Method for motion planning for a vehicle and method for training a machine learning algorithm for inferring maneuver costs in the context of motion planning for a vehicle

Abstract

The invention relates to a computer-implemented method (2) for motion planning for a vehicle (1), wherein the following steps are carried out in each of a plurality of calculation cycles: providing (21) state information relating to the vehicle (1) and environment information relating to an environment of the vehicle (1); defining (22) multiple possible maneuvers (M1, M2, M3, M4) in view of the state information and the environment information; generating (23) a particular feature vector for each maneuver (M1, M2, M3, M4); on the basis of the particular feature vector, inferring (24) costs for each maneuver (M1, M2, M3, M4) with respect to an optimization problem which comprises one or more trajectory-based cost functionals, by means of a machine learning algorithm trained for this purpose; and, according to the inferred costs, selecting (25) one or more of the maneuvers (M1, M2, M3, M4), preferably together with a specific sequence, for a detailed cost evaluation on the basis of the optimization problem.

Description

[0001] 24-2131

[0002] 1 / 22

[0003] Methods for motion planning for a vehicle and methods for training a machine learning algorithm to infer maneuver costs within the framework of motion planning for a vehicle

[0004] The invention relates to a method for motion planning for a vehicle and a method for training a machine learning algorithm to infer maneuver costs within the framework of motion planning for a vehicle. Furthermore, the invention relates to computing devices and computer programs for executing such methods, as well as computer-readable storage media comprising instructions for executing such methods.

[0005] Various automated driving functions are known from current technology, which increase safety and / or comfort for the driver and / or vehicle occupants. Depending on the degree of automation, automated driving functions can take over certain driving tasks or even control the vehicle at least partially autonomously.

[0006] Within the scope of this description, the term "automated driving" or "automated driving function" encompasses automated driving with any degree of automation. Examples of automation levels include assisted, partially automated, conditionally automated, highly automated, and fully automated driving (listed in order of increasing automation level). The five automation levels mentioned above correspond to SAE Levels 1 to 5 of the SAE J3016 standard (SAE - Society of Automotive Engineering) as of April 30, 2021. In assisted driving (SAE Level 1), the system performs longitudinal or lateral control in certain driving situations, with the expectation that the driver will handle all remaining aspects of the dynamic driving tasks.In partially automated driving (SAE Level 2), the system takes over longitudinal and lateral control in certain driving situations, although the driver must continuously monitor the system, as in assisted driving. In conditionally automated driving (SAE Level 3), the system takes over longitudinal and lateral control in certain driving situations without the driver needing to continuously monitor the system; however, the driver must be able to take over vehicle control within a certain timeframe if requested by the system. In highly automated driving (HAF; SAE Level 4), the system takes over vehicle control in certain driving situations, even if the driver does not respond to a request to intervene, thus eliminating the driver as a fallback option. In fully automated driving (VAF; SAE Level 5), the system can handle all aspects of the vehicle.

[0007] 2 / 22

[0008] dynamic driving tasks can be performed under any road and environmental conditions that can also be mastered by a human driver.

[0009] In accordance with the above, some automated driving functions intervene directly in the vehicle's control or take over the vehicle's control completely – permanently or at least temporarily. Examples include longitudinal control assistance systems, lateral control assistance systems, and assistance systems with coupled longitudinal and lateral control. Such automated driving functions can, for example, implement so-called hands-off or feet-off functionality, which relieves the driver of certain steering or pedal operation tasks (e.g., lane keeping, following other vehicles, etc.), up to and including conditionally automated driving or even highly automated driving.

[0010] Automated driving functions of the type mentioned above generally require automatic motion planning by the vehicle's control system, the result of which is a (target) trajectory. In this context, a trajectory is understood as a planned, temporally defined sequence of relevant vehicle state variables. These vehicle state variables can include, in particular, the vehicle's spatial coordinates, so that trajectory planning defines a suitable path in space and time along which the vehicle is to be controlled. The trajectory should be compatible with a goal specified by a higher-level driving strategy (also referred to as maneuver planning), such as maintaining a desired speed and / or changing lanes.

[0011] A common requirement for trajectory planning is that the trajectory must not only be feasible and safe (especially collision-free) with regard to limitations imposed by the vehicle's surroundings and actuators, but also that it should be as comfortable as possible for the driver and / or vehicle occupants, for example, by avoiding excessive acceleration, deceleration, and / or jerking. Generally, therefore, an optimal trajectory must be determined with respect to several different criteria. These various criteria sometimes place conflicting demands on the trajectory design, so trajectory planning involves dynamic optimization in the sense of finding a compromise between differently weighted requirements for the desired trajectory.

