IT202300022488A1 - DRIVING TRAJECTORY PLANNER - Google Patents

Description

DESCRIPTION

of the patent for industrial invention entitled:

DRIVING TRAJECTORY PLANNER

TECHNICAL SECTOR OF THE INVENTION

The present invention relates in general to the automotive field, in particular to a vehicle dynamics control system. In particular, the present invention relates to the planning of driving paths for an autonomous driving system.

STATE OF THE ART

As is well known, planning driving trajectories is a complex task that involves the collection and analysis of data from sensors and other vehicle detection systems. In the automotive field, various autonomous driving systems are used to calculate driving trajectories for a vehicle to follow.

In particular, several known automotive systems are designed to employ rule-based systems or finite-state machines to perform decision-making processes, for example based on nth-order logic statements, and to compute driving trajectories based on geometric or kinematic formulas and/or based on constraint-based optimizations.

It is also known that a driving system is used that allows a predefined number of optimization steps to be performed in order to determine the optimal solution in a predetermined set of driving trajectories; furthermore, it is known that the driving system is designed to simultaneously determine a fallback trajectory in case the optimization steps prove insufficient to achieve convergence.

Furthermore, it is known to use automotive systems configured to assign a score to each trajectory of a set of driving trajectories and to select a trajectory belonging to that set based on the assigned scores.

Furthermore, it is known that a vehicle system can be designed to calculate driving trajectories using artificial neural networks trained on data relating to driving trajectories traveled by expert human drivers.

OBJECT AND SUMMARY OF THE INVENTION

The Applicant was able to observe that the solutions according to the prior art, although satisfactory in certain aspects, are open to improvement.

In particular, a rule-based system does not allow for the calculation of driving trajectories to handle complex scenarios, such as heavy traffic; furthermore, an operator or designer must manually define the rules of such a rule-based system.

Furthermore, the Applicant was able to observe that the known optimization solutions configured to operate in real time assume a quadratic (rather than multimodal) problem, and therefore can easily converge to local minima.

The Applicant was also able to observe that, if a vehicle system is designed to calculate, or plan, a driving trajectory through the use of geometric or kinematic formulas, a vehicle following such a trajectory is not able to manage complex behaviors and take into account the comfort of the driver of the vehicle.

The Applicant further observed that automotive systems designed to plan driving trajectories using artificial neural networks are typically less susceptible to the influence of vehicle safety constraints (for example, avoiding collisions with other vehicles) than other systems. Furthermore, the Applicant observed that the data on which such artificial neural networks are trained often suffer from covariate shift (covariate shift).

The aim of the present invention is therefore to provide a software for planning driving paths that allows for improving, at least in part, the solutions of the prior art.

According to the present invention, a software for planning driving paths is provided as claimed in the appended claims.

BRIEF DESCRIPTION OF THE DRAWINGS

Figure 1 shows a block diagram of an electronic vehicle dynamics control system according to the present invention.

Figure 2 shows a block diagram of a training method of a driving trajectory planning artificial neural network according to the present invention.

Figure 3 shows a block diagram of a training method of an artificial neural network for evaluating driving trajectories according to an embodiment of the present invention.

DESCRIPTION OF PREFERRED FORMS AND EMBODIMENTS OF THE INVENTION

The present invention will now be described in detail with reference to the accompanying figures to enable a skilled person to construct and use it. Various modifications to the described embodiments will be immediately apparent to those skilled in the art, and the general principles described may be applied to other embodiments and applications without departing from the scope of the present invention, as defined in the appended claims. Therefore, the present invention should not be considered limited to the described and illustrated embodiments, but should be accorded the broadest scope of protection consistent with the features described and claimed.

Unless otherwise defined, all technical and scientific terms used herein have the same meanings commonly used by persons of ordinary skill in the art pertaining to the present invention. In the event of a conflict, this specification, including the definitions provided, shall prevail. Furthermore, the examples provided are for illustrative purposes only and as such should not be considered limiting.

In particular, the block diagrams included in the attached figures and described below are not to be understood as representing structural characteristics, or construction limitations, but must be interpreted as representing functional characteristics, i.e. intrinsic properties of the devices and defined by the effects obtained, or functional limitations, and which can be implemented in different ways, therefore in such a way as to protect the functionality of the device itself (ability to function).

To facilitate understanding of the embodiments described herein, specific embodiments will be referenced and specific language will be used to describe them. The terminology used herein is intended to describe specific embodiments only and is not intended to limit the scope of the present invention.

