Neural network based modeling and simulation of non-stationary traffic objects for testing and development of autonomous vehicle systems
A system performs modeling and simulation of non-stationary traffic entities for testing and development of modules used in an autonomous vehicle system. The system uses a machine learning based model that predicts hidden context attributes for traffic entities that may be encountered by a vehicle in traffic. The system generates simulation data for testing and development of modules that help navigate autonomous vehicles. The generated simulation data may be image or video data including representations of traffic entities, for example, pedestrians, bicyclists, and other vehicles. The system may generate simulation data using generative adversarial neural networks.
1. A method for generating simulation data for testing and development of systems for navigating autonomous vehicles, the method comprising:
generating a feature vector, the feature vector representing one or more characteristics of a traffic entity representing a non-stationary object in traffic in which a vehicle is driving;
providing the feature vector to a rendering engine to generate an image including the traffic entity, wherein characteristics of the traffic entity correspond to the feature vector;
accessing a trained neural network configured to receive an input image displaying the traffic entity as input and to generate a value of a hidden context attribute of the traffic entity displayed in the input image;
receiving a target value of the hidden context attribute for the traffic entity;
generating a target image that when provided as input to the trained neural network, outputs a value of the hidden context attribute that is within a threshold of the target value of the hidden context attribute, the generating comprising, iteratively performing steps of:
perturbing the feature vector;
providing the perturbed feature vector to the rendering engine to generate an updated image;
applying the trained neural network to the updated image to generate a predicted value of the hidden context attribute; and
comparing the predicted value of the hidden context attribute with the target value of the hidden context attribute;
responsive to the predicted value of the hidden context attribute being within the threshold of the target value of the hidden context attribute, using the updated image as the target image; and
using the target image as simulation data for testing of a motion planner, the motion planner configured to adjust motion of an autonomous vehicle according to a value of the hidden context attribute of one or more traffic entities encountered by the autonomous vehicle while driving in traffic.
2. The method of claim 1 , wherein the characteristics include one or more of a type of the traffic entity, an orientation of the traffic entity, a position of the traffic entity, and a motion of the traffic entity.
3. The method of claim 1 , wherein the feature vector comprises the target value of the hidden context attribute.
4. The method of claim 1 , wherein the perturbed feature vector is determined using gradient descent.
5. The method of claim 1 , wherein the hidden context attribute represents a state of mind of a user represented by the traffic entity.
6. The method of claim 1 , wherein the hidden context attribute represents a task that a user represented by the traffic entity is planning on accomplishing.
7. The method of claim 1 , wherein the hidden context attribute represents a degree of awareness of the autonomous vehicle by a user represented by the traffic entity.
8. The method of claim 1 , wherein the hidden context attribute represents a goal of a user represented by the traffic entity, wherein the user expects to achieve the goal within a threshold time interval.
9. The method of claim 1 , wherein adjusting motion of the autonomous vehicle comprises:
generating signals for controlling the autonomous vehicle based on one or more motion parameters and the hidden context attribute of each of the one or more traffic entities; and
sending the generated signals to controls of the autonomous vehicle.
10. A non-transitory computer readable storage medium storing instructions for generating simulation data for testing and development of systems for navigating autonomous vehicles, wherein the instructions when executed by a computer processor, cause the computer processor to perform steps comprising:
generating a feature vector, the feature vector representing one or more characteristics of a traffic entity representing a non-stationary object in traffic in which a vehicle is driving;
providing the feature vector to a rendering engine to generate an image including the traffic entity, wherein characteristics of the traffic entity correspond to the feature vector;
accessing a trained neural network configured to receive an input image displaying the traffic entity and to generate a value of a hidden context attribute of the traffic entity displayed in the input image;
receiving a target value of the hidden context attribute for the traffic entity;
generating a target image that when provided as input to the trained neural network, outputs a value of the hidden context attribute that is within a threshold of the target value of the hidden context attribute, the generating comprising, iteratively performing steps of:
perturbing the feature vector;
providing the perturbed feature vector to the rendering engine to generate an updated image;
applying the trained neural network to the updated image to generate an updated value of the hidden context attribute; and
comparing the updated value of the hidden context attribute with the target value of the hidden context attribute;
responsive to the updated value of the hidden context attribute being within the threshold of the target value of the hidden context attribute, using the updated image as the target image; and
using the updated image as simulation data for testing of a motion planner, the motion planner configured to adjust motion of an autonomous vehicle according to a value of the hidden context attribute of one or more traffic entities encountered by the autonomous vehicle while driving in traffic.
11. The non-transitory computer readable storage medium of claim 10 , wherein the characteristics of the traffic entity include one or more of a type of the traffic entity, an orientation of the traffic entity, a position of the traffic entity, and a motion of the traffic entity.
12. The non-transitory computer readable storage medium of claim 10 , wherein the feature vector comprises the target value of the hidden context attribute.
13. The non-transitory computer readable storage medium of claim 10 , wherein the perturbed feature vector is determined using gradient descent.
14. The non-transitory computer readable storage medium of claim 10 , wherein the hidden context attribute represents a state of mind of a user represented by the traffic entity.
15. The non-transitory computer readable storage medium of claim 10 , wherein the hidden context attribute represents a task that a user represented by the traffic entity is planning on accomplishing.
16. The non-transitory computer readable storage medium of claim 10 , wherein the hidden context attribute represents a degree of awareness of the autonomous vehicle by a user represented by the traffic entity.
17. The non-transitory computer readable storage medium of claim 10 , wherein the hidden context attribute represents a goal of a user represented by the traffic entity, wherein the user expects to achieve the goal within a threshold time interval.
18. The non-transitory computer readable storage medium of claim 10 , wherein instructions for adjusting motion of the autonomous vehicle cause the computer processor to perform steps comprising:
generating signals for controlling the autonomous vehicle based on one or more motion parameters and the hidden context attribute of each of the one or more traffic entities; and
sending the generated signals to controls of the autonomous vehicle.
19. A computer system for generating simulation data for testing and development of systems for navigating autonomous vehicles, comprising:
a computer processor; and
a non-transitory computer readable storage medium storing instructions that when executed by the computer processor, cause the computer processor to perform steps comprising:
generating a feature vector, the feature vector representing one or more characteristics of a traffic entity representing a non-stationary object in traffic in which a vehicle is driving;
providing the feature vector to a rendering engine to generate an image including the traffic entity, wherein characteristics of the traffic entity correspond to the feature vector;
accessing a trained neural network configured to receive an input image displaying the traffic entity and to generate a value of a hidden context attribute of the traffic entity displayed in the input image;
receiving a target value of the hidden context attribute for the traffic entity;
generating a target image that when provided as input to the trained neural network, outputs a value of the hidden context attribute that is within a threshold of the target value of the hidden context attribute, the generating comprising, iteratively performing steps of:
perturbing the feature vector;
providing the perturbed feature vector to the rendering engine to generate updated image;
applying the trained neural network to the updated image to generate an updated value of the hidden context attribute; and
comparing the updated value of the hidden context attribute with the target value of the hidden context attribute;
responsive to the updated value of the hidden context attribute being within the threshold of the target value of the hidden context attribute, using the updated image as the target image; and
using the updated image as simulation data for testing of a motion planner, the motion planner configured to adjust motion of an autonomous vehicle according to a value of the hidden context attribute of one or more traffic entities encountered by the autonomous vehicle while driving in traffic.
20. The computer system of claim 19 , wherein the one or more characteristics of the traffic entity include one or more of a type of the traffic entity, an orientation of the traffic entity, a position of the traffic entity, and s motion of the traffic entity.