Patent Application
20220041182
This application claims priority to European Patent Application Number 20189455.7, filed Aug. 4, 2020, the disclosure of which is hereby incorporated by reference in its entirety herein.
The present disclosure relates to the field of vehicle autonomous driving, and more precisely to a method and a system for training a planning module, typically a motion planning module, for vehicle autonomous driving.
An autonomous driving system can be described by three levels.
The first level relates to the perception. The vehicle is sensing its environment with multiple sensors like cameras, radars and lidars. It can receive additional inputs, for example from the high-definition maps.
Based on the perception information, at a second level, a motion planning module plans a trajectory for the vehicle. This motion planning module comprises sub-modules including a behavior planning module and a trajectory planning module.
The behavior planning module uses algorithms that receive inputs from the perception level and determines high-level decisions for the vehicle, like “change lane to the left/right”, “follow lane with a given speed”, “adjust lane bias to given value”, “accelerate”, or other decision. This behavior planning intends to give a high-level description of how the vehicle is planned to drive.
Instead of the behavior planning module, a tactical or maneuver planning module can be used to give more detailed outputs with additional information, for example: follow lane with a current speed, follow lane with a given speed equal to the current speed increased or decreased by a given speed delta (e.g. 2 m/s), follow lane with a speed equal to a given value (e.g. 32 m/s), follow lane with a given lane bias adjustment (e.g. 0.5 m), change lane to the left with multiple variants like the current speed, a speed change, an acceleration change, a final lane bias, etc.
The trajectory planning module uses algorithms that take the high-level decisions provided by the behavior planning module (or by the tactical or the maneuver planning module) and outputs low-level control signals for the vehicle. The vehicle does not understand the high-level decisions that need to be translated into executable steering instructions. This module generates doable trajectories for the vehicle.
Finally, the autonomous driving system comprises a third level including a control module having the function of controlling the vehicle to keep it on the planned trajectory.
In other words, behavior planning (or tactical planning or maneuver planning) inside the autonomous driving system is responsible for providing high-level tactical decisions, which the autonomous vehicle should eventually take to reach a desired destination in the most efficient way. These high-level decisions are then used by trajectory planning algorithms to generate a safe and doable trajectory for the vehicle and, ultimately, the generated trajectory is executed by a control module of the vehicle.
In order to develop or train the behavior planning module, it is known to use a reinforcement learning method. For that purpose, an agent (vehicle) is put in an environment, for example in a simulation tool. At the beginning, the agent does not know anything about the environment and does not understand the concepts of lane change, speed change, or other concepts. It is allowed to perform an exploration inside the simulation by taking specific actions in an environment in order to maximize a cumulative reward. The agent is continuously trying combinations of different actions until it finds a meaningful policy on how to steer this agent in order to achieve a desired goal, for example to travel while still keeping at the speed limit. However, such a training method of the behavior planning algorithms, based on a simulation tool, may lead to a behavior planning module that is not satisfying because it generates behaviors of the vehicle that are not human-like behaviors and consequently may result in dangerous situations on the road.
In order to improve this drawback, it is known to train the behavior planning by imitation learning. Such a method can combine the reinforcement learning based on training an agent solely in a simulation and a supervised learning. It uses a set of labeled training data. However, this technique has the drawback that it requires to collect a large amount of training data from a large scale fleet of testing vehicles.
There is a need to improve the situation. More precisely, there is a need to provide a method for collecting training data suitable for training a vehicle autonomous driving system in a more human-like manner without requiring important training efforts.
The present disclosure concerns a method of collecting training data suitable for training an autonomous driving system of a vehicle, comprising the steps of: providing a scenario database storing predefined road scenarios, providing user devices with a simulation game for controlling a vehicle agent in a road scenario, at each of a plurality of user devices, running the simulation game, playing a road scenario from the scenario database on a display of the user's device, controlling the vehicle agent in the simulation with steering actions entered by the user, and generating a human demonstration of the played road scenario. The steps further include collecting, as training data, a plurality of human demonstrations of the stored road scenarios played on the plurality of user devices and storing them in a demonstration database.