Figure 1 shows a block diagram of an electronic vehicle dynamics control system 1 for, or integrated into, a motor vehicle (not illustrated here) according to the present invention. The electronic vehicle dynamics control system 1 is configured to determine, or plan, a driving path (block 3) for the motor vehicle to follow, and to control the operation of one or more components of the motor vehicle 4 to ensure that the latter follows the determined driving path.

In particular, the automotive system 1 comprises a vehicle electronic control unit (ECU) storing, and configured to execute, a driving trajectory planning software. In detail, according to one aspect of the present invention, the driving trajectory planning software, or a module of such software, is also stored and is executable by electronic processing resources external to the vehicle so that the latter, by executing the software, are configured to determine and provide the ECU (not illustrated here) with data, for example one or more trained artificial neural networks, to plan (block 3) driving trajectories.

As regards the planning (block 3) of driving trajectories, it is emphasized that what matters are the operations that must be implemented to achieve this functionality and not the hardware and software architectures with which these operations are implemented, to the point that these could be implemented through a concentrated architecture, i.e. by a single electronic device, or through a distributed cooperative architecture, i.e. distributed among several electronic devices communicating and cooperating with each other according to a proprietary logical architecture that the manufacturer of the electronic vehicle dynamics control system decides to adopt.

Note that, from now on and in the following, the term driving trajectory of a motor vehicle refers to a sequence, or vector, of information indicating the position of the motor vehicle for one or more instants in time.

The driving path planning software is storable in, and executable by, electronic processing resources, in particular including at least the ECU, and is designed so that, when executed, the latter are configured to calculate or determine different driving paths (block 5), to select one of them on the basis of one or more predefined criteria (block 7). Furthermore, the electronic processing resources are configured to control the operation of the motor vehicle to ensure that it follows the selected driving path; in particular, to control the operation of one or more motor vehicle components 4 (for example, a steering system, an engine, a brake or other components not mentioned herein) to ensure that the motor vehicle follows that driving path.

Furthermore, the electronic processing resources are configured to receive automotive quantities provided by an automotive sensor platform 2, preferably comprising a plurality of automotive sensors, designed to detect automotive quantities.

In particular, these vehicle parameters are indicative of the static, and appropriately dynamic, road environment in which the vehicle is located; in particular, they are indicative of the presence and position, and optionally also of additional attributes, of any road obstacles (for example, vehicles, pedestrians, cyclists, and animals), and of information indicative of the dynamics and position of the vehicle. Conveniently, the vehicle parameters also include information relating to the kinematics and position of a number of road obstacles. In particular, the vehicle sensor platform 2 includes sensors for determining the position of the vehicle, for example, Lidar (Light Detection and Ranging) sensors and/or GNSS (Global Navigation Satellite System) sensors.

Furthermore, according to an aspect of the present invention, the electronic processing resources are designed to receive a number of digital road maps of different instants of time even from resources external to the motor vehicle; for example, from a data set, or database, comprising such information.

Furthermore, the electronic processing resources are configured to calculate, or plan, one or more relatively simple driving trajectories based on the received vehicle data and on the basis of geometric and/or kinematic rules; in particular, to ensure that such driving trajectories are kinematically realizable. By way of example and not limitation, the electronic processing resources are configured to calculate relatively simple driving trajectories based on clotoid functions and/or piecewise functions (splines).

According to an embodiment of the present invention, the electronic processing resources are further configured to calculate one or more relatively simple fallback driving trajectories based on the received vehicle data, based on geometric and/or kinematic rules, and based on one or more safety conditions to be satisfied; in particular, wherein the use of the safety conditions allows the calculation of driving trajectories that satisfy by definition one or more predefined safety criteria (for example, but not limited to, to at least prevent the vehicle from colliding with road obstacles). Wherein, in particular, such safety conditions to be satisfied include traffic law compliance conditions and road obstacle collision conditions. The traffic law or driving area compliance conditions are satisfied if a vehicle following the calculated driving trajectory would not violate one or more traffic laws when driving within the relevant driving area; For example, these conditions would not be met if the driving path extended toward a one-way street. However, the conditions for collision with road obstacles are met if a vehicle following the calculated driving path would not collide with any road obstacles. Conveniently, these fallback driving paths are braking paths that must be performed, or followed, in order to avoid colliding with road obstacles or violating one or more traffic regulations.

Figure 2 shows a block diagram of a training method for a driving trajectory planning artificial neural network 10.

