Right-of-way-based semantic coverage and automatic labeling for trajectory generation in autonomous systems

1 . A computer-implemented method comprising:

receiving a set of trajectories generated by at least one trajectory generation source, the set of trajectories generated for a scene having an autonomous vehicle (AV) interaction with a road agent;

labeling a ground truth trajectory from the set of trajectories with a ground truth label; classifying each remaining trajectory of the set of trajectories as at least one of an assert trajectory or a yield trajectory;

for an assert group comprising the assert trajectories, assigning an auxiliary assert label to an assert trajectory having a highest selection score from a trajectory selection source;

for a yield group comprising the yield trajectories, assigning an auxiliary yield label to a yield trajectory having a highest selection score from the trajectory selection source; and

utilizing the ground truth label, the auxiliary assert label, and the auxiliary yield label to train a trajectory generation model, wherein training of the trajectory generation model comprises using mode selectors corresponding to each label to provide secondary supervision trajectories when computing a loss for the trajectory generation model.

2 . The computer-implemented method of claim 1 , wherein the AV interaction comprises at least one of an overtake interaction or a commit interaction.

3 . The computer-implemented method of claim 1 , wherein the at least one trajectory generation source comprises the trajectory generation model, a heuristics generation source, or a non-convex solver and costing source.

4 . The computer-implemented method of claim 1 , wherein classifying each remaining trajectory of the set of trajectories as at least one of the assert trajectory or the yield trajectory further comprises comparing each trajectory to set of assert metrics or yield metrics comprising at least one of a speed difference between an AV and the road agent or determination of whether the AV or the road agent arrives to a commit region first.

5 . The computer-implemented method of claim 1 , wherein the trajectory selection source comprises a trajectory selection model.

6 . The computer-implemented method of claim 1 , wherein the mode selectors select a candidate trajectory from a set of candidate training trajectories that is closest to a label corresponding to a mode selector.

7 . The computer-implemented method of claim 6 , wherein an individual weighted Huber loss is computed for each selected candidate trajectory, and wherein individual weighted Huber losses are combined to generate a final combined loss for training of the trajectory generation model.

8 . The computer-implemented method of claim 1 , further comprising:

receiving a collection of generated trajectories from a machine learning model;

classifying a trajectory of the collection of generated trajectories as an assert trajectory classification responsive to the trajectory satisfying a set of assert definition metrics;

classifying a trajectory of the collection of generated trajectories as a yield trajectory classification responsive to the trajectory satisfying a set of yield definition metrics; and

determining a right-of-way-based semantic mode coverage metric based on assert trajectory classifications and yield trajectory classifications as compared to a total number of the generated trajectories in the collection.

9 . The computer-implemented method of claim 8 , wherein the right-of-way-based semantic mode coverage metric comprises at least one of a percentage of the assert trajectory classifications and the yield trajectory classifications, respectively, within the collection of generated trajectories or identifying whether there is at least one of each assert trajectory classification and yield trajectory classification in the collection of generated trajectories per each AV interaction.

10 . An apparatus comprising:

one or more hardware processors to:

receive a set of trajectories generated by at least one trajectory generation source, the set of trajectories generated for a scene having an autonomous vehicle (AV) interaction with a road agent;

label a ground truth trajectory from the set of trajectories with a ground truth label;

classify each remaining trajectory of the set of trajectories as at least one of an assert trajectory or a yield trajectory;

for an assert group comprising the assert trajectories, assign an auxiliary assert label to an assert trajectory having a highest selection score from a trajectory selection source;

for a yield group comprising the yield trajectories, assign an auxiliary yield label to a yield trajectory having a highest selection score from the trajectory selection source; and

utilize the ground truth label, the auxiliary assert label, and the auxiliary yield label to train a trajectory generation model, wherein training of the trajectory generation model comprises using mode selectors corresponding to each label to provide secondary supervision trajectories when computing a loss for the trajectory generation model.

11 . The apparatus of claim 10 , wherein the AV interaction comprises at least one of an overtake interaction or a commit interaction.

12 . The apparatus of claim 10 , wherein the one or more processors to classify each remaining trajectory of the set of trajectories as at least one of the assert trajectory or the yield trajectory further comprises the one or more processors to compare each remaining trajectory to set of assert metrics or yield metrics comprising at least one of a speed difference between an AV and the road agent or determination of whether the AV or the road agent arrives to a commit region first.

13 . The apparatus of claim 10 , wherein the mode selectors select a candidate trajectory from a set of candidate training trajectories that is closest to a label corresponding to a mode selector.

14 . The apparatus of claim 13 , wherein an individual weighted Huber loss is computed for each selected candidate trajectory, and wherein individual weighted Huber losses are combined to generate a final combined loss for training of the trajectory generation model.

15 . The apparatus of claim 10 , wherein the one or more processors are further to:

receive a collection of generated trajectories from a machine learning model;

classify a trajectory of the collection of generated trajectories as an assert trajectory classification responsive to the trajectory satisfying a set of assert definition metrics;

classify a trajectory of the collection of generated trajectories as a yield trajectory classification responsive to the trajectory satisfying a set of yield definition metrics; and

determine a right-of-way-based semantic mode coverage metric based on assert trajectory classifications and yield trajectory classifications as compared to a total number of the generated trajectories in the collection.

16 . A non-transitory computer-readable medium having stored thereon instructions that, when executed by one or more processors, cause the one or more processors to:

receive a set of trajectories generated by at least one trajectory generation source, the set of trajectories generated for a scene having an autonomous vehicle (AV) interaction with a road agent;

label a ground truth trajectory from the set of trajectories with a ground truth label;

classify each remaining trajectory of the set of trajectories as at least one of an assert trajectory or a yield trajectory;

for an assert group comprising the assert trajectories, assign an auxiliary assert label to an assert trajectory having a highest selection score from a trajectory selection source;

for a yield group comprising the yield trajectories, assign an auxiliary yield label to a yield trajectory having a highest selection score from the trajectory selection source; and

utilize the ground truth label, the auxiliary assert label, and the auxiliary yield label to train a trajectory generation model, wherein training of the trajectory generation model comprises using mode selectors corresponding to each label to provide secondary supervision trajectories when computing a loss for the trajectory generation model.

17 . The non-transitory computer-readable medium of claim 16 , wherein the AV interaction comprises at least one of an overtake interaction or a commit interaction.

18 . The non-transitory computer-readable medium of claim 16 , wherein the one or more processors to classify each remaining trajectory of the set of trajectories as at least one of the assert trajectory or the yield trajectory further comprises the one or more processors to compare each trajectory to set of assert/yield metrics or yield metrics comprising at least one of a speed difference between an AV and the road agent or determination of whether the AV or the road agent arrives to a commit region first.

19 . The non-transitory computer-readable medium of claim 16 , wherein the mode selectors select a candidate trajectory from a set of candidate training trajectories that is closest to a label corresponding to a mode selector, and wherein an individual weighted Huber loss is computed for each selected candidate trajectory, and wherein individual weighted Huber losses are combined to generate a final combined loss for training of the trajectory generation model.

20 . The non-transitory computer-readable medium of claim 16 , wherein the one or more processors are further to:

receive a collection of generated trajectories from a machine learning model;

classify a trajectory of the collection of generated trajectories as an assert trajectory classification responsive to the trajectory satisfying a set of assert definition metrics;

classify a trajectory of the collection of generated trajectories as a yield trajectory classification responsive to the trajectory satisfying a set of yield definition metrics; and

determine a right-of-way-based semantic mode coverage metric based on assert trajectory classifications and yield trajectory classifications as compared to a total number of the generated trajectories in the collection.