[0012] Known algorithms for solving such dynamic optimization problems are based on minimizing a cost functional. The cost functional is typically a function of several time-dependent vehicle state variables (for example, 24-2131).

[0013] 3 / 22

[0014] The system consists of a time-dependent state vector and several time-dependent control variables (e.g., summarized in a time-dependent control vector). The control variables can, for example, quantify the manipulated variables of the vehicle's actuators and thus be, for example, forces (or accelerations) or torques (or angular accelerations).

[0015] When this document refers to a cost functional, it generally means a functional that outputs a value for evaluating a given trajectory with respect to predetermined criteria. A suitable term would also be "performance functional," since, depending on the specific design of the optimization algorithm, the goal is not necessarily to minimize costs, but may also be to maximize positive values.

[0016] Solving the optimization problem involves determining the time course of the control variables—and consequently, the resulting time course of the vehicle state variables—that minimizes the cost functional up to a specific time horizon (i.e., time-integrated to a final time, which, as mentioned in section II, may itself be variable and the subject of optimization). This typically requires consideration of several boundary and constraint conditions, such as initial conditions, final conditions, and inequality restrictions. For example, the requirement of collision avoidance with objects in the vehicle's environment and / or adherence to predefined actuator control limits can be formulated as such constraints.

[0017] As an important basis for trajectory planning, an environmental model is typically generated from fused sensor data and, if necessary, other data such as map information from a database. This environmental model describes the current and / or future predicted vehicle environment to be considered during trajectory planning. It includes, for example, static and / or moving objects such as stationary obstacles or other road users. Stationary obstacles can be, for example, parts of the road infrastructure, such as guardrails or solid lane markings on the road surface, which could easily be driven over physically, but are prohibited by law. Other moving road users can be, for example, vehicles ahead or approaching from the opposite direction, or pedestrians or animals in the vehicle's vicinity.24-2131

[0018] 4 / 22

[0019] It is a common procedure in motion planning to first define several possible maneuvers based on the current environmental model and current state information regarding the vehicle (e.g., its speed). These maneuvers, which can also be understood as trajectory planning tasks, define the boundary conditions and objectives for the vehicle's movement over a specific time horizon. For example, the planning horizon can extend a few seconds into the future (e.g., in the range of 5 to 20 seconds, such as 6 seconds or 15 seconds). Subsequently, for each maneuver, a suitable trajectory is planned using an optimization process (trajectory planning). Each defined maneuver and its associated trajectory are evaluated using several cost functionals to determine the final decision for the best maneuver.Some of these cost functionals take the trajectory or quantities contained within the trajectory as input(s) (so-called trajectory-based cost functionals), while others only require data from the environment model and the predefined maneuver as input, but not the trajectory itself (so-called maneuver-based cost functionals). The most cost-effective of these alternative trajectories is passed to the subsequent chain of effects and the vehicle's actuators in each calculation cycle.

[0020] In this motion planning process, the trajectory planning optimization procedure is by far the most computationally intensive step. Therefore, depending on available computing resources, it may not be possible to evaluate all maneuvers in a single system calculation cycle. Instead, it might be possible, for example, to plan and evaluate only two maneuvers in detail from the pool of all possible maneuvers in the current environment in each calculation cycle: firstly, the maneuver most recently rated as optimal, whose trajectory was sent to the actuators, and secondly, another maneuver, which is selected from the other possible maneuvers, for example, using a round-robin scheduling method. Depending on the total number of maneuvers to be evaluated, this approach may take several calculation cycles to evaluate and select the most cost-effective maneuver.The vehicle may therefore behave suboptimally for a period of time.

[0021] The development of advanced automated driving functions, particularly for increasingly urban environments and correspondingly congested traffic situations, presents numerous new challenges that necessitate the evaluation of a greater number of trajectories during motion planning. If the evaluation of these 24-2131

[0022] 5 / 22

[0023] If a larger number of trajectories are distributed over several calculation cycles according to the approach described above, this can lead to the vehicle behaving suboptimally for an even longer period of time (or over a larger number of calculation cycles).

[0024] Against this background, it is an object of the present invention to provide an effective method for motion planning for a vehicle which addresses the challenges described above.

[0025] This problem is solved by the computer-implemented methods, computing devices, computer programs, and computer-readable (storage) media according to the independent claims. Advantageous embodiments are specified in the dependent claims.