The electronic processing resources are configured to implement a planning artificial neural network 10 trained to compute, or plan, one or different driving trajectories that mimic different driving trajectories performed by experienced drivers, whether humans or artificial intelligence autonomous driving systems.

In particular, the electronic processing resources are configured to train the planning artificial neural network 10 to cause the driving trajectories computed by the planning artificial neural network 10 to mimic different driving trajectories previously performed by experienced drivers, and preferably to provide such trained network output. In particular, the planning artificial neural network 10 is trained based on an imitation mechanism, or loss function, 14 that may be deterministic or stochastic; specifically, applying stochasticity to a neural network may extend the ability to sample driving trajectories when the neural network performs inference. In particular, by employing a stochastic imitation mechanism 14, the planning artificial neural network 10 is trained to plan a plurality of driving trajectories for each training epoch.

The electronic processing resources are configured to calculate one or more relatively complex driving trajectories based on the received vehicle data, optionally also based on received digital road maps, and based on a trained artificial neural planning network 10.

According to an embodiment of the present invention, the planning artificial neural network 10 comprises a feature (or entity) embedding layer 11 (for example, the PointNet artificial neural network); wherein, in particular, the feature embedding layer 11 comprises a Multi-Layer Perceptron (MLP) neural network optionally but preferably followed by a number of aggregation operations.

Furthermore, the feature embedding layer 11 is followed by a social embedding layer 12, or social embedding, conveniently comprising a graph neural network (GNN) or a multi-head attention network based on the Self-Attention mechanism.

Further, in turn, the social embedding layer 12 is followed by a decoding layer 13 designed to output a relatively complex driving trajectory based on the output provided by the social embedding layer 12. According to one aspect of the present invention, the decoding layer 13 comprises one of an MLP neural network, a Recurrent Neural Network (RNN), or a different neural network that employs the self-attention mechanism.

In detail, the artificial neural planning network 10 is trained on data indicative of expert driver trajectories collected during test driving maneuvers performed by expert drivers, preferably human drivers, on one or more motor vehicles. Specifically, the data indicative of expert driver trajectories are collected by storing the vehicle parameters received from the vehicle sensor platforms 2 of one or more motor vehicles, conveniently with autonomous driving functions deactivated, with which test maneuvers were performed.

Furthermore, the data indicative of trajectories of expert drivers with which to train the planning artificial neural network 10 include different samples 8, or examples, associated with reference labels 9 (i.e. ground truth) for the prediction.

In particular, a sample 8 of the data is indicative of, and optionally includes, information relating to the static road environment (such as the position of a number of static entities), and conveniently relating to the dynamic road environment, at one or different instants in time; in detail, the information relating to the dynamic road environment includes one or more of the position, kinematics, and possibly other attributes relating to one or different road obstacles and to the vehicle.

Furthermore, in detail, a label 9 is indicative of information relating to the static road environment, and conveniently relating to the dynamic road environment, at one or different moments in time subsequent to the moment in time associated with sample 8 of said label 9.

In particular, a sample 8 is a representation of the road environment and a label 9 is an expert driving trajectory. In greater detail, a sample 8 is a vector of information relating to the road environment and a label 9 is a vector comprising, for different time instants (conveniently in the future with respect to the current time instant), one or different pieces of information indicative of the position and kinematics of the motor vehicle and possibly also the position and kinematics of road obstacles. According to a different aspect of the present invention, such representations of the road environment are two-dimensional maps of the road environment or three-dimensional matrices comprising one or more maps (conveniently two-dimensional) of the road environment at different time instants; wherein, each map of the road environment comprises a number of channels, or pixels, indicative of different pieces of information relating to the perceived road environment.

According to one aspect of the present invention, the electronic processing resources are configured to determine and collect the samples 8 via an automotive sensor platform 2 and to determine the labels 9 associated therewith based on a different precision automotive sensor platform 2 of a motor vehicle with which to perform test maneuvers, comprising precision sensors (e.g., precision GNSS sensors) for detecting labels 9. According to a further aspect of the present invention, the electronic processing resources are designed to determine and collect the labels 9 based on the same automotive sensor platform 2 used to collect the samples 8 of the data indicative of expert driver trajectories.

In more detail, the electronic processing resources are configured to collect a number of labels 9 based on the precision automotive sensory platform 2 and additional labels 9 based on the automotive sensory platform 2 designed to collect samples 8 of data indicative of expert driver trajectories; in this way, it is possible to construct a large data set indicative of expert driver trajectories comprising a plurality of training data and a set of precise data, collected via the precision automotive sensory platform 2, suitable for validation and for carrying out tests relating to the generalization capacity of the neural network used.

