Patent Application
20200004255
The present disclosure relates generally to autonomous vehicle control technologies, and particularly to a method for generating validated control commands for an autonomous road vehicle. It also relates to an arrangement for generating validated control commands for an autonomous road vehicle and an autonomous vehicle comprising such an arrangement.
In order to perform travel to some intended destination autonomous road vehicles usually need to process and interpret large amounts of information. Such information may include visual information, e.g. information captured from cameras, information from radars or lidars, and may also include information obtained from other sources, such as from GPS devices, speed sensors, accelerometers, suspension sensors, etc.
During the travel of an autonomous road vehicle decisions needs to be made in real time according to the available information. Such decisions may include decisions to perform braking, acceleration, lane-changing, turns, U-turns, reversing and the like, and the autonomous road vehicle is controlled according to the decision-making result.
WO2017125209 A1 discloses a method for operating a motor vehicle system which is designed to guide the motor vehicle in different driving situation classes in a fully automated manner, wherein - a computing structure comprising multiple analysis units is used to ascertain control data to be used from surroundings data describing the surroundings of the motor vehicle and ego data describing the state of the motor vehicle as driving situation data for guiding the motor vehicle in a fully automated manner and to use the control data in order to guide the motor vehicle, - each analysis unit ascertains output data from output data of at least one other analysis unit and/or driving situation data, and—at least some of the analysis units are designed as a neural net at least partly on the basis of software. At least some of the analysis units designed as a neural net are produced dynamically from a configuration object which can be configured using configuration parameter sets during the runtime, wherein—a current driving situation class is ascertained from multiple specified driving situation classes, each driving situation class being assigned at least one analysis function, using at least some of the driving situation data,—configuration parameter sets assigned to the analysis functions of the current driving situation class are retrieved from a database, and—analysis units which carry out the analysis function and which have not yet been provided are produced by configuring configuration objects using the retrieved configuration parameter sets.
WO2017120336 A2 discloses systems and methods for navigating an autonomous vehicle using reinforcement learning techniques. In one implementation, a navigation system for a host vehicle may include at least one processing device programmed to: receive, from a camera, a plurality of images representative of an environment of the host vehicle; analyze the plurality of images to identify a navigational state associated with the host vehicle; provide the navigational state to a trained navigational system; receive, from the trained navigational system, a desired navigational action for execution by the host vehicle in response to the identified navigational state; analyze the desired navigational action relative to one or more predefined navigational constraints; determine an actual navigational action for the host vehicle, wherein the actual navigational action includes at least one modification of the desired navigational action determined based on the one or more predefined navigational constraints; and cause at least one adjustment of a navigational actuator of the host vehicle in response to the determined actual navigational action for the host vehicle.
US2017357257 A1 discloses a vehicle control method and device and a method and device for obtaining a decision model. The vehicle control method includes the steps of obtaining current external environment information and map information in real time in the running process of an unmanned vehicle; and determining vehicle state information corresponding to the external environment information and the map information which are obtained every time according to the decision model which is obtained through pre-training and embodies the correspondence relationship among the external environment information and the map information and the vehicle state information, and controlling driving states of the unmanned vehicle according to the determined vehicle state information.
It is further known from the publication by Mariusz Bojarski, et al., “End to End Learning for Self-Driving Cars”, arXiv:1604.07316v1 [cs.CV] 25 Apr. 2016, to train a convolutional neural network (CNN) to map raw pixels from a single front-facing camera of an autonomous road vehicle directly to steering commands. With minimum training data from humans a system trained accordingly may learn to drive in traffic on local roads with or without lane markings and on highways. It may also operate in areas with unclear visual guidance, such as in parking lots and on unpaved roads. Such a system may automatically learn internal representations of the necessary processing steps, such as detecting useful road features, with only a human provided steering angle as a training signal.
An object of the present invention is to provide a method and arrangement providing for improved safety when generating control commands for an autonomous road vehicle.
The invention is defined by the appended independent claims. Embodiments are set forth in the appended dependent claims and in the figures.
According to a first aspect there is provided a method for generating validated control commands for an autonomous road vehicle that comprises: providing as input data to an end-to-end trained neural network system raw sensor data from on-board sensors of the autonomous road vehicle as well as object-level data and tactical information data; mapping, by the end-to-end trained neural network system, input data to control commands for the autonomous road vehicle over pre-set time horizons; subjecting the control commands for the autonomous road vehicle over the pre-set time horizons to a safety module arranged to perform risk assessment of planned trajectories resulting from the control commands for the autonomous road vehicle over the pre-set time horizons; validating as safe and outputting from the safety module validated control commands for the autonomous road vehicle.
In a further embodiment the method further comprises adding to the end-to-end trained neural network system a machine learning component.