[0026] A first aspect of the invention relates to a computer-implemented method for motion planning for a vehicle. The motion planning, and thus the method proposed here, can be carried out particularly within the framework of an automated driving function of the vehicle, such as an automated driving function according to SAE Level 2 or a higher degree of automation.

[0027] The process is executed in several calculation cycles. In this context, a calculation cycle can be understood as a data processing cycle that encompasses the complete information processing from sensor input, through environmental perception and interpretation, behavior and motion planning, to the control of the vehicle's actuators. For example, a calculation cycle can have a duration in the range of a few milliseconds, such as approximately 80 ms or 40 ms.

[0028] As a matter of procedure, certain steps are carried out in each calculation cycle, which are explained below.

[0029] First, the most up-to-date vehicle status information, such as speed and yaw rate, as well as environmental information about the vehicle's surroundings, are provided. This environmental information can include, in particular, a current and / or predicted environmental model.24-2131

[0030] 6 / 22

[0031] Based on the provided state and environmental information, several alternative maneuvers are defined. These maneuvers establish the boundary conditions and objectives of the vehicle's movement at a comparatively abstract behavioral level. They can also be understood as trajectory planning tasks, where a trajectory, to be determined within the framework of further motion planning, specifies the vehicle's movement according to the given maneuver. Examples of such maneuvers include maintaining a desired speed, synchronizing with a gap in a secondary lane to prepare for a lane change (e.g., one or more alternative behaviors, such as distance-based or speed-based planning, can be considered for each possible gap), or initiating or aborting a lane change.

[0032] For each defined maneuver, a corresponding feature vector is generated. A feature vector contains properties of the environmental model aspects that are particularly relevant to the maneuver. For example, this can include the vehicle's speed, road curvature, or the distance and / or relative speed of vehicles in the environment, such as a vehicle in front of and a vehicle behind it. Optionally, the feature vector can also include a behavioral prediction of such vehicles in the environment.

[0033] For example, the feature vector for a maneuver can include one, several, or all of the following:

[0034] - Distance and relative speed to the vehicle in front and behind the gap in the adjacent lane (preferably as floating-point numbers);

[0035] - Distance and relative speed to the vehicle in front on the current lane (preferably as floating-point numbers);

[0036] - Vehicle speed (preferably as a floating-point number);

[0037] - Indication of the existence of each of the vehicles (preferably as Boolean variables); - Type of behavior (preferably encoded by different values ​​of a floating-point number);

[0038] - remaining time to synchronize to a gap, based on a navigation destination or a time elapsed since the turn signal was activated (preferably as a floating-point number);

[0039] - Vehicle sizes (preferably as floating-point numbers).

[0040] Based on the respective feature vector, costs for each maneuver are inferred. This is done using a specially trained machine learning algorithm. The costs are 24-2131.

[0041] 7 / 22

[0042] These refer to an optimization problem that includes one or more trajectory-based cost functionals. They are cost contributions of the trajectory-based cost term(s) of the optimization problem for the respective maneuver, inferred from a trained model; in other words, they are a prediction of the costs that the trajectory-based cost term(s) would yield for a trajectory that optimally implements the maneuver.

[0043] The machine learning algorithm (ML algorithm) used can, for example, be based on an artificial neural network, such as a convolutional neural network (CNN), in a manner known per se.

[0044] It may be intended that some, several or all of the following aspects are already taken into account in the ML algorithm and during ML validation during training:

[0045] - A loss function used during training can take into account that the predictive accuracy depends on the ratio of trajectory costs (i.e., costs from trajectory-based cost functionals) to maneuver costs (i.e., costs from maneuver-based cost functionals), preferably with the tendency that the greater the trajectory costs compared to the maneuver costs, the higher the loss function should be for that training sample. This tendency can be implemented in addition to the cost function otherwise used.

[0046] - The validation metrics used can assess the influence of the cost prediction accuracy on the selection of a sequence, described in more detail below, in which detailed cost assessments are to be carried out for several maneuvers;

[0047] Regarding network architecture, focusing on efficient (i.e., particularly relatively small) architectures (e.g., with a total of 3 x 256 or 1 x 128 neurons, where the above notation refers to the "number of layers x number of neurons per layer") that can be computed on the vector units of a typical ECU target can be advantageous. The goal can be to ensure that the inference process optimally utilizes the ECU's vector unit.