Furthermore, the electronic processing resources, whether they are in-vehicle or external to the vehicle, are designed to process the samples 8 of the collected data based on time accumulation techniques, i.e. time buffering, in order to allow efficient management of the sequential or temporal information of the samples 8.

For example, the electronic processing resources process a sample 8 of the data in such a way that the information relating to the road environment understood by the latter is associated with time instants whose extremes differ by a predefined number of seconds, for example two seconds, from each other; that is, a sample 8 is defined by accumulating two seconds of information indicative of the position and kinematics, relating to obstacles and road entities (including the motor vehicle), of time instants in the past with respect to the current time instant.

In particular, the EPRs are also designed to process the labels 9 through non-causal operations, i.e. operations that do not depend on the cause and effect relationships between the data, in order to train the artificial neural planning network 10 with them. In greater detail, the EPRs are designed to smooth, for example using a Butterworth filter, the trajectories of a number of road obstacles and the vehicle present in a representation of the road environment in both temporal directions in order to exclude any possible temporal delays. Furthermore, the EPRs are configured to filter, or associate with each other, the trajectories of the road obstacles and the vehicle and the static entities of the static road environment based on the size of the static entities and the coherence in time and/or space for the driving trajectories, determined subsequently.

According to an embodiment of the present invention optionally independent of other aspects or embodiments of the present invention, the artificial neural planning network 10 is also trained on the basis of an evaluation model 17 (a scoring model) configured to evaluate driving trajectories on the basis of one or more factors, in particular conditions, of evaluation of the road environment; in this way, the artificial neural planning network 10, if trained, is configured to calculate one or more relatively complex driving trajectories also as a function of a number of evaluation scores assigned to such relatively complex driving trajectories or to segments (indicative of at least one position assumed by the motor vehicle at a given moment in time) of the latter.

Specifically, road environment assessment factors are indicative of one or more of the following: collisions with road obstacles, traffic violations (such as extending the trajectory into unauthorized road areas, such as one-way streets), or deviations from an experienced driving trajectory, or ground truth. Optionally, the electronic processing resources are also designed to communicate with other vehicles or road obstacles in the vicinity of the vehicle and to receive data relating to the driving trajectories of other vehicles, both performed and predicted, in order to determine whether the vehicle will collide with or conflict with such vehicles. For example, the data received relating to the driving trajectories of other vehicles includes information indicative of the position and speed of such vehicles at various points in time.

In particular, the electronic processing resources are further configured to implement an evaluation artificial neural network 15 trained to determine an evaluation model 17 for evaluating one or more driving trajectories based on one or more driving trajectory evaluation factors, in particular as a function of the ability of such driving trajectories to comply with one or more traffic regulations and/or to avoid collisions with road obstacles. Wherein, in particular, such evaluation artificial neural network 15 is an encoder-decoder neural network trained to map the produced driving trajectory to an evaluation score; in particular, such evaluation artificial neural network 15 is a convolutional neural network (CNN), and even more specifically, being defined by an encoder-decoder architecture, it is a fully convolutional neural network (FCN), for example. a UNet convolutional neural network.

Figure 3 shows a block diagram of a training method of a driving trajectory evaluation artificial neural network 15 according to an embodiment of the present invention.

Furthermore, the electronic processing resources are configured to determine the evaluation model 17 based on the evaluation artificial neural network 15 trained on the basis of a set of training data 8 and respective labels (not illustrated here); in detail, the evaluation artificial neural network 15 is trained to output the evaluation model 17. The evaluation artificial neural network 15 can be trained using different approaches and methodologies.

In particular, the training dataset 8 comprises one or more maps, for example vectors or images, indicative of the road environment; preferably, representing a bird's-eye view, indicative of the perception of the road environment.

Furthermore, each map in the training dataset includes information indicative of the road environment perceived at one or more time instants, conveniently prior to the current time instant, and optionally also information indicative of the road environment expected at one or more time instants subsequent to the current time instant. Specifically, a map in the training dataset presents one or more trajectories (of the vehicle and road obstacles) and optionally predictions (for future time instants) of trajectories of road obstacles, conveniently vehicles, carried out and transmitted by such obstacles or road vehicles. By way of example and not limitation, a map in this set presents one or more trajectories of past, present, and/or future time instants of road obstacles and the trajectory of the vehicle indicative of its position at the current time instant and, where applicable, at time instants prior to the current one.

