Data-driven warm start selection for optimization-based trajectory planning

1. A method of trajectory planning for an autonomous vehicle, comprising:

generating a plurality of candidate trajectories, the plurality of candidate trajectories including a sequence of waypoints for the autonomous vehicle associated with corresponding time instances;

computing features of a path associated with each of the plurality of candidate trajectories;

computing scores for the computed features of the plurality of candidate trajectories, wherein the computed scores are based on parameter values of the plurality of candidate trajectories;

evaluating the computed scores for each of the plurality of candidate trajectories, and selecting, based on the evaluation of the computed scores of the plurality of candidate trajectories, a trajectory from among the plurality of candidate trajectories to be used as a warm-start trajectory for trajectory optimization and applying the warm-start trajectory to develop a final trajectory for the autonomous vehicle; and

operating the autonomous vehicle in accordance with the final trajectory.

2. The method of claim 1 , wherein computing features of the path associated with each of the plurality of candidate trajectories comprises feature mapping to map each trajectory ξ into a multi-dimensional feature vector.

3. The method of claim 1 , further comprising using a set of candidate trajectories of the plurality of candidate trajectories and scene features of the set of candidate trajectories as training data to train a machine learning module to predict scores for the set of candidate trajectories.

4. The method of claim 1 , further comprising generating a decision tree to determine an index for each of a plurality of selected warm-start trajectory candidates offline and, during driving, using the decision tree to select the warm-start trajectory.

5. The method of claim 1 , wherein the computing scores for the computed features of the plurality of candidate trajectories comprises predicting scores off-line based on data examples, and using this prediction on-line to select a subset of a smaller subset of the plurality of candidate trajectories, and selecting the warm-start trajectory from among the selected subset of the plurality of candidate trajectories.

6. The method of claim 1 , further comprising using a machine learning module to predict a base ranking function used to compute the scores.

7. The method of claim 6 , wherein inputs to the machine learning module comprise the plurality of candidate trajectories and the computed features for the plurality of candidate trajectories.

8. The method of claim 6 , wherein the using the machine learning module comprises:

computing optimal trajectories based on the plurality of candidate trajectories;

computing scores for each of the optimal trajectories; and

generating a training data set.

9. The method of claim 8 , wherein the developing the final trajectory for the autonomous vehicle comprises using the selected warm-start trajectory to initialize the trajectory optimization and using the trajectory optimization to determine an optimal trajectory.

10. The method of claim 8 , wherein the training data set comprises training samples for planning problems generated and solved during an autonomous test drive.

11. The method of claim 1 , further comprising selecting warm-start candidate trajectories to be included in the plurality of candidate trajectories, the selecting comprising ranking warm-start candidate trajectories during in-vehicle autonomous operation or driving tests, prioritizing the warm-start candidate trajectories according to their respective rankings, and choosing a subset of ranked warm-start candidate trajectories as the plurality of candidate trajectories based on their respective rankings.

12. A vehicle control system for trajectory planning for an autonomous vehicle, comprising:

a processor; and

a memory coupled to the processor to store instructions, which when executed by the processor, cause the processor to perform operations, the operations comprising:

computing features of a path associated with each of a plurality of candidate trajectories;

computing scores for the computed features of the plurality of candidate trajectories, wherein the computed scores are based on parameter values of the plurality of candidate trajectories;

evaluating the computed scores for each of the plurality of candidate trajectories, and selecting, based on the evaluation of the computed scores of the plurality of candidate trajectories, a trajectory from among the plurality of candidate trajectories to be used as a warm-start trajectory for trajectory optimization and applying the warm-start trajectory to develop a final trajectory for the autonomous vehicle; and

operating the autonomous vehicle in accordance with the final trajectory.

13. The vehicle control system of claim 12 , wherein the computing features of the path associated with each of the plurality of candidate trajectories comprises feature mapping to map each trajectory ξ into a multi-dimensional feature vector.

14. The vehicle control system of claim 12 , wherein the operations further comprise using candidate trajectories and their associated scene features as training data to train a machine learning module to predict scores for the candidate trajectories.

15. The vehicle control system of claim 12 , wherein the operations further comprise generating a decision tree to determine an index for each of a plurality of selected warm-start trajectory candidates offline and, during driving, using the decision tree to select the warm-start trajectory.

16. The vehicle control system of claim 12 , wherein the computing scores for the computed features of the plurality of candidate trajectories comprises predicting scores off-line based on data examples, and using this prediction on-line to select a subset of a smaller quantity of the plurality of candidate trajectories, and selecting the warm-start trajectory from among the selected subset of the plurality of candidate trajectories.

17. The vehicle control system of claim 12 , wherein the operations further comprise using a machine learning module to predict a base ranking function used to compute the scores.

18. The vehicle control system of claim 17 , wherein inputs to the machine learning module comprise the plurality of candidate trajectories and the computed features for the plurality of candidate trajectories.

19. The vehicle control system of claim 17 , wherein the using the machine learning comprises:

computing optimal trajectories based on the plurality of candidate trajectories;

computing scores for each of the optimal trajectories; and

generating a training data set.

20. The vehicle control system of claim 19 , wherein the developing the final trajectory for the autonomous vehicle comprises using the selected warm-start trajectory to initialize the trajectory optimization and using the trajectory optimization to determine an optimal trajectory.

21. The vehicle control system of claim 19 , wherein the training data set comprises training samples for planning problems generated and solved during an autonomous test drive.

22. The vehicle control system of claim 12 , wherein the operations further comprise selecting warm-start candidate trajectories to be included in the plurality of candidate trajectories, the selecting comprising ranking warm-start candidate trajectories during in-vehicle autonomous operation or driving tests, prioritizing the warm-start candidate trajectories according to their respective rankings, and choosing a subset of ranked warm-start candidate trajectories as the plurality of candidate trajectories based on their respective rankings.

23. The method of claim 1 , further comprising:

generating a plurality of candidate maneuvers each defined by a corresponding candidate trajectory from the plurality of candidate trajectories to achieve the candidate maneuvers;

wherein the path is further associated with each of the plurality of candidate trajectories for the candidate maneuvers.

24. The vehicle control system of claim 12 , wherein a plurality of candidate maneuvers are each defined by a corresponding candidate trajectory from the plurality of candidate trajectories to achieve the candidate maneuvers, and wherein the path is further associated with each of the plurality of candidate trajectories for the candidate maneuvers.