Method and system for predictive control of vehicle using digital images

Patent metadata

Field	Value
Publication number	US-11934191-B2
Application number	US-202016921523-A
Country	US
Kind code	B2
Filing date	Jul 6, 2020
Priority date	Jul 5, 2019
Publication date	Mar 19, 2024
Grant date	Mar 19, 2024

Abstract

Methods and systems for predictive control of an autonomous vehicle are described. Predictions of lane centeredness and road angle are generated based on data collected by sensors on the autonomous vehicle and are combined to determine a state of the vehicle that are then used to generate vehicle actions for steering control and speed control of the autonomous vehicle.

First claim

The invention claimed is:

1. A method for predictive control of an autonomous vehicle, the method comprising:
receiving a digital image representing an environment of the autonomous vehicle and vehicle data representing a speed of the vehicle;
determining a current state of the autonomous vehicle and the environment of the autonomous vehicle, the current state being a set of data that includes the digital image and the vehicle data;
generating, using a respective set of lane centeredness general value functions (GVFs), a first set of predictions representing future lane centeredness of the vehicle over a set of respective multiple future time horizons, the first set of predictions being generated by the respective set of lane centeredness GVFs based on only the digital image and the vehicle data in the current state;
generating, using a respective set of road angle GVFs, a second set of predictions representing future road angle of the vehicle over the same or different set of respective multiple future time horizons, the second set of predictions being generated by the respective set of road angle GVFs based on only the digital image and the vehicle data in the current state;
wherein the respective set of lane centeredness GVFs is implemented using at least one neural network and maps the digital image and the vehicle data in the current state to the first set of predictions over the set of respective multiple future time horizons;
wherein the respective set of road angle GVFs is implemented using the same at least one neural network or at least one other neural network and maps the digital image and the vehicle data in the current state to the second set of predictions over the same or different set of respective multiple future time horizons; and
generating, based on the first set of predictions over the set of respective multiple future time horizons and the second set of predictions over the same or different set of respective multiple future time horizons, a vehicle action.

2. The method of claim 1 , further comprising learning the respective set of lane centeredness GVFs and the respective set of road angle GVFs by:
generating the first set of predictions by the respective set of lane centeredness GVFs based on the current state at a current time step and generating the second set of predictions by the respective set of road angle GVFs based on the current state at the current time step;
generating the vehicle action based on the first and second sets of predictions;
executing the vehicle action and sampling a next state at a next time step;
computing a cumulant based on the current state, the executed vehicle action and the next state; and
updating the respective set of lane centeredness GVFs and the respective set of road angle GVFs implemented based on the cumulant.

3. The method of claim 1 , further comprising learning the respective set of lane centeredness GVFs and the respective set of road angle GVFs by:
receiving a dataset containing vehicle action, digital images and vehicle data at respective time steps;
determining a state at each respective time step that includes the digital image and vehicle data at each respective time step; and
updating the respective set of lane centeredness GVFs and the respective set of road angle GVFs based on cumulants computed using the vehicle action and state at each respective time step.

4. The method of claim 1 , wherein the generated vehicle action is one of a steering control action to change a steering angle of the vehicle, a speed control action to change a target speed of the vehicle, and a steering and speed control action to change both a steering angle and a target speed of the vehicle.

5. The method of claim 1 , wherein the vehicle action is generated by a predefined proportional-integral-derivative (PID) controller.

6. The method of claim 1 , wherein the vehicle action is generated by a controller using reinforcement learning (RL).

7. The method of claim 1 , wherein the respective set of lane centeredness GVFs and the respective set of road angle GVFs are implemented together as a single neural network.

8. A vehicle control system for controlling an autonomous vehicle, the vehicle control system comprising:
a processor system configured to execute instructions of a predictive control system to cause the predictive control system to:
receive a digital image representing an environment of the autonomous vehicle and vehicle data representing speed of the vehicle;
determine a current state of the autonomous vehicle and the environment of the vehicle, the current state being a set of data that includes the digital image and the vehicle data;
generate, using a respective set of lane centeredness general value functions (GVFs), a first set of predictions representing future lane centeredness of the vehicle over a set of respective multiple future time horizons, the first set of predictions being generated by the respective set of lane centeredness GVFs based on only the digital image and the vehicle data in the current state;
generate, using a respective set of road angle GVFs, a second set of predictions representing future road angle of the vehicle over the same or different set of respective multiple future time horizons, the second set of predictions being generated by the respective set of road angle GVFs based on only the digital image and the vehicle data in the current state;
wherein the respective set of lane centeredness GVFs is implemented using at least one neural network that and maps the digital image and the vehicle data in the current state to the first set of predictions over the set of respective multiple future time horizons;
wherein the respective set of road angle GVFs is implemented using the same at least one neural network or at least one other neural network and maps the digital image and the vehicle data in the current state to the second set of predictions over the same or different set of respective multiple future time horizons; and
generate, based on the first set of predictions over the set of respective multiple future time horizons and the second set of predictions over the same or different set of respective multiple future time horizons, a vehicle action.

9. The vehicle control system of claim 8 , wherein the processor system is further configured to execute instructions to learn the respective set of lane centeredness GVFs and the respective set of road angle GVFs by:
generating the first set of predictions by the respective set of lane centeredness GVFs based on the current state at a current time step and generating the second set of predictions by the respective set of road angle GVFs based on the current state at the current time step;
generating the vehicle action based on the first and second sets of predictions;
executing the vehicle action and sampling a next state at a next time step;
computing a cumulant based on the current state, the executed vehicle action and the next state; and
updating the respective set of lane centeredness GVFs and the respective set of road angle GVFs implemented based on the cumulant.

10. The vehicle control system of claim 8 , wherein the processor system is further configured to execute instructions to learn the respective set of lane centeredness GVFs and the respective set of road angle GVFs by:
receiving a dataset containing vehicle actions, digital images, and vehicle data at respective time steps;
constructing a state at each respective time step that includes the digital image and vehicle data at each respective time step; and
updating the respective set of lane centeredness GVFs and the respective set of road angle GVFs based on cumulants computed usi

Assignees

• Graves Daniel Mark
• Huawei Tech Co Ltd

Inventors

• Graves Daniel Mark

Classifications

• G05D1/021 Primary

  specially adapted to land vehicles · CPC title

• B60W30/0956

  the prediction being responsive to traffic or environmental parameters · CPC title

• B60W50/0097

  Predicting future conditions · CPC title

• B60W30/12 Primary

  Lane keeping · CPC title

• B60W2520/10

  Longitudinal speed · CPC title