In particular, the planning artificial neural network 10, being also trained on the basis of the evaluation model 17, allows for the calculation of relatively complex driving trajectories that are more likely to be selected by the evaluation model 17, to control the dynamics of the vehicle compared to driving trajectories calculated on the basis of an artificial neural network not trained on the basis of the evaluation model 17; in particular, it is desired that such trajectories be able to manage relatively complex behaviors while preserving safety and driving comfort.

In particular, the planning artificial neural network 10 is trained based on a reinforcement learning technique to reward driving trajectory planning, performed by the planning artificial neural network 10, based on evaluation scores determined as a function of the evaluation model 17.

In more detail, the planning artificial neural network 10 is trained using an actor-critic learning technique in order to guide the training of the planning artificial neural network 10 based on one or different evaluation scores assigned to respective analyzed driving trajectories or segments of driving trajectories; wherein, the actor is the planning artificial neural network 10 and the critic, also called state-action variable network, is the evaluation artificial neural network 15. In more detail, the actor-critic technique is used to train the planning artificial neural network 10 (in this case, the actor) based on one or different feedback provided by the evaluation model 17 (here considered to be the critic).

In particular, the underlying assumption of the implemented technique is that the prediction, or planning, horizon of the planning artificial neural network 10 is smaller than the prediction horizon of the evaluation model 17 provided by the evaluation artificial neural network 15; where the prediction horizon is the number of time instants subsequent to the current instant over which the neural network is able to make predictions. Furthermore, it is assumed that no differentiation is necessary between evaluations, or costs, calculated on a driving trajectory that covers a relatively low number of time instants and between evaluations calculated on a different driving trajectory that covers a relatively high number of time instants.

According to one aspect of the present invention, the planning artificial neural network 10 is trained by cyclically, or repeatedly, alternating learning to imitate the driving trajectories performed by expert drivers and reinforcement learning, in particular by rewards, based on the evaluation model 17 produced by the evaluation artificial neural network 15 at a previous training epoch (conveniently the training of the evaluation artificial neural network 15); in particular, in order to mitigate the forgetfulness phenomenon that can occur in the training of artificial neural networks when the latter are trained on new data.

In greater detail, the planning artificial neural network 10 is trained by cyclically performing, for a number of training epochs (or steps), training to imitate the driving trajectories performed by expert drivers and reinforcement learning based on the evaluation model 17 produced by the evaluation artificial neural network 15; in particular, based on the evaluation model 17 produced by the evaluation artificial neural network 15 in a previous epoch, conveniently the closest, to the current training epoch. In particular, the electronic processing resources are also designed to transmit, or make accessible, the prediction performed by the planning artificial neural network 10 to the evaluation artificial neural network 15 after each training epoch, or predefined number of epochs.

The evaluation artificial neural network 15 is trained, in order to determine the evaluation model 17, by cyclically carrying out (for a predefined number of epochs) two different learning processes in an alternating manner; that is, a reward-based learning process, in particular a value-based technique (typical of the actor-critic technique), on the basis of a predefined reward scheme 18 and on the basis of a prediction made by the planning artificial neural network 10, for example a plurality of driving trajectories, at a previous training epoch and a different and second learning process to determine the evaluation model 17. Wherein, in particular, the second learning process is a training process based on a technique for differentiating (block 20), as a function of one or more evaluation factors of driving trajectories, expert driving trajectories compared to non-expert driving trajectories (for example via a maximum margin loss function).

In more detail, the electronic processing resources are designed to train, by modifying the weights of, the evaluation artificial neural network 15, at each training epoch of this network, carrying out a first backpropagation based on the value-based technique and carrying out a second and different backpropagation based on the technique for differentiating (block 20) driving trajectories.

In particular, the electronic processing resources are further designed to transmit the prediction, one or several evaluation scores or the evaluation model 17, performed by the evaluation artificial neural network 15 to the planning artificial neural network 10 after each training epoch, or predefined number of epochs.

In particular, the electronic processing resources are designed to employ reinforcement learning to train the evaluation artificial neural network 15 to determine (or predict), for a number of potential outputs, or actions, that may be performed at a given time, respective rewards associated therewith (for one or several time points subsequent to the current time point) based on a reward mapping, or reward scheme 18; wherein, such reward mapping is designed to correlate a segment of a driving trajectory, indicative of the action the motor vehicle is expected to perform at a given time point, to a reward based on a plurality of predefined driving comfort conditions and/or driving safety conditions.

