Predicting yielding likelihood for an agent
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for predicting how likely it is that a target agent in an environment will yield to another agent when the pair of agents are predicted to have overlapping future paths. In one aspect, a method comprises obtaining a first trajectory prediction specifying a predicted future path for a target agent in an environment; obtaining a second trajectory prediction specifying a predicted future path for another agent in the environment; determining that, at an overlapping region, the predicted future path for the target agent overlaps with the predicted future path for the other agent; and in response: providing as input to a machine learning model respective features for the target agent and the other agent; and obtaining the likelihood score as output from the machine learning model.
1. A method performed by an on-board system of a vehicle, the method comprising:
obtaining, by the on-board system of the vehicle, a first trajectory prediction specifying a predicted future path for a target agent in a vicinity of the vehicle in an environment, the target agent being different from the vehicle;
obtaining, by the on-board system of the vehicle, a second trajectory prediction specifying a predicted future path for another agent in the vicinity of the vehicle in the environment, the other agent being different from both the vehicle and the target agent;
determining by the on-board system of the vehicle that, at an overlapping region in the environment, the predicted future path for the target agent overlaps with the predicted future path for the other agent; and
in response:
providing, by the on-board system of the vehicle, as input to a machine learning model respective features that describe the target agent and the other agent, wherein the machine learning model is implemented within the on-board system of the vehicle and is configured to predict a likelihood score representing a predicted likelihood that the target agent will yield to the other agent when entering the overlapping region, given that the target agent will follow the predicted future path for the target agent and that the other agent will follow the predicted future paths for the other agent; and
obtaining, by the on-board system of the vehicle, the likelihood score as output from the machine learning model.
2. The method of claim 1 , further comprising, if the likelihood score is greater than a predetermined threshold value:
modifying the first trajectory prediction for the target agent in the environment to account for the target agent yielding to the other agent.
3. The method of claim 1 , wherein the respective features for the target agent comprise (i) features that describe the target agent and (ii) features that describe the predicted future path for the target agent.
4. The method of claim 1 , wherein the respective features for the other agent comprise (i) features that describe the other agent from a perspective of the target agent and (ii) features that describe the predicted future path for the other agent from the perspective of the target agent.
5. The method of claim 1 , wherein the respective features for the target agent and the other agent comprise pairwise features that describe interactions between the target agent and the other agent.
6. The method of claim 1 , further comprising:
providing as input to the machine learning model respective features for the target agent and the vehicle in the environment, the vehicle having a planned future path generated by the on-board system of the vehicle that overlaps with the predicted future path of the target agent at a same or different overlapping region in the environment, wherein the respective features comprise pairwise features that describe interactions between the target agent and the vehicle; and
obtaining as output of the machine learning model a likelihood score representing a predicted likelihood that the target agent will yield to the vehicle when entering the same or different overlapping region.
7. The method of claim 1 , wherein the machine learning model is a gradient boosting decision forest model.
8. A method of training a machine learning model having a plurality of model parameters, the method comprising:
generating training data for the machine learning model, wherein the machine learning model is configured to receive as input (i) a first plurality of features for a target agent in an environment including features that describe a predicted future path for the target agent and (ii) a second plurality of features for another agent in the environment including features that describe a predicted future path for the other agent, and to process the first plurality of features and the second plurality of features in accordance with the plurality of model parameters to generate an output that assigns a likelihood score to the target agent, the likelihood score representing a predicted likelihood that the target agent will yield to the other agent, given that the target agent will follow the predicted future path for the target agent and that the other agent will follow the predicted future path for the other agent, and wherein generating the training data comprisesi+g:
obtaining a plurality of candidate training examples, wherein each candidate training example includes features for a corresponding target agent and a corresponding other agent at a corresponding time point, wherein at the corresponding time point the target agent and the other agent had respective predicted future paths that overlapped at an overlapping region; and
identifying as a positive training example each candidate training example for which:
both agents followed the respective predicted paths after the corresponding time point, and
the other agent entered the overlapping region prior to the target agent; and
training the machine learning model on the training data.
9. The method of claim 8 , wherein generating training data for the machine learning model further comprises:
identifying as a negative training example each candidate training example for which:
both agents followed the respective predicted paths after the corresponding time point, and
the target agent entered the overlapping region prior to the other agent.
10. The method of claim 8 , wherein obtaining the plurality of candidate training examples comprises:
determining, for each candidate training example, that the predicted future path for the corresponding target agent has a lower or equal precedence at the overlapping region compared to the predicted future path for the corresponding other agent.
