Trajectory planning based on extracted trajectory features
A trajectory planning system can be used to select a trajectory for an autonomous vehicle. The trajectory planning system may generate multiple trajectories and extract features from the generated trajectories. The trajectory planning system may evaluate the trajectories based on the extracted features and select a trajectory for the vehicle based on the evaluation. The selected trajectory may be used to control the vehicle.
1 . A method, comprising:
obtaining scene data associated with a scene of a vehicle;
generating a plurality of trajectories for the vehicle based on the scene data;
extracting a plurality of trajectory features from a particular trajectory of the plurality of trajectories;
generating an image using the scene data;
determining a plurality of scene features from the image using a first machine learning model;
determining feature scores for individual features of the plurality of trajectory features, wherein the feature score of an individual feature is determined based at least on:
combining the individual feature with at least a portion of the plurality of scene features to result in a combined feature set; and
applying a portion of a second machine learning model to the combined feature set including the individual feature and the at least a portion of the plurality of scene features to generate the feature score for the individual feature;
wherein distinct portions of the second machine learning model are trained and used to separately process distinct features of the plurality of trajectory features; and
wherein determining feature scores for individual features of the plurality of trajectory features includes determining distinct feature scores for at least two individual features of the plurality of trajectory features, each distinct feature score based at least in part on a respective individual feature combined with at least a portion of the plurality of scene features;
determining a trajectory score for the particular trajectory based on the feature scores for individual features of the plurality of trajectory features;
selecting a first trajectory from the plurality of trajectories based on the trajectory score; and
causing the vehicle to be controlled based on the first trajectory.
2 . The method of claim 1 , wherein the plurality of trajectories includes a completed trajectory.
3 . The method of claim 1 , wherein obtaining the scene data includes receiving at least one of map data associated with a map corresponding to the scene, route data associated with a route for the vehicle, object data associated with at least one object identified in the scene, location data associated with a location of the vehicle.
4 . The method of claim 1 , wherein generating the plurality of trajectories comprises simulating a plurality of groups of actions to perform in sequence.
5 . The method of claim 4 , wherein the groups of actions comprise at least one of accelerating, modifying a heading, decelerating, or maintaining velocity.
6 . The method of claim 1 , wherein evaluating the particular trajectory comprises:
determining a feature score for each feature of the plurality of features,
weighting each of the plurality of feature scores, and
combining the plurality of weighted feature scores to determine a trajectory score for the particular trajectory.
7 . The method of claim 1 , wherein the image is a birds-eye-view image.
8 . The method of claim 1 , wherein evaluating the particular trajectory comprises:
determining a feature score for each feature of the plurality of features, and
combining the plurality of feature scores to determine a trajectory score for the particular trajectory.
9 . A system, comprising:
a data store storing computer-executable instructions; and
a processor configured to execute the computer-executable instructions, wherein execution of the computer-executable instructions causes the system to:
obtain scene data associated with a scene of a vehicle;
generate a plurality of trajectories for the vehicle based on the scene data;
extract a plurality of trajectory features from a particular trajectory;
generate an image using the scene data;
determine a plurality of scene features from the image using a first machine learning model;
determine feature scores for individual features of the plurality of trajectory features, wherein the feature score of an individual feature is determined based at least on instructions to:
combine the individual feature with at least a portion of the plurality of scene features to result in a combined feature set; and
apply a portion of a second machine learning model to the combined feature set including the individual feature and the at least a portion of the plurality of scene features to generate the feature score for the individual feature;
wherein distinct portions of the second machine learning model are trained and used to separately process distinct features of the plurality of trajectory features; and
wherein determining feature scores for individual features of the plurality of features includes determining distinct feature scores for at least two individual features of the plurality of features, each distinct feature score based at least in part on a respective individual feature combined with at least a portion of the plurality of scene features;
determining a trajectory score for the particular trajectory based on feature scores for individual features of the plurality of trajectory features;
select a first trajectory from the plurality of trajectories based on the trajectory score; and
cause the vehicle to be controlled based on the first trajectory.
10 . The system of claim 9 , wherein to evaluate the particular trajectory, execution of the computer-executable instructions further cause the system to:
determine a feature score for each feature of the plurality of features,
weight each of the plurality of feature scores, and
combine the plurality of weighted feature scores to determine a trajectory score for the particular trajectory.
11 . The system of claim 9 , wherein the plurality of trajectories includes a completed trajectory.
12 . The system of claim 9 , wherein to evaluate the particular trajectory, execution of the computer-executable instructions further cause the system to:
determine a feature score for each feature of the plurality of features, and
combine the plurality of feature scores to determine a trajectory score for the particular trajectory.
13 . Non-transitory computer-readable media comprising computer-executable instructions that, when executed by a computing system, causes the computing system to:
obtain scene data associated with a scene of a vehicle;
generate a plurality of trajectories for the vehicle based on the scene data;
extract a plurality of trajectory features from a particular trajectory;
generate an image using the scene data;
determine a plurality of scene features from the image using a first machine learning model;
determine feature scores for individual features of a plurality of trajectory features, wherein the feature score of an individual feature is determined based on instructions to:
combine the individual feature with at least a portion of the plurality of scene features to result in a combined feature set; and
apply a portion of a second machine learning model to the combined feature set including the individual feature and the at least a portion of the plurality of scene features to generate the feature score for the individual feature;
wherein distinct portions of the second machine learning model are trained and used to separately process distinct features of the plurality of trajectory features; and
wherein determining feature scores for individual features of the plurality of trajectory features includes determining distinct feature scores for at least two individual features of the plurality of trajectory features, each distinct feature score based at least in part on a respective individual feature combined with at least a portion of the plurality of scene features;
determine a trajectory score for the particular trajectory based on feature scores for individual features of the plurality of trajectory features;
select a first trajectory from the plurality of trajectories based on the trajectory score; and
cause the vehicle to be controlled based on the first trajectory.
14 . The non-transitory computer-readable media of claim 13 , wherein to evaluate the particular trajectory, execution of the computer-executable instructions further cause the computing system to:
determine a feature score for each feature of the plurality of features,
weight each of the plurality of feature scores, and
combine the plurality of weighted feature scores to determine a trajectory score for the particular trajectory.
15 . The non-transitory computer-readable media of claim 13 , wherein the image is a birds-eye-view image.
16 . The non-transitory computer-readable media of claim 13 , wherein the plurality of trajectories includes a completed trajectory.
17 . The non-transitory computer-readable media of claim 13 , wherein to evaluate the particular trajectory, execution of the computer-executable instructions further cause the computing system to:
determine a feature score for each feature of the plurality of features, and
combine the plurality of feature scores to determine a trajectory score for the particular trajectory.