Adversarial scenarios for safety testing of autonomous vehicles

Patent metadata

Field	Value
Publication number	US-11977386-B2
Application number	US-202218057079-A
Country	US
Kind code	B2
Filing date	Nov 18, 2022
Priority date	Jun 10, 2020
Publication date	May 7, 2024
Grant date	May 7, 2024

Abstract

Techniques to generate driving scenarios for autonomous vehicles characterize a path in a driving scenario according to metrics such as narrowness and effort. Nodes of the path are assigned a time for action to avoid collision from the node. The generated scenarios may be simulated in a computer.

First claim

What is claimed is:

1. A method for modifying the behavioral logic of an autonomous vehicle, the method comprising:
assigning a narrowness metric to each of one or more paths in a driving scenario, a path comprising a sequence of driving states of the autonomous vehicle;
wherein the narrowness metric for a particular state in a path is determined based on a number of child states of the particular state;
assigning an effort metric to the paths;
assigning an amount of time for action by the autonomous vehicle to avoid collision along the paths; and
updating the behavioral logic of the autonomous vehicle based on a scoring of the paths determined from an inverse of the narrowness metric, the effort metric, and the time for action.

2. The method of claim 1 , wherein each node of the paths is assigned a node effort metric.

3. The method of claim 2 , wherein:
the effort metric of a path is a sum of the node effort metrics for the path; and
the node effort metric is determined as a sum of steering and acceleration values to reach the node from a parent node of the path.

4. The method of claim 1 , wherein the paths are generated by introducing perturbations into the driving scenario.

5. A non-transitory computer-readable storage medium, the computer-readable storage medium including instructions that when executed by a computer, cause the computer to:
characterize a path in a driving scenario according to narrowness and effort of the path;
wherein the narrowness of the path is determined by a number of children at each node of the path;
assign to each node of the path a minimum amount of time for action by a vehicle to avoid collision from the node; and
apply the narrowness, effort, and minimum time for action to simulate the path in the computer.

6. The computer-readable storage medium of claim 5 , the computer-readable storage medium including instructions that when executed by a computer, cause the computer to:
score the path based at least in part on the effort of the path.

7. A method for computer simulation of a driving scenario for an autonomous vehicle, the method comprising:
configuring the driving scenario with a set of road descriptors on which a set of actors operate, each actor assigned a behavioral policy, and wherein at least one actor is the autonomous vehicle;
deriving a multivariate metric for a number of safe driving paths in the driving scenario;
assigning a narrowness metric to each of the paths in the driving scenario, wherein the narrowness metric is determined based on a number of child nodes of each node of the path;
assigning an effort metric to nodes of the paths; and
assigning to each node a minimum amount of time for action by the autonomous vehicle to avoid collision from the node;
executing the driving scenario on the paths in the computer simulation; and
improving a behavioral policy of the autonomous vehicle based on results of the computer simulation.

8. The method of claim 7 further comprising:
generating unsafe paths into the driving scenario by introducing perturbations into the driving scenario.

9. The method of claim 7 further comprising:
scoring the driving scenario based on performance of the autonomous vehicle in the computer simulation;
prioritizing the driving scenario based on the scoring; and
applying the prioritization to improve the behavioral policy of the autonomous vehicle.

10. The method of claim 7 wherein the effort for a path is computed as a sum of node efforts in the path.

11. The method of claim 7 wherein the multivariate metric comprises:
a total number of paths to evaluate from a beginning of the driving scenario to an end of the driving scenario, wherein the total number of paths forms a tree comprising branches;
a number of no-collision driving paths from the beginning of the driving scenario to the end of the driving scenario;
an average effort for the no-collision driving paths;
a minimum no-collision driving path effort;
a minimum amount of time before a collision occurs in a branch of the tree comprising at least one no-collision driving path; and
an average of a minimum number of branches of the tree along each no-collision driving path.

12. The method of claim 7 wherein:
a rate of change of a state of an actor is defined as a control policy represented by

f
=

ds
dt

=

[

a
⁢

υ
⁢

tan
⁢

(
δ
)

L

]

wherein a is an acceleration of the actor, δ is a steering angle of the actor, L is a wheelbase of the actor, and ν is a scalar velocity of the actor.

13. The method of claim 7 further comprising:
identifying a control policy for an unsafe actor A;
applying the control policy for the unsafe actor to generate a trajectory that minimizes an objective function γ represented by
γ AE =D AE 2 =∥x A −x E ∥ 2 2 =(x iA −x iE ) 2 +(x jA −x jE ) 2 wherein a position of actors A and E in the driving scenario is defined by a vector x=[x i x j ] of Cartesian coordinates and wherein a square of a Euclidean distance between the unsafe actor and an autonomous vehicle E is minimized.

14. The method of claim 13 , wherein:
reducing the objective function with time based on a constraint on a derivative of γAE such that

∂

γ

AE

∂
t

<
0.

15. The method of claim 14 , wherein the constraint is applied such that the autonomous vehicle does not react to the unsafe actor and therefore the control policy of the autonomous vehicle is fixed.

16. The method of claim 15 wherein the constraint is relaxed on condition that the autonomous vehicle applies Safety Force Field (SFF).

17. The method of claim 7 wherein the behavioral policy for an actor is represented as:
β 1 α A +β 2 tan(δ A )<const where a is acceleration of the actor, and δ is a steering angle for the actor.

18. The method of claim 17 wherein additional constraints are placed on a maximum value of the steering angle and the acceleration.

19. The method of claim 7 further comprising the application of a plurality of modes, wherein:
a first mode for selecting a control policy for an unsafe actor comprises a policy with maximum acceleration and maximum steering;
a second mode for selecting the control policy for the unsafe actor comprises a policy with maximum acceleration and minimum steering; and
a third mode for selecting the control policy for the unsafe actor comprises a p

Assignees

• Nvidia Corp

Inventors

• Hari Siva Kumar Sastry
• Frosio Iuri
• Ghodsi Zahra
• Anandkumar Anima
• Tsai Timothy
• Keckler Stephen W
• Troccoli Alejandro

Classifications

• G05D1/0214 Primary

  in accordance with safety or protection criteria, e.g. avoiding hazardous areas (monitoring the location of vehicles within a certain area, e.g. forbidden or allowed areas, in traffic control systems for road vehicles G08G1/13) · CPC title

• B60W60/0015

  specially adapted for safety · CPC title

• G01S13/931

  of land vehicles · CPC title

• G05B13/027

  using neural networks only · CPC title

• G05D1/0088 Primary

  characterized by the autonomous decision making process, e.g. artificial intelligence, predefined behaviours (using knowledge based models G06N5/00) · CPC title