Probabilistic simulation sampling from agent data

Patent metadata

Field	Value
Publication number	US-11834070-B2
Application number	US-202117370924-A
Country	US
Kind code	B2
Filing date	Jul 8, 2021
Priority date	Jul 8, 2021
Publication date	Dec 5, 2023
Grant date	Dec 5, 2023

Abstract

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for determining the likelihood that a particular event would occur during a navigation interaction using simulations generated by sampling from agent data. In one aspect, a method comprises: identifying an instance of a navigation interaction that includes an autonomous vehicle and agents navigating in an environment; generating multiple simulated interactions corresponding to the instance, comprising, for each simulated interaction: identifying one or more agents; for each identified agent and for each property that characterizes behavior of the identified agent, obtaining a probability distribution for the property; sampling a respective value from each of the probability distributions; and simulating the navigation interaction in accordance with the sampled values; and determining a likelihood that the particular event would occur during the navigation interaction based on whether the particular event occurred during each of the simulated interactions.

First claim

What is claimed is:

1. A method performed by one or more computers, the method comprising:
identifying an instance of a navigation interaction that includes an autonomous vehicle and one or more agents navigating in an environment;
generating a plurality of simulated interactions corresponding to the instance, comprising, for each simulated interaction:
identifying one or more of the agents;
for each identified agent and for each of one or more properties that characterize behavior of the identified agent, obtaining data representing a probability distribution over a set of possible values for the property, wherein the one or more properties comprise a reaction time of the identified agent to a stimulus;
sampling a respective value from each of the probability distributions; and
simulating the navigation interaction using a computer simulation of the environment that, for each identified agent, selects actions performed by the identified agent at a plurality of time steps using behavior predictions for other agents in the computer simulation, and wherein the simulating comprises:
selecting, based on the sampled value for the reaction time of the identified agent, a time window after a time step at which the stimulus occurred; and
selecting actions performed by the identified agent for the time steps in the time window using prior behavior predictions from time steps prior to the stimulus occurring;

determining, for each simulated interaction, whether a particular event occurred during the simulated interaction; and
determining a likelihood that the particular event would occur during the navigation interaction based at least in part on whether the particular event occurred during each of the simulated interactions.

2. The method of claim 1 , wherein simulating the navigation interaction comprises simulating the navigation interaction such that any other agent that is not identified behaves as the other agent did in the instance of the navigation interaction.

3. The method of claim 1 , wherein, during the simulation, selecting actions performed by the identified agent for the time steps in the time window is in response to (i) state data characterizing a respective state of the environment at the time step and (ii) behavior predictions for other agents in the simulation.

4. The method of claim 3 , wherein selecting actions performed by the identified agent for the time steps in the time window using prior behavior predictions from time steps prior to the stimulus occurring comprises:
setting the behavior predictions for the time steps in the time window to the behavior predictions obtained at a most recent time step prior to the time step at which the stimulus occurred; and
obtaining new behavior predictions starting from the first time step after the time window has elapsed.

5. The method of claim 1 , wherein for each identified agent and for each of one or more properties that characterize behavior of the identified agent, obtaining data representing a probability distribution over a set of possible values for the property comprises:
the probability distribution being generated from logged data characterizing agents navigating in the environment.

6. The method of claim 5 , wherein for each identified agent and for each of the one or more properties, the probability distribution over the set of possible values for the property represents an estimate of a range of possible responses of agents controlled by humans to a stimulus.

7. The method of claim 1 , wherein the particular event is a collision between two agents in the environment.

8. The method of claim 7 , wherein the particular event is one of a low severity collision or a high severity collision.

9. The method of claim 1 , further comprising:
identifying, in the instance, a time point at which the stimulus occurs by identifying an interaction between two agents that satisfies one or more criteria.

10. A system comprising:
one or more computers; and
one or more storage devices storing instructions that, when executed by the one or more computers, cause the one or more computers to perform operations for determining the likelihood that a particular event would occur during a navigation interaction using one or more simulations of the navigation interaction, each simulation generated by sampling from agent data, the operations comprising:
identifying an instance of a navigation interaction that includes an autonomous vehicle and one or more agents navigating in an environment;
generating a plurality of simulated interactions corresponding to the instance, comprising, for each simulated interaction:
identifying one or more of the agents;
for each identified agent and for each of one or more properties that characterize behavior of the identified agent, obtaining data representing a probability distribution over a set of possible values for the property, wherein the one or more properties comprise a reaction time of the identified agent to a stimulus;
sampling a respective value from each of the probability distributions; and
simulating the navigation interaction using a computer simulation of the environment that, for each identified agent, selects actions performed by the identified agent at a plurality of time steps using behavior predictions for other agents in the computer simulation, and wherein the simulating comprises:
selecting, based on the sampled value for the reaction time of the identified agent, a time window after a time step at which the stimulus occurred; and
selecting actions performed by the identified agent for the time steps in the time window using prior behavior predictions from time steps prior to the stimulus occurring;

determining, for each simulated interaction, whether a particular event occurred during the simulated interaction; and
determining a likelihood that the particular event would occur during the navigation interaction based at least in part on whether the particular event occurred during each of the simulated interactions.

11. The system of claim 10 , wherein simulating the navigation interaction comprises simulating the navigation interaction such that any other agent that is not identified behaves as the other agent did in the instance of the navigation interaction.

12. The method of claim 10 , wherein, during the simulation, selecting actions performed by the identified agent for the time steps in the time window is for each identified agent, a respective action in response to (i) state data characterizing a respective state of the environment at the time step and (ii) behavior predictions for other agents in the simulation.

13. The method of claim 10 , wherein selecting actions performed by the identified agent for the time steps in the time window using prior behavior predictions from time steps prior to the stimulus occurring comprises:
setting the behavior predictions for the time steps in the time window to the behavior predictions obtained at a most recent time step prior to the time step at which the stimulus occurred; and
obtaining new behavior predictions starting from the first time step after the time window has elapsed.

14. The system of claim 10 , wherein for each identified agent and for each of one or more properties that characterize behavior of the identified agent, obtaining data representing a probability distribution over a set of possible values for the property comprises:
the probability distribution being generated from logged data characterizing agents navigating in the environment.

15. The system of claim 14 , wherein for each identified agent and for each of the

Assignees

• Waymo Llc

Inventors

• Engstrom Johan
• Christophe Emmanuel
• Lee Joseph
• Supeene Isaac
• Mathias Razvan

Classifications

• B60W60/0011 Primary

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

• B60W30/0956

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

• G06V40/20

  Movements or behaviour, e.g. gesture recognition (recognition of facial expressions G06V40/16) · CPC title

• B60W2554/4046

  Behavior, e.g. aggressive or erratic · CPC title

• B60W2554/4049

  Relationship among other objects, e.g. converging dynamic objects · CPC title