MULTI-HEADED RECURRENT NEURAL NETWORK (RNN) FOR MULTI-CLASS TRAJECTORY PREDICTIONS

摘要

Described herein are various technologies that pertain to controlling an AV based upon a neural-network-generated prediction of a trajectory (e.g., one or more future positions) of an object in the driving environment of the AV. With more specificity, described herein are various technologies pertaining to generating predictions of a trajectory of an object by way of a multi-headed recurrent neural network (RNN). The multi-headed RNN is configured to output probability data that indicates a probability of an object moving to a position at a future point in time when the object has each of a plurality of different object characteristics. The multi-headed RNN is further configured to use state information of the multi-headed RNN from prior predictions of a trajectory of an object in connection with generating a future prediction of a trajectory of the object.