Furthermore, the electronic processing resources are designed to employ the reward learning technique to cause the evaluation artificial neural network 15 to be or become trained to provide an output, or perform an action, associated with an expected reward that is greater than or equal to the expected rewards of additional, different outputs; that is, so that such output maximizes the expected value of the sum of rewards for time points subsequent to the current time point with respect to additional potential outputs that the evaluation artificial neural network 15 could provide.

Furthermore, the electronic processing resources are designed to employ the technique to differentiate expert driving trajectories (block 20), i.e., driving trajectories that would have assigned a relatively high evaluation score, from non-expert driving trajectories, i.e., driving trajectories that would have assigned a relatively low evaluation score, to cause the evaluation artificial neural network 15 to become trained to evaluate one or more driving trajectories based on one or more evaluation factors. Furthermore, in particular, the evaluation artificial neural network 15 is trained by determining the evaluation model 17, in particular one or more cost maps of the evaluation model 17, based on one or more of the relatively simple driving trajectories 19; in particular, the evaluation artificial neural network 15 is trained by determining the evaluation model 17, in particular one or more cost maps of the evaluation model 17, based on one or more of the relatively simple driving trajectories 19; in particular, the evaluation artificial neural network 15 is trained by determining the evaluation model 17, in particular one or more cost maps of the evaluation model 17, based on one or more of the relatively simple driving trajectories 19; trained using an importance sampling technique by integrating the training data with relatively simple driving trajectories 19 and optionally with one or more driving trajectories followed by the vehicle.

In particular, the EHRs are designed to guide the learning of the network as a function of one or more costs calculated, via an overall loss function, for a set of training data, including one or more maps of the road environment, and for relatively simple driving trajectories 19 (supplemental to the training data). In more detail, the EHRs are designed to train the evaluation artificial neural network 15 by modifying one or more network weights based on one or more residual errors of the neural network, calculated by comparing one or more predictions of the neural network (performed on a subset of the training data) and respective comparison labels; and also based on one or more cost mappings to correlate one or more relatively simple driving trajectories 19 to one or different cost maps of an evaluation model 17.

In more detail, the evaluation artificial neural network 15 is trained to calculate a residual error based on the similarity, or dissimilarity, of the prediction made with respect to the associated label; wherein, this error, and therefore a similarity index, is calculated via a max-margin loss function in order to train the evaluation artificial neural network 15 to associate a relatively high evaluation score for an expert driving trajectory, and to ensure a relatively low evaluation score for an inexperienced driving trajectory. In particular, the evaluation artificial neural network 15 is trained by back-propagating one or more relatively simple guiding trajectories 19 in order to determine the evaluation model 17, in particular one or more cost maps of an evaluation model 17. In detail, the electronic processing resources, in order to train the artificial neural network for evaluation 15, are configured to back-propagate one or more relatively simple trajectories 19 together with the residual errors calculated in such a way as to determine one or more cost maps of an evaluation model 17. In particular, the electronic processing resources are configured to determine whether one or more relatively simple trajectory segments 19 correspond to one or more cells, each associated with a cost, of a cost map of an evaluation model 17 to be modified and output; and if the segments correspond to such cells, such resources are also configured to guide the training of the network also on the basis, conveniently, of the sum of the costs associated with the segments of relatively simple trajectories 19.

Furthermore, according to an aspect of the present invention, the evaluation model 17 produced by the artificial neural evaluation network 15 is a cost mapping, or a volume, or a three-dimensional cost matrix, indicative of the disadvantages associated with following a driving trajectory. In greater detail, in order to determine the evaluation score, the evaluation model 17 is configured to map one or more pieces of information relating to a road environment, and relating to, preferably included in, a driving trajectory to be followed, to one or more costs or disadvantages of following that trajectory.

In greater detail, the evaluation model 17 is a cost volume comprising one or more cost maps (or two-dimensional cost matrices), each comprising a plurality of cells, or channels; in particular, each of these cells is associated with a cost determined on the basis of the evaluation factors of the road environment. In greater detail, each of these cells is associated with a cost relating to a region, at a given moment in time, of a representation of the road environment presenting the driving trajectory to be evaluated; in particular, it is associated with a driving trajectory segment. Furthermore, in particular, the electronic processing resources are designed to associate each trajectory segment to be evaluated with a different cell of a cost map; in which, the cell corresponds, in spatial terms, to the trajectory segment.

