Trajectory tracking for vehicle lateral control using neural network

• US 10,737,717 B2
• Filed: 02/14/2018
• Issued: 08/11/2020
• Est. Priority Date: 02/14/2018
• Status: Active Grant

• Current Assignee: GM Global Technology Operations LLC (General Motors Company)
• Original Assignee: GM Global Technology Operations LLC (General Motors Company)
• Inventors: Peng, Farui, Litkouhi, Bakhtiar B.
• Primary Examiner(s): Black, Thomas G
• Assistant Examiner(s): Lewandroski, Sara J
• Application Number:
  US15/896,352
• Publication Number:
  US 20190248411A1
• Time in Patent Office: 909 Days
• CPC Class Codes: B62D 15/025 Active steering aids, e.g. helping the driver by actively influencing the steering system after environment evaluation B62D1/28 takes precedence
• CPC Class Codes: B62D 6/002 computing target steering angles for front or rear wheels B62D7/159 takes precedence
• CPC Class Codes: G01S 19/14 specially adapted for specific applications
• CPC Class Codes: G01S 19/45 by combining measurements of signals from the satellite radio beacon positioning system with a supplementary measurement
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent

Abstract

A processor-implemented method in a vehicle for performing automatic vehicle lateral control using a neural network for trajectory tracking is provided. The method includes receiving a desired vehicle trajectory; constructing, from the desired trajectory, desired waypoint data for a small number of desired waypoints at different look ahead distances in front of the center of gravity of the vehicle; and generating a steering angle command using a feedforward artificial neural network (ANN) as a function of vehicle speed and the desired waypoint data without modeling vehicle or tire dynamics. The vehicle steering actuator system provides steering control to cause vehicle steering to attempt to achieve the steering angle command.

Claims

1. A processor-implemented method in a real world vehicle for performing automatic vehicle lateral control using a neural network for trajectory tracking, the method comprising:

receiving a desired trajectory for the real world vehicle;

constructing, from the desired trajectory, desired waypoint data for less than ten desired waypoints at different look ahead distances in front of the center of gravity of the real world vehicle, wherein the constructed desired waypoint data comprises a relative lateral position and heading with respect to a vehicle centered coordinate system for each waypoint; and

generating a steering angle command for a steering system in the real world vehicle from a feedforward artificial neural network (ANN) that has no more than one hidden layer, the ANN having been trained using real world training data collected while the real world vehicle or a similar type of vehicle was traveling in a projected travel path of the real world vehicle by;

collecting real world training data comprising real world vehicle position data in a GPS coordinate system, real world vehicle heading data, real world vehicle speed data, and real world driver steering angle data, wherein the real world training data was collected while the real world vehicle or a similar type of vehicle was manually driven along straight roads, intersections, curvy roads, roundabouts, and while making U-turns;

converting the collected GPS position data to a vehicle centered coordinate system;

labeling the real world training data; and

computing and correcting the ANN model based on an error metric to determine a final weight for each node in the ANN;

wherein vehicle speed and the desired waypoint data for less than ten waypoints are input to the ANN, the steering angle command is output from the ANN, the ANN does not model vehicle dynamics, and the ANN does not model tire dynamics.

2. The method of claim 1, wherein the ANN can be used for vehicle lateral control with a range of vehicle speeds that includes a zero speed.

3. The method of claim 1, wherein the constructing desired waypoint data includes calculating less than ten waypoints at different look ahead distances in front of the center of gravity of the vehicle from the desired trajectory wherein the desired trajectory comprises more than ten waypoints or a desired polynomial equation.

