Decision-making for autonomous vehicle

Patent metadata

Field	Value
Publication number	US-2023391362-A1
Application number	US-202318452449-A
Country	US
Kind code	A1
Filing date	Aug 18, 2023
Priority date	Aug 31, 2022
Publication date	Dec 7, 2023
Grant date	—

Abstract

The disclosed techniques include: obtaining traveling state information of the autonomous vehicle; obtaining, in response to confirming that there is a traffic object within a preset range for the autonomous vehicle, true motion state information of the traffic object; determining a first control decision for indicating whether the autonomous vehicle is to avoid the traffic object based on the traveling state information of the autonomous vehicle and the true motion state information of the traffic object; predicting first predictive motion state information of the traffic object within a first preset duration based on the true motion state information of the traffic object and the first control decision; and determining a second control decision for the autonomous vehicle based on the traveling state information of the autonomous vehicle and the first predictive motion state information of the traffic object.

First claim

1 . A decision-making method for an autonomous vehicle, comprising:
detecting a traveling state of the autonomous vehicle to obtain travel state information of the autonomous vehicle; detecting, in response to detecting that there is a traffic object within a first range for the autonomous vehicle, a motion state of the traffic object to obtain a true motion state information of the traffic object; determining a first control decision based on the traveling state information of the autonomous vehicle and the true motion state information of the traffic object, wherein the first control decision indicates whether the autonomous vehicle is to avoid the traffic object; predicting first predictive motion state information of the traffic object within a first duration based on the true motion state information of the traffic object and the first control decision; and determining a second control decision for the autonomous vehicle based on the traveling state information of the autonomous vehicle and the first predictive motion state information of the traffic object.

2 . The method according to claim 1 , further comprising:
predicting second predictive motion state information of the traffic object within a second duration based on the true motion state information of the traffic object and the second control decision; and determining a third control decision for the autonomous vehicle based on the traveling state information of the autonomous vehicle and the second predictive motion state information of the traffic object.

3 . The method according to claim 1 , wherein the detecting the motion state of the traffic object includes detecting the motion state of one or more of a pedestrian, a non-motor vehicle or a motor vehicle.

4 . The method according to claim 1 , wherein the determining the first control decision based on the traveling state information of the autonomous vehicle and the true motion state information of the traffic object comprises:
inputting the traveling state information of the autonomous vehicle and the true motion state information of the traffic object into an automatic driving decision-making model to obtain the first control decision output by the automatic driving decision-making model, and wherein, the determining the second control decision for the autonomous vehicle based on the traveling state information of the autonomous vehicle and the first predictive motion state information of the traffic object comprises: inputting the traveling state information of the autonomous vehicle and the first predictive motion state information of the traffic object into the automatic driving decision-making model to obtain the second control decision output by the automatic driving decision-making model.

5 . The method according to claim 1 , wherein the predicting the first predictive motion state information of the traffic object within the first duration based on the true motion state information of the traffic object and the first control decision comprises:
inputting the true motion state information of the traffic object and the first control decision into a motion state prediction model to obtain the first predictive motion state information output by the motion state prediction model.

6 . The method according to claim 1 , further comprising:
obtaining a traveling scenario information, wherein the traveling scenario information indicates a motion restriction information for the autonomous vehicle and the traffic object in a traveling scenario, and wherein, the determining the first control decision based on the traveling state information of the autonomous vehicle and the true motion state information of the traffic object comprises: determining the first control decision based on the traveling state information of the autonomous vehicle, the true motion state information of the traffic object and the traveling scenario information, and wherein, the predicting the first predictive motion state information of the traffic object within the first duration based on the true motion state information of the traffic object and the first control decision comprises:
predicting the first predictive motion state information of the traffic object within the first duration based on the true motion state information of the traffic object, the first control decision and the traveling scenario information.

7 . The method according to claim 4 , wherein the predicting the first predictive motion state information of the traffic object within the first duration based on the true motion state information of the traffic object and the first control decision comprises:
inputting the true motion state information of the traffic object and the first control decision into the motion state prediction model to obtain the first predictive motion state information output by the motion state prediction model, and wherein, the automatic driving decision-making model and the motion state prediction model is trained through following actions:
obtaining first sample traveling state information of the autonomous vehicle and first sample motion state information of the traffic object;
inputting the first sample traveling state information of the autonomous vehicle and the first sample motion state information of the traffic object into the decision-making model for the autonomous vehicle to obtain a first predictive control decision output by the decision-making model for the autonomous vehicle, wherein the decision-making model for the autonomous vehicle is configured to obtain the first predictive control decision through the following actions:
inputting the first sample traveling state information of the autonomous vehicle and the first sample motion state information of the traffic object into the automatic driving decision-making model to obtain a fourth control decision output by the automatic driving decision-making model, wherein the fourth control decision indicates whether the autonomous vehicle is to avoid the traffic object;
inputting the first sample motion state information of the traffic object and the fourth control decision into the motion state prediction model to obtain third predictive motion state information of the traffic object within a third duration output by the motion state prediction model; and
inputting the first sample traveling state information of the autonomous vehicle and the third predictive motion state information of the traffic object into the automatic driving decision-making model to obtain the first predictive control decision output by the automatic driving decision-making model;
evaluating the first predictive control decision in a simulation scenario in which the first predictive control decision is executed by the autonomous vehicle to obtain an evaluation result of the first predictive control decision for the simulation scenario, wherein the evaluation result indicates whether a traveling behavior of the autonomous vehicle executing the first predictive control decision meets a threshold; and
adjusting parameters of the decision-making model for the autonomous vehicle based on the evaluation result.

8 . The method according to claim 7 , wherein the actions further comprise:
determining a label for the first predictive control decision based on the evaluation result, wherein the label indicates whether the first predictive control decision is a positive sample or a negative sample; obtaining second sample traveling state information of the autonomous vehicle and second sample motion state information of the traffic object when the autonomous vehicle executes the first predictive control decision in the simulation scenario; inputting the second sample traveling state information of the autonomous vehicl

Assignees

• Apollo Intelligent Driving Technology Beijing Co Ltd

Inventors

• Liu Changchun
• Su Yabo
• Pan Yifeng
• Li Yixian
• Deng Han
• Yang Jing
• Wang Chao
• Zhu Zhenguang
• Peng Liang

Classifications

• B60W60/001 Primary

  Planning or execution of driving tasks · CPC title

• B60W30/0956

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

• B60W2554/4029

  Pedestrians · CPC title

• B60W60/0011 Primary

  involving control alternatives for a single driving scenario, e.g. planning several paths to avoid obstacles · CPC title

• B60W60/00272

  relying on extrapolation of current movement · CPC title