[0048] Depending on the inferred costs, one or more maneuvers, preferably in a specific order, are selected for a detailed cost evaluation based on the optimization problem. This detailed cost evaluation can be performed by solving the optimization problem in one or more subsequent computation cycles. Preferably, the maneuver(s) are selected, 24-2131

[0049] 8 / 22

[0050] which prove to be the most cost-effective maneuver(s) according to the inferred costs.

[0051] In line with the foregoing, the invention is based on the idea of ​​first learning the results of the trajectory-based cost functional offline using a machine learning (ML) model and then predicting them computationally efficiently online in the vehicle for a large number of potential maneuvers. These ML-based predictions can then be combined with maneuver-based cost functionals to predict the total cost of a maneuver. These total cost predictions can ultimately be used to selectively evaluate the most promising maneuvers in detail at an early stage (instead of, for example, using a frequently suboptimal round-robin approach), including the computationally intensive trajectory planning.

[0052] In an embodiment of the method in which the costs for the several alternative maneuvers are inferred in each calculation cycle, it may be provided that only a single alternative is selected for detailed cost evaluation based on the inferred costs.

[0053] However, it is also possible to implement ML-based maneuver cost inferences in a specific order for detailed cost evaluation, rather than in every calculation cycle. In such an implementation, several maneuvers are selected along with a specific sequence for detailed cost evaluation, which is then carried out successively in subsequent calculation cycles according to that sequence.

[0054] According to one embodiment of the invention, the method further provides that a detailed cost evaluation based on the given optimization problem is performed in a current and / or in one or more (e.g., immediately) subsequent computation cycles for several candidate maneuvers. The candidate maneuvers include, in particular, the maneuver(s) selected according to the method. If the cost evaluation for the several selected maneuvers is performed in different subsequent computation cycles, it is preferably carried out in the order determined by the method.

[0055] Performing the detailed cost evaluation can be achieved by planning a detailed trajectory for each of the several candidate maneuvers based on the 24-2131.

[0056] 9 / 22

[0057] This can be done using a common optimization method. The expert is familiar with mathematical and algorithmic methods for solving such dynamic optimization problems using one or more cost functionals. The minimization of the cost functional to determine a suitable trajectory can then be carried out in a suitable manner known to the expert. For example, known mathematical methods of the calculus of variations, such as using a Hamiltonian function and the resulting Euler-Lagrange equations, can be used. Alternatively or additionally, known local optimization methods, which may be implemented in gradient-free and / or gradient-based algorithms, can be used to solve such optimization problems. Furthermore, methods of global optimization, such as...This can be implemented using optimization methods analogous to natural processes or dynamic programming. For this purpose, numerous ready-made solver algorithms are available to the specialist.

[0058] For example, optimization using Quadratic Programming can be provided, whereby a separate solution of the longitudinal and the lateral planning problem can be performed.

[0059] The procedure can further include selecting one of the meticulously planned trajectories based on the results of the detailed cost assessment. The selected trajectory will then form the basis for controlling the vehicle.

[0060] According to one embodiment, the candidate maneuvers always include the maneuver for which a detailed trajectory (implementing the corresponding maneuver) for vehicle control was last selected (i.e., in an immediately preceding calculation cycle). In other words, the most recent maneuver, whose associated (target) trajectory was (or is currently) used as the basis for vehicle control, can be included again as a candidate maneuver in the evaluation. This aspect is based on the idea that a maneuver last evaluated as optimal is highly likely to remain the optimal maneuver, with the associated trajectory simply needing to be updated in the subsequent calculation cycle, i.e., adapted in detail to an updated environmental model and extended a certain distance into the future according to a temporal trajectory planning horizon.24-2131.

[0061] 10 / 22

[0062] According to a further embodiment of the invention, the method also includes generating control signals for actuators to control the vehicle with the selected trajectory as the target trajectory.

[0063] The selection of the maneuver or maneuvers (preferably including the specific sequence) for the detailed cost assessment can also be carried out depending on costs that have been inferred for one or more of the maneuvers in one or more past calculation cycles and / or depending on a result of a detailed cost assessment carried out in one or more past calculation cycles of one or more (corresponding to or implementing the maneuver(s)) detailed planned trajectories.A relative weighting of the aforementioned influencing factors can be provided for the selection decision, which depends on the calculation method (ML-based inference or detailed cost evaluation), preferably with the tendency that a value based on a detailed cost evaluation by solving the optimization problem tends to be weighted more heavily than a value that has been inferred on the basis of the ML model.