Specifically, the electronic processing resources are configured to calculate the evaluation score of a driving trajectory by performing an additive calculation of the cells (associated with a region of the road environment representation), cell by cell, of different cost maps of the evaluation model 17 corresponding to different time points, preferably sequentially. Furthermore, the algorithm for calculating the evaluation score is preferably additive and parallelizable.

In particular, the electronic processing resources are configured to calculate the evaluation score based on a total cost determined by performing an additive calculation of one or more cost map cells corresponding, in spatial terms, to a number of characteristic points of the motor vehicle following the driving trajectory to be evaluated; wherein, such characteristic points of the motor vehicle are identifiable in a region of the representation of the road environment or, in particular, in a segment of the driving trajectory to be evaluated. According to one aspect of the present invention, the electronic processing resources are configured to calculate the evaluation score by performing the additive calculation of one or more cells corresponding to the center of the motor vehicle of different cost maps of different time points. According to a further aspect of the present invention, the number of characteristic points corresponds to the four main vertices, or corners, of the motor vehicle.

According to a preferred embodiment of the present invention, the electronic processing resources are configured to calculate a number of relatively simple guidance trajectories and relatively complex guidance trajectories in order to construct a single set of guidance trajectories, based on the calculated trajectories, designed to cover a wide range of purposes.

In detail, the set of calculated driving trajectories is stored in a tree data structure, conveniently a binary tree structure or a min-heap, in order to reduce the computational cost required to assign (block 6) evaluation scores to the driving trajectories of the set; in more detail, the computational cost required for the iteration of k best evaluated trajectories would correspond to

which would be lower than the cost required to iterate through all n

trajectories

Furthermore, in greater detail, the electronic processing resources are configured to mark (fi ag) the fallback guidance trajectories so that they can be differentiated from the other computed guidance trajectories, allowing them to be extracted from the set of guidance trajectories without the need to employ additional computational resources. Optionally, additional guidance trajectories, both relatively simple and relatively complex, are also marked so that they can be differentiated from the other guidance trajectories.

Furthermore, according to the preferred embodiment of the present invention, the electronic processing resources are configured to assign an evaluation score (block 6), to each of the calculated driving trajectories, based on an evaluation model 17 produced at the end of the training of the evaluation artificial neural network 15; wherein, the evaluation artificial neural network 15 is trained at least on the basis of the technique for differentiating expert driving trajectories from non-expert driving trajectories, optionally it corresponds to the evaluation artificial neural network 15 trained for the execution of the actor-critic technique.

Additionally, the electronic processing resources are configured to map a driving trajectory (whether relatively complex, relatively simple, or a fallback) to an evaluation score based on evaluation model 17.

In detail, the electronic processing resources are configured to sort, in particular in the tree data structure, the driving trajectories calculated on the basis of the evaluation scores assigned to the driving trajectories; in particular, to sort the driving trajectories so that the evaluation scores assigned to them are arranged in decreasing order.

The electronic processing resources are configured to select (block 7) one of the relatively simple and relatively complex driving trajectories calculated on the basis of one or more predefined criteria; in particular, on the basis of the evaluation model 17 and possibly, but preferably, on the basis of one or more different safety conditions to be satisfied. According to a different embodiment of the present invention, the electronic processing resources are configured to assign an evaluation score (block 6) to the relatively complex driving trajectories and to select one of these relatively complex driving trajectories.

Furthermore, in particular, the electronic processing resources are configured to determine whether, for a driving trajectory, one or several safety conditions to be satisfied are satisfied or not satisfied; and to select (block 7) one of the fallback driving trajectories, in the event that one or several safety conditions are not satisfied for a predefined number of driving trajectories; in particular, for the first k best-rated driving trajectories, which are assigned evaluation scores greater than or equal to the evaluation scores assigned to further calculated driving trajectories.

In particular, the electronic processing resources are configured to select (block 7) one of the fallback driving trajectories based on the evaluation scores assigned to them; in greater detail, to select (block 7) a fallback driving trajectory with an evaluation score greater than or equal to the evaluation scores of the other fallback driving trajectories. Furthermore, in particular, the different safety conditions will be satisfied for this fallback driving trajectory by definition.

In more detail, the electronic processing resources are configured to determine that the safety conditions are satisfied for a number k of trajectories that are best estimated from the ordered driving trajectories.

If the safety conditions are determined to be satisfied for a driving path (whether relatively complex or relatively simple), the electronic processing resources are further configured to select (block 7) that driving path and preferably to stop the verification for the further calculated driving paths that are part of the set of driving paths.

