Assessing surprise for autonomous vehicles

Patent metadata

Field	Value
Publication number	US-12017686-B1
Application number	US-202117399418-A
Country	US
Kind code	B1
Filing date	Aug 11, 2021
Priority date	Aug 11, 2021
Publication date	Jun 25, 2024
Grant date	Jun 25, 2024

Abstract

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for computing a backward looking surprise metric for autonomously driven vehicles. One of the methods includes obtaining first data representing one or more previously predicted states of an agent along one or more predicted trajectories of the agent at a first time step. Second data representing one or more states of the agent at a subsequent time step is obtained. A surprise score is computed from a measure of a difference between the first data computed for the one or more predicted trajectories for the prior time step and the second data computed for the one or more predicted states for the subsequent time step.

First claim

The invention claimed is:

1. A method comprising:
continually generating, at each time step of a plurality of time steps, data representing one or more predicted trajectories of an agent based on current state of the agent at a time step when the predicted trajectories are generated;
obtaining one or more values of one or more respective state variables;
obtaining first data representing one or more previously predicted states of the agent along the one or more predicted trajectories of the agent at a first time step, wherein the one or more previously predicted states were generated based on a prior state of the agent at a prior time step that occurred before a current time step;
obtaining, based on the one or more state variables, second data representing one or more states of the agent at a subsequent time step; and
computing a surprise score including computing a measure of a difference between the first data obtained for the one or more predicted trajectories for the prior time step and the second data obtained for the one or more states for the subsequent time step, wherein computing the surprise score comprises computing a likelihood of a posterior state based on a prior probability distribution, wherein the first data comprises the prior probability distribution and wherein the second data comprises the posterior state.

2. The method of claim 1 , wherein generating the one or more predicted trajectories is based on a window of one or more previous states before a time step of the current state.

3. The method of claim 1 , wherein computing the surprise score comprises computing a measure of a difference between a predicted state of the agent and an actual state of the agent, wherein the first data comprises the predicted state and wherein the second data comprises the actual state.

4. The method of claim 1 , wherein computing the surprise score comprises computing a measure of a difference between a posterior probability distribution and the prior probability distribution over states, wherein the second data comprises the posterior probability distribution over states.

5. The method of claim 1 , wherein computing the surprise score comprises computing a log likelihood of the posterior state based on the prior probability distribution, wherein the first data represents the prior probability distribution and wherein the second data represents the posterior state.

6. The method of claim 1 , wherein generating the data representing the one or more predicted trajectories comprises:
generating a plurality of predicted trajectories; and
generating a representative trajectory from the plurality of predicted trajectories.

7. The method of claim 6 , wherein generating the representative trajectory from the plurality of predicted trajectories comprises computing a weighted average trajectory using a likelihood associated with each of the plurality of predicted trajectories.

8. The method of claim 6 , wherein generating the representative trajectory from the plurality of predicted trajectories comprises selecting a trajectory having a maximum likelihood.

9. The method of claim 1 , wherein generating the data representing the one or more predicted trajectories comprises jointly representing a plurality of trajectories at each time step using a probability distribution.

10. The method of claim 9 , wherein each probability distribution computed for a plurality of predicted trajectories is a mixture model having component distributions for each of the predicted trajectories.

11. The method of claim 9 , wherein the first data comprises a first probability distribution, the second data comprises a second probability distribution, wherein computing the measure of the difference between the first data and the second data comprises computing a measure of divergence between the second probability distribution and the first probability distribution.

12. The method of claim 1 , wherein computing the second data comprises computing a predicted state of the agent in a future time step along a plurality of predicted trajectories based on the current state of the agent.

13. The method of claim 1 , wherein the method is performed by a computer-system on-board an autonomous or a semi-autonomous vehicle.

14. The method of claim 1 , wherein the state variables comprise separate lateral and longitudinal components,
and wherein computing the surprise score comprises computing separate lateral and longitudinal scores.

15. The method of claim 1 , further comprising performing a parameter sweeping process to generate multiple surprise scores using only a single inference pass of a generative model, wherein a first parameter of the parameter sweeping process is which time step is used as the prior time step, and wherein a second parameter of the parameter sweeping process is a number of steps to predict into a future time step.

16. The method of claim 1 , wherein the method is performed by a computer-system in an offline status in a location different from an autonomous or a semi-autonomous vehicle.

17. A system comprising:
one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
continually generating, at each time step of a plurality of time steps, data representing one or more predicted trajectories of an agent based on current state of the agent at a time step when the predicted trajectories are generated;
obtaining one or more values of one or more respective state variables;
obtaining first data representing one or more previously predicted states of the agent along the one or more predicted trajectories of the agent at a first time step, wherein the one or more previously predicted states were generated based on a prior state of the agent at a prior time step that occurred before a current time step;
obtaining, based on the one or more state variables, second data representing one or more states of the agent at a subsequent time step; and
computing a surprise score including computing a measure of a difference between the first data obtained for the one or more predicted trajectories for the prior time step and the second data obtained for the one or more states for the subsequent time step, wherein computing the surprise score comprises computing a likelihood of a posterior state based on a prior probability distribution, wherein the first data comprises the prior probability distribution and wherein the second data comprises the posterior state.

18. The system of claim 17 , wherein generating the one or more predicted trajectories is based on a window of one or more previous states before a time step of the current state.

19. One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
continually generating, at each time step of a plurality of time steps, data representing one or more predicted trajectories of an agent based on current state of the agent at a time step when the predicted trajectories are generated;
obtaining one or more values of one or more respective state variables;
obtaining first data representing one or more previously predicted states of the agent along the one or more predicted trajectories of the agent at a first time step, wherein the one or more previously predicted states were generated based on a prior state of the agent at a prior ti

Assignees

• Waymo Llc

Inventors

• Dinparastdjadid Azadeh
• Engstrom Johan
• Chen Haoyu
• Supeene Isaac
• Wang Menghui

Classifications

• B60W60/0027 Primary

  using trajectory prediction for other traffic participants · CPC title

• G07C5/0808 Primary

  Diagnosing performance data (testing of vehicles G01M17/00; testing of electrical installation on vehicles G01R31/005) · CPC title

• B60W2554/404

  Characteristics · CPC title

• B60W2554/4026

  Cycles · CPC title

• B60W2554/4029

  Pedestrians · CPC title