The use of the imitation game by a plurality of user devices to create human demonstrations of predefined road scenarios allows to easily obtain a large amount of training data suitable to train a planning module of an autonomous driving system of a vehicle, for example by using an imitation learning algorithm or method. Such a method for collecting training data does not require factual vehicle tests for experimenting the road scenarios, that are difficult to achieve, expensive and time consuming, and allows to obtain a massive amount of training data and handle a broad range of road scenarios. It can be easily deployed to a large number of users.
Advantageously, a distribution server selects the road scenarios to be distributed to the user devices.
The method can comprise a step of anonymizing the human demonstrations before transmitting them to the demonstration database.
Each human demonstration of a road scenario generated by a user device can comprise successive samples that each include data indicating a current state on the road in the simulation and, if applicable, data indicating a steering action entered by the user that is currently being operated.
The road scenario can comprise a goal for the vehicle agent, and the human demonstration of the played road scenario can be completed either when the vehicle agent has reached this goal, or when the vehicle agent has missed this goal.
When the road scenario is played by the simulation game, predetermined steering actions to control the vehicle agent can be proposed to the user. These predetermined steering actions can depend on the road scenario presented to the user. They can also be actions that are the responsibility of a planning module to be trained. In this way, the collected training data are more relevant for training this planning module.
The simulation game can have a reward function and calculate a reward information depending on the actions entered by the user. The reward information can be included in the human demonstration of the road scenario that is transmitted to the demonstration database.
The present disclosure also concerns a method of training a planning module for an autonomous driving system of a vehicle, comprising the steps of creating a demonstration database by collecting training data by the method previously defined, and training the planning module by using the human demonstrations stored in the demonstration database.
The planning module can be trained by executing an imitation learning method and using the demonstration database.
The step of training can comprise training at least one of a behavior planning module and a trajectory planning module of a motion planning module for an autonomous driving vehicle.
The present disclosure can also concern a method of producing an autonomous driving vehicle, comprising the steps of training a reference planning module for an autonomous driving system of a vehicle by executing the training method above defined, configuring a planning module to be consistent with the trained reference planning module, providing the autonomous driving vehicle with the configured planning module.
The present disclosure also concerns a system for collecting training data suitable for training a vehicle autonomous driving system, comprising a scenario database storing a plurality of predefined road scenarios, a demonstration database, a plurality of user devices that each runs the simulation game, plays a road scenario from the scenario database on a display, controls a vehicle agent in the simulation with steering actions entered by the user, and generates a human demonstration of the played road scenario, and a collection server that collects, as training data, a plurality of human demonstrations of the stored road scenarios played on the plurality of user devices and storing them in a demonstration database.
The simulation game can be arranged to propose predefined actions to the user to control the vehicle agent, when a road scenario is played, said predefined actions being actions that are the responsibility of the planning module to be trained in autonomous driving.
Other features, purposes and advantages of the disclosure will become more explicit by means of reading the detailed statement of the non-restrictive embodiments made with reference to the accompanying drawings.
In this embodiment, the system 100 comprises a scenario database 10, a demonstration database 20, a collection server 30, a training server 40 and a plurality of user devices 60. The collection server 30 and the training server 40 each have access to the demonstration database 20. The two servers 30, 40, the scenario database 10 and the user devices 60 are all connected to a communication network 200.
The scenario database 10 stores a plurality of predefined road scenarios to be presented to the user's devices 60, for example by downloading or by streaming. A distribution server 50 can be in charge of distributing the road scenarios from the database 10 to the user devices 60. A road scenario corresponds to a situation that takes place on a road and can include a specific goal for a vehicle (called “agent”) controllable by a user. Different types of goal can be defined. A first type of goal can be a lane-based goal, for example to be on a given lane in a given distance. A second type of goal can be related to the use of the speed limit, for example to travel with the highest possible speed. A third type of goal can be related to the traffic flow smoothness, for example to optimize some scene-based criterium like the average speed of all the vehicles on the road. A fourth type of goal can be related to the passenger's comfort, for example to minimize the vehicle's acceleration and jerk. Other types of goal could be defined.
