Systems and methods for training and controlling an artificial neural network with discrete vehicle driving commands
AssigneesPorsche AG, Audi AG
InventorMuneeb SALEEM
Life of this patent
Estimating the term…
Abstract
Systems, devices, and methodologies are provided for training and controlling a neural network. The neural network is trained using definitive and random training modes to train neurons in a monolithic network. The neural network output is used to control an autonomous or semi-autonomous vehicle.
Claims20
A method for training a monolithic neural network for controlling a vehicle, the method comprising:
- inputting first and second definitive driving commands into the neural network,
- inputting first and second driving maneuvers corresponding to the first and second definitive driving commands into the neural network with the first and second definitive driving commands as part of a definitive training mode,
- means for inducing confusion into the neural network by inputting a unique random command into the neural network along with first and second driving maneuvers so that the neural network controls an autonomous vehicle by identifying open road space.
An vehicle comprising:
- one or more sensors mounted to the vehicle and configured to capture data about the vehicle surroundings in real-time;
- and
- an autonomous or semi-autonomous driving system having end to end neural network having a single trained module operably trained in definitive and neutral commands and configured process the sensed data and determine and output a driving trajectory,
- wherein the driving system executes an automatic driving maneuver of the vehicle smoothly based on the driving trajectory.
A system for automatically and autonomously executing smooth driving maneuvers comprising:
- a vehicle;
- an autonomous driving system integrated into the vehicle for having a command controller that controls and executing smooth driving maneuvers;
- and
- means for continuously determining a trajectory of the vehicle by using neutral command reference data so that an output control command is continuously updated and that the vehicle performs a smooth driving maneuver in response to the continuously updated control command.
A method for training an end-to-end artificial neural network in an vehicle to control an autonomous driving system, the method comprising:
- executing a definitive training mode in which a first defined command and a first driving maneuver is input to a model in the neural network, and a second defined command and a second driving maneuver is input in the neural network;
- and
- executing a random training mode in which plurality neutral commands and a plurality of randomly selected first and second driving maneuvers are input into the model in the neural network along,
- wherein the trained neural network includes a single model made up of a plurality of neurons and the neurons trained in the random mode contribute to the production of a safe and smooth trajectory for the vehicle, which is then input to the vehicle resulting in a smooth execution of the determined driving maneuver.
Dependencies are inferred from the claim text.