Autonomous vehicles featuring machine-learned yield model

1. A computing system that controls an autonomous vehicle, the computing system comprising:

a machine-learned yield model configured to receive and process feature data descriptive of objects perceived by the autonomous vehicle and, in response, provide yield decisions for the autonomous vehicle relative to the objects;

one or more processors; and

one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the computing system to perform operations, the operations comprising:

obtaining first feature data descriptive of a first object perceived by the autonomous vehicle, wherein obtaining the first feature data descriptive of the first object perceived by the autonomous vehicle comprises determining gap data descriptive of a plurality of gaps associated with one or more additional vehicles perceived by the autonomous vehicle, wherein the plurality of gaps comprise one or more temporal gaps that are predicted to occur;

inputting the first feature data into the machine-learned yield model;

receiving a first yield decision as an output of the machine-learned yield model, wherein the first yield decision describes a yield behavior for the autonomous vehicle relative to the first object, wherein the first yield decision instructs the autonomous vehicle to proceed through a particular gap of the plurality of gaps, wherein receiving the first yield decision as the output of the machine-learned yield model comprises

obtaining, through implementation of the machine-learned yield model, a plurality of scores respectively for the plurality of gaps; and

selecting the particular gap of the plurality of gaps based at least in part on the plurality of scores;

generating a motion plan that executes the yield behavior described by the first yield decision output by the machine-learned yield model, wherein generating the motion plan comprises generating a gap penalty for the particular gap of the plurality of gaps, wherein the gap penalty is used to guide motion of the autonomous vehicle through the particular gap; and

controlling motion of the autonomous vehicle based at least in part on the motion plan.

2. The computing system of claim 1 , wherein:

the first object comprises a traffic signal; and

the first feature data descriptive of the first object comprises feature data descriptive of a transition timing associated with the traffic signal.

3. The computing system of claim 1 , wherein the first feature data descriptive of the first object comprises at least one of a required deceleration and a required acceleration associated with the first object.

4. The computing system of claim 1 , wherein:

the first object comprises a traffic signal; and

the first feature data descriptive of the first object comprises feature data descriptive of at least one of:

a stop line location associated with the traffic signal;

a cross traffic line location associated with the traffic signal; and

an exit point location associated with the traffic signal.

5. The computing system of claim 1 , wherein determining gap data descriptive of the plurality of gaps associated with the one or more additional vehicles comprises:

obtaining one or more predicted trajectories respectively generated for the one or more additional vehicles by the autonomous vehicle;

identifying one or more yielding lanes through which the autonomous vehicle will traverse according to a proposed path; and

determining the plurality of gaps for the one or more yielding lanes based at least in part on the one or more predicted trajectories respectively generated for the one or more additional vehicles.

6. The computing system of claim 1 , wherein the operations further comprise:

generating a plurality of gap labels on which the machine-learned yield model is trained, wherein generating the plurality of gap labels comprises:

determining a plurality of example gaps;

determining a plurality of trajectories that respectively traverse through the plurality of example gaps;

determining a plurality of scores respectively for the plurality of trajectories; and

labelling as a positive training example each example gap for which the corresponding trajectory has a score that is greater than a stop score associated with a stop plan and does not violate one or more constraints; and

labelling as a negative training example each example gap for which the corresponding trajectory has a score that is less than the stop score associated with the stop plan or violates at least one of the one or more constraints.

7. The computing system of claim 1 , wherein the machine-learned yield model comprises a machine-learned yield classifier model.

8. The computing system of claim 1 , wherein the machine-learned yield model comprises a machine-learned random forest model or a machine-learned neural network.

9. The computing system of claim 1 , wherein the operations further comprise:

training the machine-learned yield model based at least in part on log data annotated with yield labels, wherein the log data describes yield behaviors performed by autonomous vehicles during previously conducted real-world driving sessions.

10. The computing system of claim 1 , wherein the operations further comprise:

generating a plurality of synthetic autonomous vehicle trajectories;

simulating an outcome for each synthetic autonomous vehicle trajectory;

assigning a yield label for each synthetic autonomous vehicle trajectory based at least in part on the outcome for such synthetic autonomous vehicle trajectory; and

training the machine-learned yield model based at least in part on the yield labels assigned to the synthetic autonomous vehicle trajectories.

11. An autonomous vehicle, comprising:

a machine-learned yield model configured to receive and process feature data descriptive of objects perceived by the autonomous vehicle and, in response, provide yield decisions for the autonomous vehicle relative to the objects;

one or more processors; and

one or more non-transitory computer-readable media that store instructions that, when executed by the one or more processors, cause the one or more processors to perform operations, the operations comprising:

obtaining first feature data descriptive of a first object perceived by the autonomous vehicle, wherein obtaining the first feature data descriptive of the first object perceived by the autonomous vehicle comprises determining gap data descriptive of a plurality of gaps associated with one or more additional vehicles perceived by the autonomous vehicle, wherein the plurality of gaps comprise one or more temporal gaps that are predicted to occur;

inputting the first feature data into the machine-learned yield model;

receiving a first yield decision as an output of the machine-learned yield model, wherein the first yield decision describes a yield behavior for the autonomous vehicle relative to the first object, wherein the first yield decision instructs the autonomous vehicle to proceed through a particular gap of the plurality of gaps, wherein receiving the first yield decision as the output of the machine-learned yield model comprises

obtaining, through implementation of the machine-learned yield model, a plurality of scores respectively for the plurality of gaps; and

selecting the particular gap of the plurality of gaps based at least in part on the plurality of scores;

generating a motion plan that executes the yield behavior described by the first yield decision output by the machine-learned yield model, wherein generating the motion plan comprises generating a gap penalty for the particular gap of the plurality of gaps, wherein the gap penalty is used to guide motion of the autonomous vehicle through the particular gap; and

controlling motion of the autonomous vehicle based at least in part on the motion plan.

12. The autonomous vehicle of claim 11 , wherein:

the first object comprises a traffic signal; and

the first feature data descriptive of the first object comprises feature data descriptive of a transition timing associated with the traffic signal.

13. The autonomous vehicle of claim 11 , wherein the first feature data descriptive of the first object comprises at least one of a required deceleration and a required acceleration associated with the first object.

14. A computer-implemented method, comprising:

obtaining, by a computing system that comprises one or more computing devices, first feature data descriptive of a first object perceived by an autonomous vehicle, wherein obtaining the first feature data descriptive of the first object perceived by the autonomous vehicle comprises determining, by the computing system, gap data descriptive of a plurality of gaps associated with one or more additional vehicles perceived by the autonomous vehicle, wherein the plurality of gaps comprise one or more temporal gaps that are predicted to occur;

inputting, by the computing system, the first feature data into a machine-learned yield model;

receiving, by the computing system, a first yield decision as an output of the machine-learned yield model, wherein the first yield decision describes a yield behavior for the autonomous vehicle relative to the first object, wherein the first yield decision instructs the autonomous vehicle to proceed through a particular gap of the plurality of gaps, wherein receiving the first yield decision as the output of the machine-learned yield model comprises

obtaining, through implementation of the machine-learned yield model, a plurality of scores respectively for the plurality of gaps; and

selecting the particular gap of the plurality of gaps based at least in part on the plurality of scores;

generating a motion plan that executes the yield behavior described by the first yield decision output by the machine-learned yield model, wherein generating the motion plan comprises generating a gap penalty for the particular gap of the plurality of gaps, wherein the gap penalty is used to guide motion of the autonomous vehicle through the particular gap; and

controlling motion of the autonomous vehicle based at least in part motion plan.