Each user has a simulation game application 61 installed in his user device 60. The user device 60 has a display device 62 and is configured to run the simulation game and play a road scenario from the scenario database 10 on the display device 62. In the simulation game 61, the user can control an agent (a vehicle) in the road scenario that is being presented and played, by entering steering actions that are operated in the simulation. The simulation game application 61 has the additional function of generating a human demonstration of the road scenario that is being played in the simulation. A human demonstration is a set of data that describes how the road scenario does under control of the steering actions entered by the user in the simulation. It can comprise a sequence of samples taken with a predefined frequency (at successive time points), for example 10 samples per second. Each sample includes current information of the simulation, describing the current situation in the road scenario that is being played in the simulation. This current information can include data indicating a current state of the road defined by various pieces of information like the number of lanes, the number of vehicles (other than the vehicle agent), respective positions, speeds and accelerations of the vehicles in the scenario, or other information. It can also include data indicating a steering action entered by the user that is currently being executed in the simulation. The user device 60 can be a smartphone, a tablet, a computer or any other device or terminal.
The demonstration database 20 stores the human demonstrations of the predefined road scenarios stored in the scenario database 10 and that have been presented to and played by user devices 60.
The collection server 30 has the function of collecting these human demonstrations from the user devices 60 and storing the collected human demonstrations into the demonstration database 20.
The training server 40 is for training a planning module of an autonomous driving system for a vehicle.
As previously explained, an autonomous driving system of a vehicle, for example a car, has three levels: a perception level sensing the vehicle's environment with multiple sensors like cameras, radars and lidars, and receiving additional inputs like data from high-definition maps, a motion planning level that plans the trajectory of the vehicle, and a control level that controls the vehicle to follow the planned trajectory.
The data stored in the demonstration database 20 are used by the training server 40 to train algorithm(s) of the motion planning level.
The motion planning level can have two modules. A first module can be a behavior planning module and a second module can be a trajectory planning module. Alternatively, the first module can be a tactical or maneuver planning module.
Schematically, the behavior planning module is responsible for providing high-level driving decisions (e.g. “follow lane”, “change to left/right lane” or “accelerate” or “slow down”, . . . ) based on the inputs provided by the perception level (typically sensor data and map data). When a tactical or maneuver planning module is used instead of the behavior planning module, the high-level driving decisions include additional information that defines more precisely how to drive the vehicle (e.g. “follow lane with current speed (or acceleration)”, “follow lane with the current speed (or acceleration) increased or decreased by a given speed (acceleration) delta”, “follow the lane with a given lane bias adjustment”, “change lane to the left with multiple variants like current speed, a speed (or acceleration) change, a final lane bias, other additional driving information”, “abort the current maneuver”, “emergency braking maneuver”, “emergency braking maneuver with the escape lane change to the right”, other additional driving information).
The trajectory planning module takes the high-level decisions as inputs and is responsible for outputting low level signals for controlling the vehicle so as to generate an executable trajectory. In other words, the trajectory planning module translates the high-level decisions from the behavior (or tactical or maneuver) planning module into executable steering instructions and thus generates a doable trajectory for the vehicle.
The respective responsibilities of the first module and second module (i.e., the behavior (or tactical or maneuver) planning module and the trajectory planning module) can vary. But, from a global point of view, the motion planning module receives inputs from the perception level and outputs steering instructions to control the vehicle so that the latter follows a planned trajectory.
In the present embodiment, the training server 40 has the function of training the behavior (or tactical or maneuver) planning module. The training can be executed by an imitation learning algorithm or method and use the human demonstration data stored in the demonstration database 20 as training data.
The method of collecting training data suitable for training an autonomous driving system of a vehicle will now be described with reference to
The method comprises a step S0 of providing a scenario database 10 storing a plurality of predefined road scenarios and a step S1 of providing each of a plurality of users with a simulation game application 61 for controlling a vehicle agent in a road scenario, the user having a device 60 configured to run the simulation game application 61 installed therein.
In a step S2, a user device 60 runs the simulation game application 61. The simulation game application 61 connects to the scenario database 10 with the user device 60 to receive a predefined road scenario stored therein, for example by download or streaming, in a step S3.