In a yet further embodiment the method further comprises providing as a feedback to the end-to-end trained neural network system validated control commands for the autonomous road vehicle validated as safe by the safety module.
In a still further embodiment the method comprises providing as raw sensor data at least one of: image data; speed data and acceleration data, from one or more on-board sensors of the autonomous road vehicle.
In an additional embodiment the method further comprises providing as object-level data at least one of: the position of surrounding objects; lane markings and road conditions.
In yet an additional embodiment the method further comprises providing as tactical information data at least one of: electronic horizon (map) information, comprising current traffic rules and road geometry, and high-level navigation information.
According to a second aspect there is provided an arrangement for generating validated control commands for an autonomous road vehicle, that comprises: an end-to-end trained neural network system arranged to receive an input of raw sensor data from on-board sensors of the autonomous road vehicle as well as object-level data and tactical information data; the end-to-end trained neural network system further being arranged to map input data to control commands for the autonomous road vehicle over pre-set time horizons; a safety module arranged to receive the control commands for the autonomous road vehicle over the pre-set time horizons and perform risk assessment of planned trajectories resulting from the control commands for the autonomous road vehicle over the pre-set time horizons; the safety module further being arranged to validate as safe and output validated control commands for the autonomous road vehicle.
In a further embodiment the arrangement further comprises that the end-to-end trained neural network system further comprises a machine learning component.
In a yet further embodiment the arrangement further is arranged to feedback to the end-to-end trained neural network system validated control commands for the autonomous road vehicle validated as safe by the safety module.
In a still further embodiment the arrangement further comprises the end-to-end trained neural network system being arranged to receive as raw sensor data at least one of: image data; speed data and acceleration data, from one or more on-board sensors of the autonomous road vehicle.
In an additional embodiment the arrangement further comprises the end-to-end trained neural network system being arranged to receive as object-level data at least one of: the position of surrounding objects; lane markings and road conditions.
In yet an additional embodiment the arrangement further comprises the end-to-end trained neural network system being arranged to receive as tactical information data at least one of: electronic horizon (map) information, comprising current traffic rules and road geometry, and high-level navigation information.
Also, here envisaged is an autonomous road vehicle that comprises an arrangement for generating validated control commands as set forth herein.
The above embodiments have the beneficial effects of providing an end-to-end solution for improved safety holistic decision making for autonomous road vehicles in complex traffic environments.
In the following, embodiments herein will be described in greater detail by way of example only with reference to attached drawings, in which:
In the following will be described some example embodiments of a method and arrangement 1 for generating validated control commands for an autonomous road vehicle 3. The autonomous road vehicle 3 may be a car, a truck, a bus etc. The autonomous road vehicle 3 may further be a fully autonomous (AD) vehicle or a partially autonomous vehicle with advanced driver-assistance systems (ADAS).
According to the proposed method, as illustrated in
Raw sensor data 5 may be provided as at least one of: image data; speed data and acceleration data, from one or more on-board sensors 6 of the autonomous road vehicle 3. Thus, raw sensor data 5 may for example be images from the autonomous road vehicle's vision system and information about speed and accelerations, such as e.g. obtainable by tapping into the vehicle's Controller Area Network (CAN) or similar. The image data can include images from surround view cameras and also consist of several recent images.
Object-level data 7 may e.g. be provided as at least one of: the position of surrounding objects; lane markings and road conditions.
Tactical information data 8 may e.g. be provided as at least one of: electronic horizon (map) information, comprising current traffic rules and road geometry, and high-level navigation information. Electronic horizon (map) information about current traffic rules and road geometry enables the neural network system 4 to take things such as allowable speed, highway exits, roundabouts, and distances to intersections into account when making decisions.
In order to guide the neural network system 4 in situations when lane-changes, highway exits, or similar actions are needed, high-level navigation information should to be included as input to the neural network system 4. This way, the neural network system 4 will be able to take user preference into account, for example driving to a user specified location.
Raw sensor data 5, such as image data, and object level data 7 do not have to be synchronous. Previously sampled object level data 7 may e.g. be used together with current image data.
The neural network system 4 may either be based on convolutional neural networks (CNNs) or on recurrent neural networks (RNNs). End-to-end training of the neural network 4 is to be done prior to using the neural network 4 in accordance with the proposed method and should be performed end-to-end in either a supervised fashion, where the neural network 4 gets to observe expert driving behavior for a very large amount of possible traffic scenarios and/or by means of reinforcement learning in a simulated environment.
In the supervised case, using a very large and diverse data set will enable a model to become as general as possible in operating range. In a reinforcement setting, the neural network system 4 will be trained in a simulated environment where the neural network system 4 will try to improve good behavior and suppress bad behavior. Doing this in simulations will allow the training of the neural network system 4 to be more exploratory, without imposing risk onto people or expensive hardware.