[0064] When selecting the maneuver(s) (preferably along with their specific sequence) for the detailed cost assessment, it is also possible to consider how long ago (i.e., how many calculation cycles) it has been since the respective influencing factor(s) mentioned in the preceding paragraph were determined. In other words, the "age" of the respective values ​​of the influencing factor(s) can be taken into account. A relative weighting of the aforementioned influencing factors can be provided for the selection decision, depending on the age of the respective values, preferably with the tendency that a value (e.g., a cost contribution inferred in a previous calculation cycle) is weighted less the older it is.

[0065] For example, in each calculation cycle, the detailed maneuver costs planned in past calculation cycles, with their corresponding age and weighting, as well as the costs inferred by the machine learning model in the current and past calculation cycles, with their corresponding age and weighting, can be considered when selecting the maneuver(s) to be planned in detail. Preferably, the respective values ​​are also weighted according to the calculation method (machine learning-based inference or detailed cost assessment). Overall, such an approach with different cost contributions weighted according to age and calculation method can represent and account for epistemic uncertainties in the selection of the most promising maneuvers.Epistemic uncertainties regarding the accuracy of the costs inferred by the machine learning model can play a role here and can be estimated either generally or anew in each calculation cycle. This form of uncertainty models how strongly the machine learning model should be trusted in cost prediction.

[0066] Considering the age of maneuver costs inferred or determined in detail in previous calculation cycles when selecting a maneuver is particularly advantageous in embodiments where maneuver costs are inferred in each calculation cycle and only a single maneuver is selected for detailed cost evaluation. Without considering the age of the maneuvers, in such an embodiment, the same maneuver could be selected repeatedly, and alternative maneuvers might not be considered at all.

[0067] In embodiments where the ML-based inference of maneuver costs is not performed in every calculation cycle, but only in some calculation cycles and then in advance for some subsequent calculation cycles, considering the age of the maneuver costs inferred or determined in detail in past calculation cycles can also be advantageous as an optional add-on.

[0068] It is also within the scope of the invention that the optimization problem can be designed in such a way that consistent behavior of the vehicle, i.e., specifically avoiding a change of a target gap, tends to be rewarded at the level of maneuver costs.

[0069] A second aspect of the invention relates to a computer-implemented method (hereinafter also referred to as a training method) for training a machine learning algorithm (i.e., an ML model) to infer maneuver costs within the context of motion planning for a vehicle. In particular, the maneuver costs within the motion planning process according to the first aspect of the invention can be inferred using an ML model trained according to this method.

[0070] The training procedure involves defining an optimization problem for motion planning, where the optimization problem includes at least one trajectory-based cost functional, and defining a number of possible maneuvers that a vehicle can perform with respect to its environment. Furthermore, simulated and / or real-world datasets are provided for each maneuver, with each dataset containing state information about a real or simulated vehicle and environmental information about the surroundings of the real or simulated vehicle. For each dataset, the respective costs are then calculated by evaluating the trajectory-based cost functional of the optimization problem. In contrast to in-vehicle execution, the software can be configured differently for this purpose, allowing for more computation time per cycle.This allows all maneuvers, including trajectories, to be calculated in detail in each calculation cycle. Therefore, all evaluated cost functionals are available in the dataset for each calculation cycle.

[0071] Subsequently, a respective feature vector is generated for each of the datasets and ground truth datasets are provided, with each ground truth dataset comprising one of the feature vectors in combination with the associated calculated costs from the trajectory-based cost functionals.

[0072] Based on these ground-truth datasets, the machine learning algorithm is ultimately trained to infer the costs of the trajectory-based cost functional of the optimization problem for maneuvers characterized by feature vectors. It goes without saying that when applying the trained algorithm, the feature vectors characterizing the maneuvers may differ from the feature vectors derived from the datasets mentioned above (in the values ​​of the vector entries).

[0073] In other words, the training process can be described as follows: In an offline learning phase, all possible variations of the scenario are simulated or derived from real-world data. For each of these maneuvers, a corresponding feature vector is extracted in each computation cycle, and the associated trajectory-based cost functionals are calculated. This data, consisting of feature vectors and calculated trajectory-based costs, is stored in a dataset. Based on this dataset, a machine learning model can then be trained to predict the various trajectory-based costs based on the feature vector.

[0074] The fully trained machine learning model can be integrated into vehicle software so that, during the operation of an automated driving function, in accordance with the motion planning method according to the first aspect of the invention, a cost prediction for each possible maneuver can be generated in each calculation cycle. Based on this cost prediction, the time elapsed since the last trajectory planning for a given maneuver, and the maneuver-based cost functionals, a new decision can be made in each calculation cycle as to which maneuvers should be subjected to detailed evaluation, including trajectory planning.