The road scenario presented to the user device 60 can be selected among the plurality of predefined road scenarios stored in the database, either by the user or, advantageously, by the distribution server 50 in charge of distributing the road scenarios to users. The distribution of the predefined road scenarios from the database 10 to the plurality of users 60 should be done so as to collect a massive amount of training data related to a broad range of road scenarios played by a large number of user devices. For example, the distribution server 50 distributes each predefined road scenario of the scenario datable 10 to a number of user devices that is equal or more than a predetermined threshold.
In a step S4, the simulation game application 61 runs on the user device 60 and plays the received road scenario 63 that is displayed on the display device 62, as represented in
In the step S4, the user controls the agent 630 (vehicle) by entering steering actions that are operated in the simulation. A user interface can allow the user to enter predetermined steering actions. For example, it includes a plurality of action buttons displayed on the display device 62, such as the buttons 631a to 631c in
The steering actions of the user could be entered by any other means of user interface. This could be voice commands or tactile commands.
When the road scenario is being played and the user controls the vehicle agent 630 by entering one or more steering actions, the simulation game application 61 generates a human demonstration of the played road scenario, in a step S5. For that purpose, at successive sampling time points, the user device 60 saves current information of the simulation. For each time point, the current information can include data describing the current state of the simulation (number of lanes, number of vehicles, position, speed, acceleration of each vehicle . . . ) and, if applicable, a current steering action entered by the user that is being operated in the simulation. The samples can be created and saved with a predetermined frequency. For example, ten samples are created per second. The sequence of successive samples forms the human demonstration of the played road scenario.
The road scenario played by the user device 60 is completed when the vehicle agent 630 has reached the goal 633 or has missed it. The complete human demonstration is saved by the user device 60.
In a step S6, the collection server 30 collects the (complete) human demonstration of the road scenario presented to the user device 60 and stores it into the demonstration database 20.
The steps S2 to S6 are executed by a plurality of users and for a plurality of different road scenarios, in order to collect, as training data, a plurality of human demonstrations of a plurality of predefined road scenarios. Advantageously, all the predefined road scenarios stored in the scenario database 10 are each presented to different users (for example at least to a minimum number of users), in order to produce a demonstration database containing a sufficient quantity of human demonstrations (training data) from different users for each of the stored road scenarios.
The present disclosure also concerns a method of training a motion planning module designed for an autonomous driving vehicle, including the steps of populating a demonstration database by collecting training data as previously described with reference to
Thus, the demonstration database 20 can be used to train the algorithm(s) of a motion planning module designed for an autonomous driving vehicle, for example a behavior planning module, in a training step S7. More precisely, the planning module can be trained with an imitation learning method and using the demonstration database 20.
The demonstration database 20 could be used to train algorithm(s) of the trajectory planning module of the motion planning module, or other algorithms like those used in a simulation game or tool for driving a vehicle.
In a first variant, the application game application includes an anonymization function that anonymizes the human demonstrations before transmitting them to the demonstration database 20. However, additional contextual data can be collected together with the human demonstrations, such as a geographical information identifying for example a geographical region of the user. This allows to learn robust driving policy capable of working in different regions. Indeed, data is collected from drivers from different world regions. The driving policy can thus be trained on diversified datasets and will therefore behave more robustly.
In a second variant, the application game application includes a reward function that can calculate a reward information depending on the steering actions entered by the user in the simulation of a road scenario. The reward information can be a numerical value. It can be displayed on the display device 62 during the simulation, as shown in
Optionally, the simulation game application 61 can display on the display device 62 various data (or metrics) describing the current state of the vehicle in the simulation (e.g. position, speed, acceleration, jerk, etc.), as shown in
Optionally, the distribution server 50 and the collection server 30 can be fused. The training server 40 can also be fused with the collection server 30 and/or the distribution server 50.
The training using the demonstration database can be performed on a reference planning module (e.g. a behavior planning module). Then, this reference planning module is duplicated to be installed in a fleet of vehicles. In other words, the reference planning module are produced or configured to be consistent with the reference planning module.
The present disclosure also concerns a method of producing an autonomous driving vehicle, comprising the steps of training a reference planning module designed for an autonomous driving vehicle by executing the training method previously described, configuring a planning module to be consistent with the trained reference planning module, and providing the autonomous driving vehicle with the configured planning module
| Number | Date | Country | Kind |
|---|---|---|---|
| 20189455.7 | Aug 2020 | EP | regional |