11. An on-board system of a vehicle comprising:
one or more computers; and
one or more storage devices communicatively coupled to the one or more computers, wherein the one or more storage devices store instructions that, when executed by the one or more computers, cause the one or more computers to perform operations comprising:
obtaining, by the on-board system of the vehicle, a first trajectory prediction specifying a predicted future path for a target agent in a vicinity of the vehicle in an environment, the target agent being different from the vehicle;
obtaining, by the on-board system of the vehicle, a second trajectory prediction specifying a predicted future path for another agent in the vicinity of the vehicle in the environment, the other agent being different from both the vehicle and the target agent;
determining by the on-board system of the vehicle that, at an overlapping region in the environment, the predicted future path for the target agent overlaps with the predicted future path for the other agent; and
in response:
providing, by the on-board system of the vehicle, as input to a machine learning model respective features that describe the target agent and the other agent, wherein the machine learning model is implemented within the on-board system of the vehicle and is configured to predict a likelihood score representing a predicted likelihood that the target agent will yield to the other agent when entering the overlapping region, given that the target agent will follow the predicted future path for the target agent and that the other agent will follow the predicted future paths for the other agent; and
obtaining, by the on-board system of the vehicle, the likelihood score as output from the machine learning model.
12. The on-board system of claim 11 , wherein the operations further comprise, if the likelihood score is greater than a predetermined threshold value:
modifying the first trajectory prediction for the target agent in the environment to account for the target agent yielding to the other agent.
13. The on-board system of claim 11 , wherein the respective features for the target agent comprise (i) features that describe the target agent and (ii) features that describe the predicted future path for the target agent.
14. The on-board system of claim 11 , wherein the respective features for the other agent comprise (i) features that describe the other agent from a perspective of the target agent and (ii) features that describe the predicted future path for the other agent from the perspective of the target agent.
15. The on-board system of claim 11 , wherein the respective features for the target agent and the other agent comprise pairwise features that describe interactions between the target agent and the other agent.
16. One or more non-transitory computer storage media storing instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:
obtaining, by an on-board system of the vehicle that comprises the one or more computers, a first trajectory prediction specifying a predicted future path for a target agent in a vicinity of the vehicle in an environment, the target agent being different from the vehicle;
obtaining, by the on-board system of the vehicle, a second trajectory prediction specifying a predicted future path for another agent in the vicinity of the vehicle in the environment, the other agent being different from both the vehicle and the target agent;
determining by the on-board system of the vehicle that, at an overlapping region in the environment, the predicted future path for the target agent overlaps with the predicted future path for the other agent; and
in response:
providing, by the on-board system of the vehicle, as input to a machine learning model respective features that describe the target agent and the other agent, wherein the machine learning model is implemented within the on-board system of the vehicle and is configured to predict a likelihood score representing a predicted likelihood that the target agent will yield to the other agent when entering the overlapping region, given that the target agent will follow the predicted future path for the target agent and that the other agent will follow the predicted future paths for the other agent; and
obtaining, by the on-board system of the vehicle, the likelihood score as output from the machine learning model.
17. The non-transitory computer storage media of claim 16 , wherein the operations further comprise, if the likelihood score is greater than a predetermined threshold value:
modifying the first trajectory prediction for the target agent in the environment to account for the target agent yielding to the other agent.
18. The method of claim 6 , further comprising:
determining, based at least on the likelihood scores, whether the target agent will yield to the other agent or yield to the vehicle.
19. The on-board system of claim 11 , wherein the operations further comprise:
providing as input to the machine learning model respective features for the target agent and the vehicle in the environment, the vehicle having a planned future path generated by the on-board system of the vehicle that overlaps with the predicted future path of the target agent at a same or different overlapping region in the environment, wherein the respective features comprise pairwise features that describe interactions between the target agent and the vehicle; and
obtaining as output of the machine learning model a likelihood score representing a predicted likelihood that the target agent will yield to the vehicle when entering the same or different overlapping region.
20. The on-board system of claim 19 , wherein the operations further comprise:
determining, based at least on the likelihood scores, whether the target agent will yield to the other agent or yield to the vehicle.
21. The non-transitory computer storage media of claim 16 , wherein the operations further comprise:
providing as input to the machine learning model respective features for the target agent and the vehicle in the environment, the vehicle having a planned future path generated by the on-board system of the vehicle that overlaps with the predicted future path of the target agent at a same or different overlapping region in the environment, wherein the respective features comprise pairwise features that describe interactions between the target agent and the vehicle; and
obtaining as output of the machine learning model a likelihood score representing a predicted likelihood that the target agent will yield to the vehicle when entering the same or different overlapping region.
22. The non-transitory computer storage media of claim 21 , wherein the operations further comprise:
determining, based at least on the likelihood scores, whether the target agent will yield to the other agent or yield to the vehicle.