[0075] For a final cost-based decision on which maneuver should be implemented by the actuators, a precise cost evaluation based on trajectory planning can be used. The machine learning model no longer plays a role in this final selection step, so even a prediction error in the model will not have a catastrophic impact on the overall behavior of the vehicle.

[0076] According to a third aspect of the invention, a computing device is proposed, wherein the computing device is set up (in particular programmed) to carry out a method according to the first and / or second aspect of the invention.

[0077] In particular, if the computing device is set up to carry out the method according to the first aspect of the invention (method for motion planning), it can be part of a control system of the vehicle, which includes one or more processors (such as CPUs and / or GPUs) on which the necessary computational operations for carrying out the method are performed.

[0078] If the computing device is set up to carry out the method according to the second aspect of the invention (training method), it can, for example, be located in a development environment of a vehicle manufacturer.

[0079] A fourth aspect of the invention is a vehicle with a computing device that is set up (in particular programmed) to carry out a method according to the first aspect of the invention.

[0080] Preferably, the vehicle is a motor vehicle. A motor vehicle in this sense is a land vehicle that is moved by mechanical power and is not bound to railway tracks. A motor vehicle in this sense can be, for example, a car, motorcycle, or tractor. In particular, the vehicle can be a vehicle with one or more automated driving functions, whereby trajectory planning is carried out within the framework of the automated driving function(s), taking into account an environmental model.

[0081] A fifth aspect concerns a computer program which includes instructions which, when the computer program is executed by a computing device, cause it to execute a method according to the first and / or second aspect of the invention.

[0082] A sixth aspect of the invention relates to a computer-readable (storage) medium which includes instructions which, when executed by a computing device, cause it to execute a method according to the first and / or second aspect of the invention.

[0083] It is understood that the computing devices mentioned above in connection with the fourth, fifth and sixth aspects of the invention may in particular be computing devices according to the third aspect of the invention.

[0084] In summary, the invention enables effective motion planning for a vehicle, addressing the challenges of evaluating a large number of trajectories. By combining machine learning and optimization methods, the vehicle can efficiently assess the most relevant maneuvers and behave more optimally, thereby improving the safety and comfort of automated driving.

[0085] The invention will now be explained in more detail with reference to exemplary embodiments and the accompanying drawings.

[0086] Fig. 1 shows a schematic flowchart of a computer-implemented method for motion planning for a vehicle.

[0087] Fig. 2 shows a schematic flowchart of a computer-implemented procedure for training a machine learning algorithm to infer maneuver costs in the context of motion planning for a vehicle.24-2131

[0088] 15 / 22

[0089] Figures 3A-D schematically show several possible alternative maneuvers of a vehicle in an example scenario that can be considered during motion planning.

[0090] Fig. 1 shows a schematic flowchart of a computer-implemented method 2 for motion planning for a vehicle 1 (hereinafter also referred to as: ego-vehicle). The following steps are executed in several calculation cycles:

[0091] First, status information regarding Egovehicle 1 and environment information regarding the environment of Egovehicle 1 are provided (step 21).

[0092] Subsequently, several possible maneuvers are defined, taking into account the state and environment information (step 22).

[0093] Figures 3A-D illustrate various possible maneuvers in an example scenario: The ego-vehicle 1, with ACC and LSA (Lane and Steering Assist) activated, approaches an exit on the right but is still in the left lane of a two-lane road. In this situation, up to seven different maneuvers can be defined, for example:

[0094] Figure 3A illustrates a maneuver M1 in which the longitudinal target position is in the middle of the gap in the adjacent lane at the level of the ego vehicle 1, while the lateral target position remains in the left lane. The other vehicle 61 traveling in the right lane is considered the vehicle behind the target gap, and the other vehicle 62 is considered the vehicle in front of the target gap.

[0095] Figure 3B shows a maneuver M2 in which the longitudinal target position is in a gap further ahead in the adjacent lane (i.e., one gap further ahead than in Figure 3A), while the lateral target position also remains in the left lane. Here, the other vehicle 62 is considered the gap-rear vehicle and the other vehicle 63 the gap-front vehicle.

[0096] Figure 3C shows a maneuver M3 in which the longitudinal target position is in a gap in the adjacent lane even further ahead (i.e., two gaps further ahead than in Fig. 3A), with the lateral target position also remaining in the left lane. The foreign vehicle 63 is 24-2131

[0097] 16 / 22

[0098] The foreign vehicle 64 is considered the gap-filling rear vehicle and the gap-filling front vehicle.