Convolutional neural networks are particularly suitable for use with the proposed method as features may be learned automatically from training examples, e.g. large labeled data sets may be used for training and validation. Training data may e.g. have been previously collected by vehicles driving on a wide variety of roads and in a diverse set of lighting and weather conditions. CNN learning algorithms may be implemented on massively parallel graphics processing units (GPUs) in order to accelerate learning and inference.
The method further comprises performing, by the end-to-end trained neural network system 4 mapping 17 of the input data, such as raw sensor data 5 from on-board sensors 6 of the autonomous road vehicle 3 as well as object-level data 7 and tactical information data 8, to control commands 10 required to control travel of the autonomous road vehicle 3 over planned trajectories resulting from the control commands 10 for the autonomous road vehicle 3 over the pre-set time horizons. Thus, it is suggested to use a trained neural network 4, such as a convolutional neural network (CNN), to map the input data 5, 7, 8 directly to control commands 10.
The outputs of the end-to-end trained neural network system 4 consists of estimates of control commands 10 over a time horizon. The time horizon should be long enough to enable prediction of the system behavior, e.g., 1 second. This way the neural network system 4 learns to do long-term planning. The outputs may also be fed back into the model of the neural network system 4 to ensure smooth driving commands where the model takes previous decisions into consideration. The method may e.g. comprise automatically setting a control command for speed to an electronic horizon recommended value in case no obstacles are obscuring an intended trajectory and if the road conditions are proper. Otherwise it may comprise estimating a proper speed.
In connection with lateral control commands, the method should preferably provide for performing planned lane-changes, e.g. changing to the appropriate lane for turning in intersections at an early stage. As such, one of the outputs of the system should, in such cases, be a lane-changing signal 13. This lane-changing signal 13 should be fed back into the neural network system 4 in a recurrent fashion to allow for smooth overtakes etc.
Further in accordance with the proposed method a safety module 9 is subjected 18 to the control commands 10 for the autonomous road vehicle 3 over the pre-set time horizons. Risk assessment 19 of planned trajectories resulting from the control commands 10 for the autonomous road vehicle over the pre-set time horizons is performed by the safety module 9. Control commands 2 to be used to control travel of the autonomous road vehicle 3 are validated as safe, illustrated by the Y option in
In order to enhance the planning capability from predicted system behavior some embodiments of the method include adding to the end-to-end trained neural network system 4 a machine learning component 11. Such a machine learning component 11 adds to the method the ability to “learn”, i.e., progressively improve the performance from new data, without being explicitly programmed. Another alternative for this is through more advanced methods, such as value-iteration networks.
In some further embodiments the method further comprises providing as a feedback 12 to the end-to-end trained neural network system 4 validated control commands 2 for the autonomous road vehicle 3 validated as safe by the safety module 9. In this way it becomes possible to further train the neural network system 4 to avoid unfeasible or unsafe series of control commands. In such embodiments the safety module 9 must be available for system training.
Thus, provided hereby is a method that can act as an end-to-end solution for holistic decision making for an autonomous road vehicle 3 in complex traffic environments. The method operates on rich sensor information 5, 7, 8, allowing proper decisions to be made and planning for what control commands 2 to take for future time horizons.
The proposed arrangement 1 for generating validated control commands 2 for an autonomous road vehicle 3 comprises an end-to-end trained neural network system 4 arranged to receive an input of raw sensor data 5 from on-board sensors 6 of the autonomous road vehicle 3 as well as object-level data 7 and tactical information data 8.
The end-to-end trained neural network system 4 of the proposed arrangement 1 for generating validated control commands for an autonomous road vehicle 3 may further be arranged to receive as raw sensor data 5 at least one of: image data; speed data and acceleration data, from one or more on-board sensors 6 of the autonomous road vehicle 3. The raw sensor data 5 may for example be images from vision systems of the autonomous road vehicle 3 and information about speed and accelerations thereof, e.g. obtained by tapping into the autonomous road vehicle's Controller Area Network (CAN) or similar. The image data can include images from surround view cameras and also consist of several recent images.
The end-to-end trained neural network system 4 may still further be arranged to receive as object-level data 7 at least one of: the position of surrounding objects; lane markings and road conditions.
Furthermore, the end-to-end trained neural network system 4 may be arranged to receive as tactical information data 8 at least one of: electronic horizon (map) information, comprising current traffic rules and road geometry, and high-level navigation information. Electronic horizon (map) information about current traffic rules and road geometry enables the end-to-end trained neural network system 4 to take things such as allowable speed, highway exits, roundabouts, and distances to intersections into account when making decisions.
In order to guide the end-to-end trained neural network system 4 in situations when lane-changes, highway exits, or similar actions are needed, the system can be arranged to include as input to the end-to-end trained neural network system 4 high-level navigation information. This way, the end-to-end trained neural network system 4 will be able to take user preference into account, for example driving to a user specified location.