4. The method of claim 3, wherein the constructing desired waypoint data includes calculating less than six waypoints.

5. The method of claim 1, wherein the ANN has been trained using back propagation.

6. The method of claim 1, wherein the ANN comprises no more than a single output node.

7. A non-transitory computer readable storage medium embodying programming instruction for performing a method in a real world vehicle, the method comprising:

receiving a desired trajectory for the real world vehicle;

constructing, from the desired trajectory, desired waypoint data for less than ten desired waypoints at different look ahead distances in front of the center of gravity of the real world vehicle, wherein the constructed desired waypoint data comprises a relative lateral position and heading with respect to a vehicle centered coordinate system for each waypoint; and

generating a steering angle command for a steering system in the real world vehicle from a feedforward artificial neural network (ANN) that has no more than one hidden layer, the ANN having been trained using real world training data collected while the real world vehicle or a similar type of vehicle was traveling in a projected travel path of the real world vehicle by;

collecting real world training data comprising real world vehicle position data in a GPS coordinate system, real world vehicle heading data, real world vehicle speed data, and real world driver steering angle data, wherein the real world training data was collected while the real world vehicle or a similar type of vehicle was manually driven along straight roads, intersections, curvy roads, roundabouts, and while making U-turns;

converting the collected GPS position data to a vehicle centered coordinate system;

labeling the real world training data; and

computing and correcting the ANN model based on an error metric to determine a final weight for each node in the ANN;

wherein vehicle speed and the desired waypoint data for less than ten waypoints are input to the ANN, the steering angle command is output from the ANN, the ANN does not model vehicle dynamics, and the ANN does not model tire dynamics.

8. The non-transitory computer readable storage medium claim 7, wherein the ANN has been trained using back propagation.

9. The non-transitory computer readable storage medium claim 7, wherein the ANN comprises no more than a single output node.

10. The non-transitory computer readable storage medium claim 7, wherein the constructing desired waypoint data includes calculating less than ten waypoints at different look ahead distances in front of the center of gravity of the vehicle from the desired trajectory wherein the desired trajectory comprises more than ten waypoints or a desired polynomial equation.

11. The non-transitory computer readable storage medium claim 7, wherein the constructing desired waypoint data includes calculating less than six waypoints.

12. The non-transitory computer readable storage medium claim 7, wherein the ANN can be used for vehicle lateral control with a range of vehicle speeds that includes a zero speed.

13. A steering command generation module in an autonomously or semi-autonomously driven real world vehicle, the steering command generation module comprising one or more processors configured by programming instructions encoded on non-transient computer readable media, the steering command generation module configured to:

receive a desired trajectory for the real world vehicle;

construct, from the desired trajectory, desired waypoint data for less than ten desired waypoints at different look ahead distances in front of the center of gravity of the real world vehicle, wherein the constructed desired waypoint data comprises a relative lateral position and heading with respect to a vehicle centered coordinate system for each waypoint; and

generate a steering angle command for a steering system in the real world vehicle from a feedforward artificial neural network (ANN) that has no more than one hidden layer, the ANN having been trained using real world training data collected while the real world vehicle or a similar type of vehicle was traveling in a projected travel path of the real world vehicle by;

collecting real world training data comprising real world vehicle position data in a GPS coordinate system, real world vehicle heading data, real world vehicle speed data, and real world driver steering angle data, wherein the real world training data was collected while the real world vehicle or a similar type of vehicle was manually driven along straight roads, intersections, curvy roads, roundabouts, and while making U-turns;

converting the collected GPS position data to a vehicle centered coordinate system;

labeling the real world training data; and

computing and correcting the ANN model based on an error metric to determine a final weight for each node in the ANN;

wherein vehicle speed and the desired waypoint data for less than ten waypoints are input to the ANN, the steering angle command is output from the ANN, the ANN does not model vehicle dynamics, and the ANN does not model tire dynamics.

14. The steering command generation module of claim 13, wherein the ANN can be used for vehicle lateral control with a range of vehicle speeds that includes a zero speed.

15. The steering command generation module of claim 13, wherein the ANN has been trained using back propagation.

16. The steering command generation module of claim 13, wherein vehicle speed and the desired waypoint data for less than six waypoints are input to the ANN.

17. The steering command generation module of claim 13, wherein the ANN comprises no more than a single output node.

18. The steering command generation module of claim 13, wherein constructing desired waypoint data includes calculating less than ten waypoints at different look ahead distances in front of the center of gravity of the vehicle from the desired trajectory wherein the desired trajectory comprises more than ten waypoints or a desired polynomial equation.