Based on the above description, the advantages that the present invention allows to achieve are evident.

In particular, the present driving trajectory planning software enables the planning of safe and comfortable driving trajectories for the behavior of a self-driving vehicle in order to improve the experience of the vehicle's passengers; in particular, the present invention enables the planning of safe and comfortable driving trajectories by combining the calculation of relatively simple driving trajectories with the calculation of relatively complex driving trajectories.

Furthermore, a simple combination of calculating relatively simple driving paths and calculating relatively complex driving paths would hardly lead to an improvement in the safety and comfort of the calculated driving paths; in fact, driving paths calculated on the basis of an artificial intelligence system or algorithm, i.e., relatively complex driving paths, would often be rejected by an evaluator as unsafe.

Claims (10)

1. Driving path planning software for a motor vehicle electronic dynamics control system (1); wherein, the driving path planning software is storable in, and executable from, electronic processing resources and is designed so that, when executed, the latter become configured to: receive automotive quantities provided by an automotive sensor platform (2) designed to detect automotive quantities; implement a planning artificial neural network (10) trained to compute driving trajectories that mimic different driving trajectories performed by expert drivers, or expert driving trajectories; and compute one or more relatively complex driving trajectories based on the received automotive quantities and based on the planning artificial neural network (10); select one of the driving paths calculated on the basis of one or more predefined criteria; and check the operation of the vehicle to ensure that it follows the selected driving path; characterized in that the planning artificial neural network (10) is trained on the basis of an evaluation model (17) configured to evaluate driving trajectories on the basis of one or more driving trajectory evaluation factors; and wherein, the driving trajectory planning software is further designed so that, when executed, the electronic processing resources become configured to implement an evaluation artificial neural network (15) trained to determine an evaluation model (17) by cyclically and alternately performing two different training runs. 2. Driving trajectory planning software according to claim 1, wherein the evaluation artificial neural network (10) is trained by cyclically performing in an alternating manner a reward learning based on a predefined reward scheme (18) and a different and second learning to determine the evaluation model (17). 3. Driving trajectory planning software according to claim 1 or 2, wherein the evaluation artificial neural network (10) is trained by cyclically performing in an alternating manner: reward learning based on a predefined reward scheme (18) and on a prediction made by the planning artificial neural network (10) at a previous training epoch; and training based on a technique for differentiating, based on one or more factors for evaluating driving trajectories, expert driving trajectories from non-expert driving trajectories. 4. Driving trajectory planning software according to any preceding claim, wherein the planning artificial neural network (10) is trained by cyclically performing in an alternating manner: - learning to imitate the driving trajectories performed by expert drivers; and - a reinforcement learning based on the evaluation model (17) produced by the evaluation artificial neural network (15) at a previous training epoch. 5. Driving trajectory planning software according to any preceding claim, and designed so that, when executed, the electronic processing resources become configured to calculate one or more relatively simple driving trajectories based on the received automotive quantities and on the basis of geometric and/or kinematic rules; and wherein the evaluation artificial neural network (15) is trained by determining the evaluation model (17) based on the relatively simple driving trajectories. 6. Driving trajectory planning software according to claim 5, wherein the evaluation artificial neural network (15) is trained by back-propagating one or more of the relatively simple driving trajectories (19) in order to determine, conveniently, one or several cost maps of the evaluation model (17). 7. Driving path planning software according to any preceding claim, and designed so that, when executed, the electronic processing resources become configured to: calculate one or more relatively simple driving trajectories based on the received vehicle data and on the basis of geometric and/or kinematic rules; and select (block 7) one of the driving trajectories based on the evaluation model (17), and possibly based on one or different safety conditions to be satisfied. 8. Driving path planning software according to claim 7, and designed so that, when executed, the electronic processing resources become configured to: calculate one or different fallback driving trajectories based on the received vehicle data, on the basis of geometric and/or kinematic rules and on the basis of one or different safety conditions to be satisfied; determine whether, for a driving path, one or several safety conditions to be satisfied are satisfied or not satisfied; and select (block 7) one of the one or several fallback driving paths on the basis of the evaluation model (17), in case one or several safety conditions are not satisfied for a predefined number of driving paths. 9. Driving trajectory planning software according to any preceding claim, and designed so that, when executed, the electronic processing resources become configured to train the planning artificial neural network (10) and output the trained planning artificial neural network (10). 10. A motor vehicle comprising electronic processing resources in which the driving path planning software according to any of the preceding claims is stored.