Machine-learning systems and techniques to optimize teleoperation and/or planner decisions

Patent metadata

Field	Value
Publication number	US-9632502-B1
Application number	US-201514933602-A
Country	US
Kind code	B1
Filing date	Nov 5, 2015
Priority date	Nov 4, 2015
Publication date	Apr 25, 2017
Grant date	Apr 25, 2017

Abstract

A system, an apparatus or a process may be configured to implement an application that applies artificial intelligence and/or machine-learning techniques to predict an optimal course of action (or a subset of courses of action) for an autonomous vehicle system (e.g., one or more of a planner of an autonomous vehicle, a simulator, or a teleoperator) to undertake based on suboptimal autonomous vehicle performance and/or changes in detected sensor data (e.g., new buildings, landmarks, potholes, etc.). The application may determine a subset of trajectories based on a number of decisions and interactions when resolving an anomaly due to an event or condition. The application may use aggregated sensor data from multiple autonomous vehicles to assist in identifying events or conditions that might affect travel (e.g., using semantic scene classification). An optimal subset of trajectories may be formed based on recommendations responsive to semantic changes (e.g., road construction).

First claim

What is claimed is:

1. A method, comprising:
receiving, from one or more of a sensor or a communication interface, telemetry data associated with an event, the event being a situation or condition associated with operation of an autonomous vehicle;
obtaining policy data, the policy data including instructions for operating an autonomous vehicle according to a trajectory and the policy data including a confidence level associated with the received trajectory, the confidence level indicating a degree of certainty that the autonomous vehicle, in operating according to the trajectory, will operate safely in view of the event;
obtaining candidate trajectories responsive to the event, based on the telemetry data, each candidate trajectory having an associated confidence level;
generating, by machine-learning by a processor and based at least in part on the candidate trajectories and the telemetry data, updated policy data that includes instructions for operating the autonomous vehicle responsive to the event differently than according to the policy data; and
communicating the updated policy data to at least one autonomous vehicle via a communications interface.

2. The method of claim 1 further comprising:
determining that the confidence level associated with the received trajectory is below a confidence level threshold, indicating a non-normative operational state of the autonomous vehicle.

3. The method of claim 2 , wherein at least one of the associated confidence levels is greater than the confidence level associated with the received trajectory.

4. The method of claim 2 , wherein the candidate trajectories are obtained at least in part responsive to the confidence level associated with the received trajectory indicating the non-normative operational state.

5. The method of claim 1 , wherein communicating the updated policy data to at least one autonomous vehicle comprises communicating the updated policy data to an autonomous vehicle other than the autonomous vehicle with which the event is associated.

6. The method of claim 5 , wherein the telemetry data and policy data are obtained from the autonomous vehicle with which the event is associated.

7. The method of claim 1 , wherein the telemetry data is received and the policy data is obtained at a teleoperator computing device and wherein the teleoperator computing device is configured to communicate the updated policy data to at least one autonomous vehicle.

8. The method of claim 1 , wherein the telemetry data is received and the policy data is obtained at a simulator, the method further comprising, by the simulator:
obtaining the candidate trajectories responsive to the event by simulating the candidate trajectories based at least in part on one or more of the telemetry data or the policy data;
determining confidence levels associated with the simulated candidate trajectories;
receiving selection of one of the candidate trajectories for implementation by the autonomous vehicle; and
providing the selected trajectory to the at least one autonomous vehicle as part of the updated policy data.

9. The method of claim 8 the method further comprising, by the simulator:
conducting the generation of the updated policy data based at least in part on one or more of simulating the candidate trajectories or the selected trajectory; and
communicating one or more of the policy data or the updated policy data to one or more of:
a teleoperator computing device, or
the at least one autonomous vehicle.

10. The method of claim 1 , wherein the telemetry data is received and the policy data is obtained at a simulator configured to simulate an autonomous vehicle in a synthetic environment.

11. The method of claim 1 , wherein the telemetry data is received, the policy data is obtained, and the updated policy data is generated at an autonomous vehicle.

12. The method of claim 1 , wherein the policy data is obtained from one or more of an autonomous vehicle that captured the telemetry data or a data repository associated with a service for a fleet of autonomous vehicles.

13. The method of claim 1 , further comprising:
determining that the at least one autonomous vehicle has operated according to the teleoperation trajectory;
obtaining further telemetry data from the at least one autonomous vehicle;
determining from at least the further telemetry data that operation of the at least one autonomous vehicle according to the teleoperation trajectory and responsive to the event resulted in normative operation of the at least one autonomous vehicle;
generating, by machine-learning, the updated policy responsive to the determining that operating the at least one autonomous vehicle according to the teleoperation trajectory resulted in normative operation; and
communicating the updated policy data to at least one autonomous vehicle.

14. The method of claim 1 , wherein the policy data includes one or more of:
criteria for selecting a candidate trajectory for operating the autonomous vehicle;
rules defining constraints for generating trajectories;
threshold confidence levels for determining whether operations of the autonomous vehicle are normative;
a classifier for detecting, from the telemetry data, objects in an environment in which the autonomous vehicle is operating;
a semantic classification of an object detected from the telemetry data; or
a semantic classification associated with the event; and
wherein the updated policy data includes an update to one or more of the criteria, the rules, the threshold confidence levels, the semantic classification of the object, or the semantic classification associated with the event.

15. A system comprising:
one or more processors;
a memory having stored thereon first criteria for controlling motion of an autonomous vehicle and a planner module executable by the one or more processors and that, when executed by the one or more processors, configure the system to perform operations including:
obtaining sensor data related to operation of an autonomous vehicle and an event;
determining, based at least on the sensor data and the first criteria, a first trajectory for operation of the autonomous vehicle responsive to the event;
receiving instructions for implementing a second trajectory for operation of the autonomous vehicle responsive to the event; and
learning, from at least the instructions and the sensor data, second criteria to update the first criteria, the second criteria configuring the one or more processors to determine the second trajectory based at least on the sensor data and the instructions.

16. A system as claim 15 recites, the first criteria including one or more of:
criteria for selecting a candidate trajectory for operating the autonomous vehicle;
rules defining constraints for generating trajectories;
threshold confidence levels for determining whether operations of the autonomous vehicle are normative;
a classifier for detecting, from the sensor data, objects in an environment in which the autonomous vehicle is operating;
a semantic classification of an object detected by the perception engine from the sensor data;
a semantic classification associated with the event;
one or more candidate trajectories; or
confidence levels associated with the candidate trajectories or rules for determining confidence levels associated with the candidate trajectories;
and wherein the second criteria includes a modification of or addition to one or more of:
the criteria for selecting a candidate trajectory,
the rules defining constraints for generating trajectories,
the thres

Assignees

• Zoox Inc

Inventors

• Levinson Jesse Sol
• Sibley Gabriel Thurston
• Rege Ashutosh Gajanan

Classifications

• G06N7/01

  Probabilistic graphical models, e.g. probabilistic networks · CPC title

• B60Q1/28

  for indicating front of vehicle · CPC title

• G06Q10/06

  Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling · CPC title

• B60Q2400/40

  Welcome lights, i.e. specific or existing exterior lamps to assist leaving or approaching the vehicle · CPC title

• G08G1/166

  for active traffic, e.g. moving vehicles, pedestrians, bikes · CPC title