8. Here, this is sketched as a curve. An x-axis shows a distance 28 from the current location, and a y-axis shows a steering angle 29 of the steering system of the vehicle 1. The steering angle curve 27 here includes three segments 30, 31, 32, each of which means that the vehicle 1 travels with a specific or changing steering angle 29 for a specific distance 28. The steering angle curve 27 is provided to the parking function 3 so that the steering system of the vehicle 1 can be operated according to the steering angle curve 27 along the parking trajectory 8.

[0061] Reinforcement learning with curriculum 23 means that several parking scenarios 24, 25, 26 are generated consecutively during training. The multiple parking scenarios 24, 25, 26 increase in difficulty within a parking environment and / or decrease within a given tolerance range of the parking location 9. Here, a first level is outlined as the first parking scenario 24. The parking space 6 is particularly wide with only two distant obstacles 7. The tolerance range for the parking location 9 is therefore high. It is relatively easy for the trajectory planning algorithm 11 to calculate a possible parking trajectory 8 for this first parking scenario 24. However, a second parking scenario 25 is already of a higher level and therefore a more complex parking scenario 24, 25, 26.In the second parking scenario 25, the number of obstacles 7 has increased and the parking space 6 is narrower because the obstacles 7 that border the parking space 6 are closer together compared to the first parking scenario 24, so that the tolerance range of the parking location 9 has been reduced. A third parking scenario 26 is even more advanced and therefore more complex, with even more obstacles 7 and an even narrower parking space 6 with a particularly small tolerance range for the parking location 9. The three sketch parking scenarios 24, 25, 26 are presented successively to the agent 21 during training based on reinforcement learning.

[0062] In summary, the invention includes a parking algorithm formulated as a reinforcement learning problem. This implements reinforcement learning in combination with path planning toward a target parking position. In addition, a classic rule-based algorithm can also be provided in parallel with the artificial intelligence-based path planning. During parking, the most promising method can be selected, for example, the one with minimal steps and / or the one that calculates the fastest.

[0063] In addition, the results of the trajectory planning algorithm 11 are tested in real-life scenarios to evaluate the parking planning algorithm 11 and improve its accuracy. QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] DE 10 2021 206 413 A1

[0003] DE 10 2020 127 216 A1

[0004]

Claims

[1] Method for operating at least an assisted parking function (3) for a vehicle (1), comprising: - providing and / or receiving sensor information (10) describing a parking space (6); and - determining (S2) a parking trajectory (8) from a current location of the vehicle (1) to a parking location (9) in the parking space (6) by applying a trajectory planning algorithm (11) to the provided sensor information (10); characterized by that the trajectory planning algorithm (11) includes an artificial neural network trained by reinforcement learning with a curriculum (23). [2] Method according to claim 1, characterized by that the method includes at least assisted operation (S4) of the vehicle (1) in order to move it along the determined parking trajectory (8). [3] Method according to one of claims 1 or 2, characterized bythat reinforcement learning with the curriculum (23) includes the continuous generation of several parking scenarios (24, 25, 26) that increase in the difficulty of a parking environment and / or decrease in a given tolerance range of the parking location (9) in the parking scenario (24, 25, 26). [4] Method according to one of the preceding claims, characterized by checking (S3) whether the determined parking trajectory (8) is free of obstacles (7) by applying a collision detection algorithm (12) to the determined parking trajectory (8) and the provided sensor information (10), wherein only if the parking trajectory (8) is free of obstacles (7) is the vehicle (1) operated to move it along the parking trajectory (8). [5] Method according to claim 4, characterized bydetermining (S5) a repositioning trajectory (13) along which the vehicle (1) is movable from the current location to a starting location if the determined parking trajectory (8) is not free of obstacles (7), and reapplying (S2) the trajectory planning algorithm (11) with the starting location as the current location of the vehicle (1) to determine a new parking trajectory (8). [6] Method according to claim 5, characterized by applying a rule-based algorithm (14) to the provided sensor information (10) to determine the repositioning trajectory (13). [7] Method according to claim 5 or 6, characterized bychecking whether the determined new parking trajectory (8) is free of obstacles (7) by applying the collision detection algorithm (12) to the determined new parking trajectory (8) and the provided sensor information (10), and only if this is the case, at least assisting operation of the vehicle (1) to move it along the new parking trajectory (8). [8] Method according to one of claims 5 to 7, characterized by operating the vehicle (1) to move it along the determined repositioning trajectory (13) to the starting location before reapplying the trajectory planning algorithm (11) or after checking that the determined new parking trajectory (8) is free of obstacles (7). [9] Method according to one of the preceding claims, characterized byconverting an output of the trajectory planning algorithm (11) into a steering angle profile (27) along the parking trajectory (8) and providing the steering angle profile (27) for the parking function (3). [10] Method according to one of the preceding claims, characterized by that the determined parking trajectory (8) includes a plurality of trajectory sections, wherein each time the trajectory planning algorithm (11) is applied, one of the trajectory sections is determined, so that the trajectory planning algorithm (11) is applied a plurality of times to determine the entire parking trajectory (8). [11] Method according to one of the preceding claims, characterized by that the trajectory planning algorithm (11) determines the parking trajectory (8) with the fewest changes of direction and / or the shortest trajectory length compared to other possible parking trajectories (8). [12] Method according to one of the preceding claims, characterized bydetermining (S7) a further parking trajectory (15) by applying a rule-based trajectory planning algorithm (16) to the sensor information (10) and selecting the parking trajectory (8, 15) from the determined parking trajectory (8) and the determined further parking trajectory (15) with the fewest changes of direction and / or the shortest trajectory length. [13] Method according to claim 12, wherein the trajectory planning algorithm (11) is applied only if the rule-based trajectory planning algorithm (16) fails to determine the further parking trajectory (15). [14] Method according to one of the preceding claims, characterized by that the determined parking trajectory (8) describes a forward or reverse parking maneuver, in particular for parking in a parallel, vertical and / or herringbone parking space (6). [15] Vehicle (1) adapted to carry out a method according to any one of the preceding claims. [16] Control unit (2) for a vehicle (1), a mobile device or a server, wherein the control unit (2) is configured to carry out a method according to one of claims 1 to 14. [17] A computer program product comprising instructions which, when executed by a computer, cause the computer to carry out a method according to any one of claims 1 to 14.

Images