[0099] Figure 3D shows a maneuver M4 in which the longitudinal target position is in the middle of the gap in the adjacent lane at the level of the ego vehicle 1, with the lateral target position being the right-hand adjacent lane. As in the maneuver according to Fig. 3A, the other vehicle 61 is considered the vehicle behind the gap and the other vehicle 62 the vehicle in front of the gap. However, unlike the maneuver according to Fig. 3A, the maneuver according to Fig. 3D involves an actual lane change to the right-hand lane. Should this maneuver prove to be the most cost-effective maneuver in the further course of the motion planning process 2, it can be provided that the driver of the ego vehicle 1 is merely prompted to change lanes, but the steering and speed control are not subsequently executed automatically according to this actual lane change maneuver.

[0100] In addition to the maneuvers shown in Figures 3A-D, there are further maneuver alternatives that are not included in these merely exemplary illustrations:

[0101] For example, a maneuver could aim for a target speed below that of the adjacent lane vehicle 61 behind the ego vehicle 1, while maintaining the lateral target position on the left lane.

[0102] Another maneuver could involve a target speed below that of the adjacent lane vehicle 62 in front of the ego vehicle 1, also with the lateral target position on the left lane.

[0103] Finally, a maneuver could be considered that completely ignores the secondary track objects 61, 62, 63, 64 and only fulfills the customer function requirements of ACC and optionally also LSA.

[0104] For each of the maneuvers M1, M2, M3, M4, a corresponding feature vector is generated (step 23). When maneuvers M1, M2, M3, M4 are mentioned here and in the following, this also includes the maneuver alternatives mentioned above that are not shown in the figures.

[0105] Based on the respective feature vector, a specially trained machine learning algorithm calculates the costs for each of the maneuvers M1, M2, M3, M4 with reference to ein24-2131.

[0106] 17 / 22

[0107] Inferring an optimization problem that includes one or more trajectory-based cost functionals (Step 24).

[0108] Depending on the inferred costs, one of the maneuvers M1, M2, M3, M4 or several of the maneuvers M1, M2, M3, M4, preferably in a specific order, are selected for a detailed cost evaluation based on the optimization problem (step 25).

[0109] In a current and / or in one or more subsequent calculation cycles, a detailed cost evaluation is then carried out for several candidate maneuvers, including the selected maneuver or the several selected maneuvers, preferably taking into account the specific order if the cost evaluation for the several selected maneuvers is carried out in different subsequent calculation cycles (optional step 26).

[0110] Performing the detailed cost evaluation involves planning a detailed trajectory for each of the several candidate maneuvers based on the optimization problem. One of the planned trajectories is selected based on a result of the detailed cost evaluation (optional step 27), and this selected trajectory will then be used to control the vehicle.

[0111] Finally, control signals can be generated to control the vehicle according to the selected trajectory (optional step 28).

[0112] Fig. 2 shows a schematic flowchart of a computer-implemented method 3 for training a machine learning algorithm to infer maneuver costs within the context of motion planning for a vehicle 1, as it can be used in method 2 according to Fig. 1. Method 3 comprises the following steps:

[0113] First, an optimization problem for motion planning is defined, wherein the optimization problem includes at least one trajectory-based cost functional (Step 31).24-2131

[0114] 18 / 22

[0115] A number of possible maneuvers are defined that a vehicle can perform in relation to its environment (step 32). These may include, for example, maneuvers M1, M2, M3, M4 from Figures 3A-D or similar maneuvers.

[0116] Simulated and / or real-world data sets for each of the maneuvers M1, M2, M3, M4 are provided, with each data set including state information regarding a real or simulated vehicle and environment information regarding the environment of the real or simulated vehicle (Step 33).

[0117] For each of the data sets, the respective costs are calculated by evaluating the trajectory-based cost functional of the optimization problem (step 34).

[0118] Furthermore, a respective feature vector is extracted from each of the datasets (step 35).

[0119] Ground truth datasets are provided, with each ground truth dataset comprising one of the feature vectors in combination with the associated calculated costs from the trajectory-based cost functionals (step 36).

[0120] Finally, based on the ground truth datasets, a machine learning algorithm is trained to infer costs of the trajectory-based cost functional of the optimization problem for a maneuver M1, M2, M3, M4 characterized by a feature vector (step 37).

