Assessing surprise for autonomous vehicles

A system comprising:

• one or more computers and one or more storage devices storing instructions that are operable, when executed by the one or more computers, to cause the one or more computers to perform operations comprising:
• generating, for an initial time using a generative model that predicts vehicle states from the perspective of an observer of the vehicle, a first plurality of predicted vehicle states of the vehicle at a future point in time;
• obtaining, from a vehicle planning system of the vehicle, a first candidate trajectory for the vehicle to follow, the first candidate trajectory being one of a plurality of candidate trajectories to be evaluated;
• generating, by the generative model, a second plurality of predicted vehicle states based on the first candidate trajectory, the second plurality of predicted vehicle states representing predicted vehicle states at the future point in time assuming that the vehicle has initiated the first candidate trajectory generated by the vehicle planning subsystem, wherein the second plurality of predicted vehicle states are based on a set of characteristics for the vehicle after initiating the first candidate trajectory;
• generating a measure of observer surprise for the observer of the vehicle including comparing the first plurality of predicted vehicle states to the second plurality of predicted vehicle states based on the first candidate trajectory, the measure of observer surprise representing how surprising the vehicle following the first candidate trajectory would be to the observer;
• and
• based on the measure of observer surprise, causing the vehicle to select and follow a different candidate trajectory from the plurality of candidate trajectories.

One or more non-transitory computer storage media encoded with computer program instructions that when executed by one or more computers cause the one or more computers to perform operations comprising:

• generating, for an initial time using a generative model that predicts vehicle states from the perspective of an observer of the vehicle, a first plurality of predicted vehicle states of the vehicle at a future point in time;
• obtaining, from a vehicle planning system of the vehicle, a first candidate trajectory for the vehicle to follow, the first candidate trajectory being one of a plurality of candidate trajectories to be evaluated;
• generating, by the generative model, a second plurality of predicted vehicle states based on the first candidate trajectory, the second plurality of predicted vehicle states representing predicted vehicle states at the future point in time assuming that the vehicle has initiated the first candidate trajectory generated by the vehicle planning subsystem, wherein the second plurality of predicted vehicle states are based on a set of characteristics for the vehicle after initiating the first candidate trajectory;
• generating a measure of observer surprise for the observer of the vehicle including comparing the first plurality of predicted vehicle states to the second plurality of predicted vehicle states based on the first candidate trajectory, the measure of observer surprise representing how surprising the vehicle following the first candidate trajectory would be to the observer;
• and
• based on the measure of observer surprise, causing the vehicle to select and follow a different candidate trajectory from the plurality of candidate trajectories.