The end-to-end trained neural network system 4 should either be based on convolutional neural networks (CNNs) or on recurrent neural networks (RNNs). End-to-end training of the neural network 4 should have been done prior to using the neural network 4 in the proposed arrangement 1 and should have been performed end-to-end in either a supervised fashion, where the neural network 4 has been allowed to observe expert driving behavior for a very large amount of possible traffic scenarios and/or by means of reinforcement learning in a simulated environment.
In the supervised case, using a very large and diverse data set will enable a model to become as general as possible in operating range. In a reinforcement setting, the system will be trained in a simulated environment where the system will try to improve good behavior and suppress bad behavior. Doing this in simulations will allow the training of the system to be more exploratory, without imposing risk onto people or expensive hardware.
Convolutional neural networks are particularly suitable for use with the proposed arrangement as it enables features to be learned automatically from training examples, e.g. by using large labeled data sets for training and validation thereof. Training data may e.g. have been previously collected by vehicles driving on a wide variety of roads and in a diverse set of lighting and weather conditions. Training may alternatively be based on a pre-defined set of classes, e.g. overtaking, roundabouts etc. but also on a “general” class that captures all situations that are not explicitly defined. CNN learning algorithms may be implemented on massively parallel graphics processing units (GPUs) in order to accelerate learning and inference.
The end-to-end trained neural network system 4 should further be arranged to map input data, such as raw sensor data 5 from on-board sensors 6 of the autonomous road vehicle 3 as well as object-level data 7 and tactical information data 8, to control commands 10 required to control travel of the autonomous road vehicle 3 over planned trajectories resulting from the control commands 10 for the autonomous road vehicle 3 over the pre-set time horizons. Thus, it is suggested to use a trained neural network 4, such as a convolutional neural network (CNN), to map the input data 5, 7, 8 directly to control commands 10.
The end-to-end trained neural network system 4 outputs consists of estimates of control commands 10 over a time horizon. The time horizon should be long enough to be able to predict the system behavior, e.g., 1 second. This way the neural network system 4 learns to do long-term planning. The outputs 10 may also be fed back into the model of the neural network system 4 as to ensure smooth driving commands where the model takes previous decisions into consideration. A control command for speed may e.g. be automatically set by the neural network system 4 to an electronic horizon recommended value in case no obstacles are obscuring an intended trajectory and if the road conditions are proper. Otherwise a proper speed may be estimated by the neural network system 4.
In connection with lateral control commands, the neural network system 4 should preferably be able to perform planned lane-changes, e.g. changing to the appropriate lane for turning in intersections at an early stage. As such, one of the outputs of the system should be a lane-changing signal 13. This lane-changing signal 13 should be fed back into the system in a recurrent fashion to allow for smooth overtakes etc. At the bottom right of
A safety module 9 is arranged to receive the control commands 10 for the autonomous road vehicle 3 over the pre-set time horizons and perform risk assessment of planned trajectories resulting from the control commands 10 for the autonomous road vehicle 3 over the pre-set time horizons. The safety module 9 is also arranged to validate as safe control commands 2 to be used to control travel of the autonomous road vehicle 3 and thereafter output such validated control commands 2 to control travel of the autonomous road vehicle 3. These validated control commands 2 may e.g. be output to a vehicle control module 15. The safety module 9 may also be arranged to receive raw sensor data 5 and object-level data 7.
To enhance the planning capability from predicted system behavior the end-to-end trained neural network system 4 of the arrangement 1 may further comprise a machine learning component 11. Such a machine learning component 11 adds to the arrangement 1 the ability to “learn”, i.e., progressively improve the performance from new data, without being explicitly programmed. Another alternative for this is through more advanced methods, such as value-iteration networks.
The arrangement 1 may further be arranged to feedback 12 to the end-to-end trained neural network system 4 validated control commands 2 for the autonomous road vehicle 3 validated as safe by the safety module 9. This makes it possible to further train the neural network system 4 to avoid unfeasible or unsafe series of control commands. In such embodiments the safety module 9 must be available for system training.
Thus, provided hereby is an arrangement 1 that can act as an end-to-end solution for holistic decision making for an autonomous road vehicle 3 in complex traffic environments and in any driving situation. The neural network system 4 thereof is arranged to operate on rich sensor information 5, 7, 8, allowing it to make proper decisions and plan for what control commands 2 to take for future time horizons. The neural network system 4 of the arrangement 1 may consist of only one learned neural network for generating validated control commands 2 for performing driving of an autonomous road vehicle 3.
Also, here envisaged is an autonomous road vehicle 3 that comprises an arrangement 1 as set forth herein.
| Number | Date | Country | Kind |
|---|---|---|---|
| 18180826.2 | Jun 2018 | EP | regional |