[0121] The motion planning procedure 2 according to Fig. 1 and the training procedure 3 according to Fig. 2 work together to ensure efficient motion planning for the vehicle 1 and, as a result, to plan an optimal trajectory for the vehicle 1 that takes into account the defined optimization goals and the current environmental and driving conditions.

Claims

Patent claims 1. Computer-implemented method (2) for motion planning for a vehicle (1), wherein the following steps are performed in several calculation cycles: - Providing (21) status information regarding the vehicle (1) and environmental information regarding the vehicle's surroundings (1); - Define (22) several possible maneuvers (M1, M2, M3, M4) taking into account the state and environment information; - Generating (23) a respective feature vector for each of the maneuvers (M 1 , M2, M3, M4); - based on the respective feature vector, inferring (24) costs for each of the maneuvers (M1, M2, M3, M4) with reference to an optimization problem which includes one or more trajectory-based cost functionals, using a machine learning algorithm trained for this purpose; and - depending on the inferred costs, selecting (25) one of the maneuvers (M1, M2, M3, M4) or several of the maneuvers (M1, M2, M3, M4), preferably together with a specific order, for a detailed cost evaluation based on the optimization problem.

2. Method (2) according to claim 1, further comprising: - Performing (26), in a current and / or in one or more subsequent calculation cycles, the detailed cost evaluation for several candidate maneuvers, which include the selected maneuver (M1, M2, M3, M4) or the several selected maneuvers (M1, M2, M3, M4), preferably taking into account the specified order when the cost evaluation for the several selected maneuvers (M1, M2, M3, M4) is performed in different subsequent calculation cycles.

3. Method (2) according to claim 2, wherein performing (26) the detailed cost evaluation is accompanied by detailed planning of a respective trajectory for each of the multiple candidate maneuvers based on the optimization problem.

4. Method (2) according to claim 3, further comprising: - Selections (27) of one of the detailed planned trajectories depending on a result of the detailed cost assessment, wherein the selected trajectory is to be used as the basis for controlling the vehicle (1).

5. Method (2) according to claim 4, wherein among the candidate maneuvers is a maneuver (M1, M2, M3, M4) for which a detailed planned trajectory for controlling the vehicle was last selected.

6. Method (2) according to claim 4 or 5, further comprising: Generating (28) control signals to control the vehicle (1) according to the selected trajectory.

7. Method (2) according to one of the preceding claims, wherein the selection (25) of the maneuver (M1, M2, M3, M4) or the multiple maneuvers (M1, M2, M3, M4) for the detailed cost evaluation is also carried out depending on one or more of the following influencing factors: - inferred costs for one or more of the maneuvers (M1, M2, M3, M4) in one or more past calculation cycles; - a result of a detailed cost assessment of one or more detailed planned trajectories carried out in one or more past calculation cycles.

8. Method (2) according to claim 7, wherein when selecting (25) the maneuver (M1, M2, M3, M4) or the multiple maneuvers (M1, M2, M3, M4) for the detailed cost evaluation, it is also taken into account how long it has been since the influencing factor(s) referred to in claim 7 were determined.

9. Computer-implemented method (3) for training a machine learning algorithm to infer maneuver costs in the context of motion planning for a vehicle (1), comprising the steps: - Define (31) an optimization problem for motion planning, wherein the optimization problem includes at least one trajectory-based cost functional; - Define (32) a number of possible maneuvers (M1, M2, M3, M4) that a vehicle can perform with respect to an environment of the vehicle; - Provide (33) simulated and / or real-world data sets for each of the maneuvers (M1, M2, M3, M4), wherein the data sets each include state information relating to a real or simulated vehicle, as well as environment information relating to an environment of the real or simulated vehicle; - For each of the datasets, compute (34) costs by evaluating the trajectory-based cost functional of the optimization problem; - Generate (35) a respective feature vector for each of the datasets; - Provide (36) ground truth datasets, each ground truth dataset comprising one of the feature vectors in combination with the associated computed costs from the trajectory-based cost functionals; - based on the ground truth datasets, training (37) the machine learning algorithm to infer costs of the trajectory-based cost functional of the optimization problem for a maneuver characterized by a feature vector (M1, M2, M3, M4).

10. Computing device, wherein the computing device is designed to perform a method (2, 3) according to one of the preceding claims.

11. Computer program comprising instructions which, when the computer program is executed by a computing device, cause it to execute a method (2, 3) according to any one of claims 1 to 9.

12. Computer-readable (storage) medium comprising instructions which, when executed by a computing device, cause it to execute a method (2, 3) according to any one of claims 1 to 9.

Images